JP2022532177A - Forged face recognition methods, devices, and non-temporary computer-readable storage media - Google Patents

Abstract

FIELD OF THE DISCLOSURE The present disclosure relates to a forged face recognition method, apparatus, and computer-readable storage medium in the technical field of computers. The method of the present disclosure is a forged face recognition method comprising the steps of obtaining a small sample set of forged face images of a new category; updating the parameters of the trained face anti-counterfeiting model, wherein the parameters of the pre-trained face anti-counterfeiting model are determined based on the plurality of sets of parameters, and the parameters of the pre-trained face anti-counterfeiting model A plurality of sets are each determined based on a plurality of tasks, each task comprising a small sample set of forged face images for each of one or more existing categories and each of one or more existing categories. and inputting the face image to be recognized to the anti-forgery model of the face after parameter update, so that the face image to be recognized is the forged face image and determining whether there is.

Description

Cross-reference to related applications This application claims priority in Chinese Patent Application No. 201910387689.0 filed May 10, 2019, which is incorporated herein by reference in its entirety.

The present disclosure relates to the field of computer technology, in particular to counterfeit face recognition methods, devices, and computer readable storage media.

Face anti-counterfeiting recognition refers to the effective identification of counterfeit facial features. For example, a facial recognition system may produce several forged faces (videos, photos, masks, etc.) to attack the facial recognition system. Face anti-counterfeiting recognition aims to identify counterfeit facial features. Generally, in a face recognition system, a user needs to pass the detection of recognition for preventing forgery of a face, which is also called biological detection.

A method of recognizing counterfeit faces known to the inventor is to perform deep learning training on the classification model based on a large number of counterfeit facial images in known categories. The trained classification model may recognize that the forged face belongs to the same category as the forged face image for training.

The inventor found that in existing counterfeit face recognition methods, it was necessary to use a large number of samples to train the model so that the trained model could accurately recognize the counterfeit face. There is. With the advent of new categories of forged facial images, it is not possible to achieve reopened model training if the sample size of the forged facial images is very small, and as a result, the model is forged in the new category. The face is not recognized in a timely manner, making it impossible to effectively prevent a new category of counterfeit face attacks in a timely manner.

One technical challenge to be resolved by this disclosure is to achieve timely and effective recognition of new categories of forged facial images, and for faces when a small number of new categories of forged facial images appear. This is a way to improve the security of the recognition system.

Some embodiments of the present disclosure provide forged face recognition methods. The method is to get a small sample set of forged facial images in a new category and update the parameters of a pre-trained facial anti-counterfeiting model with a small sample set of forged facial images in the new category. Steps to be performed, where the parameters of the pre-trained facial anti-counterfeiting model are determined based on multiple sets of parameters, each determined on the basis of multiple tasks, with each task being one or more existing. Face anti-counterfeiting after steps and parameter updates, including a small sample set of each forged face image in each of the categories and a small sample set of each actual face image in one or more existing categories. It includes a step of inputting a face image to be recognized into the model and determining whether the face image to be recognized is a forged face image.

In some embodiments, the method is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. If the anti-counterfeit model contains a step to update the parameters of the anti-counterfeit model and the anti-counterfeit model is a classification model, the classification label is configured to label the category of forged face images corresponding to the classification label. If the classification label is 1 and the category contains at least one of the counterfeiting methods or models, or the anti-counterfeiting model of the face is a regression model, the classification label is a forged face image corresponding to the classification label. A second classification label configured to be labeled as a positive sample or a negative sample.

In some embodiments, a face anti-counterfeiting model is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. The step to update the parameters of is to enter the forged face image in a small sample set of forged face images in a new category into the face anti-counterfeiting model as a support sample image for one task. Steps to determine the gradient of facial anti-counterfeiting model parameters based on the output value of the face anti-counterfeiting model for each of the forged facial images and the classification label corresponding to each of the forged facial images. Includes a step to update the face anti-counterfeiting model parameters based on the facial anti-counterfeiting model parameter gradient.

In some embodiments, if the anti-counterfeit model of the face is a classification model, the output value of the anti-counterfeit model for each of the counterfeit face images is the probability that each of the counterfeit face images belongs to a different category. The gradient of the parameters of the anti-counterfeiting model of the face is determined based on the output value of the anti-counterfeiting model of the face for each of the counterfeited face images and the classification label corresponding to each of the counterfeited face images. The first of the parameters of the anti-counterfeiting model of the face is based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is a regression model, or if the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is a positive sample for each of the counterfeited face images. The gradient of the parameters of the anti-counterfeit model based on the output value of the anti-counterfeit model of the face for each of the counterfeit face images and the classification label corresponding to each of the counterfeit face images. The steps to determine the parameters of the anti-counterfeiting model of the face are based on the probability that each of the forged face images belongs to a positive sample and the second classification label that corresponds to each of the forged face images. Provide a step to determine the second gradient.

In some embodiments, the recognition method is further a step of dividing the forged face image of the existing category and the actual face image of the existing category into a plurality of tasks, and each task has a plurality of methods. And each method has multiple real facial images of one model or multiple counterfeit facial images of existing categories, and each task has a first preset number of existing categories. Each method has a second preset number of real or forged face images as a small sample set consisting of support sample images, and a third preset as a query sample image. A step with a number of real face images or a forged face image, and a fourth preset number of tasks to form a batch each time, a parameter of the face anti-counterfeiting model. Multiple sets of are determined based on multiple tasks in the batch, and the parameters of the face anti-counterfeiting model are redetermined based on multiple sets of parameters, the steps and the face anti-counterfeiting model converge. Further prepare for steps to complete pre-training of the face anti-counterfeiting model until the conditions are reached.

In some embodiments, a plurality of sets of anti-counterfeiting model parameters are determined based on a plurality of tasks in a batch, and the parameters of the anti-counterfeiting model are re-based on a plurality of sets of parameters. What is determined is to determine the third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and the face based on the third gradient for each task in the batch. Determining one set of parameters for the anti-counterfeiting model and corresponding to each of the set of parameters determined by each of the multiple tasks in the batch, the query sample image corresponding to multiple tasks, and the query sample image. It involves determining the fourth gradient corresponding to the batch based on the classification label to be used and redetermining the parameters of the anti-counterfeiting model based on the fourth gradient corresponding to the batch.

In some embodiments, when the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label. Yes, the category contains at least one of the counterfeit methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. Is.

In some embodiments, the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method.

In some embodiments, different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises a photo, video, or mask, and the model is a depth model, near. Actual facial images of different tasks, including infrared models, or red-green-blue (RGB) models, have the same model or different models, and a new category of counterfeit facial images is within at least one of the counterfeit methods or models. Unlike existing images in, existing images include forged facial images in existing categories and actual facial images in existing categories.

In another embodiment of the present disclosure, a forged face recognition device is provided. The device has a pre-trained face with a small sample acquisition module configured to acquire a small sample set of forged facial images in a new category and a small sample set of forged facial images in the new category. A model tuning module configured to update the parameters of the anti-counterfeiting model of the pre-trained facial anti-counterfeiting model into multiple sets of parameters, each determined based on multiple tasks. Determined on the basis, each task is a small sample set of each forged facial image in one or more existing categories and a small sample set of each actual facial image in one or more existing categories. Enter a model adjustment module, including, and a face image to be recognized by the face anti-counterfeiting model after parameter updates to determine if the face image to be recognized is a forged face image. It is equipped with a face recognition module configured in.

In some embodiments, the model adjustment module is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is configured to update the parameters and the anti-counterfeiting model of the face is a classification model, the classification label is configured to label the category of forged face images corresponding to the classification label. If it is the first classification label given and the category contains at least one of the counterfeiting methods or models, or if the anti-counterfeiting model of the face is a regression model, the classification label is counterfeited corresponding to the classification label. A second classification label configured to label the facial image as a positive sample or a negative sample.

In some embodiments, the model adjustment module inputs the forged face image in a small sample set of forged face images in a new category into the face anti-counterfeiting model as a support sample image for one task. Then, the gradient of the parameters of the anti-counterfeiting model of the face is determined based on the output value of the anti-counterfeiting model of the face for each of the forged facial images and the classification label corresponding to each of the forged facial images. It is configured to update the parameters of the face anti-counterfeiting model based on the gradient of the face anti-counterfeiting model parameters.

In some embodiments, when the anti-counterfeiting model of the face is a classification model, the output value of the anti-counterfeiting model for each of the counterfeit face images is the probability that each of the counterfeit face images belongs to a different category. And the model adjustment module is based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images of the anti-counterfeiting model of the face. It is configured to determine the first gradient of the parameter.

Alternatively, if the anti-counterfeiting model of the face is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a positive sample, and the model adjustment. A second of the parameters of the anti-counterfeiting model of the face, based on the probability that each of the forged facial images belongs to a positive sample and the second classification label that corresponds to each of the forged facial images. It is configured to determine the gradient.

In some embodiments, the recognition device is to divide the forged face image of an existing category from the actual face image of an existing category into a plurality of tasks, each task having a plurality of methods. Each method has multiple real facial images of one model or multiple forged facial images of existing categories, and each task has a first preset number of existing categories. Each method has a second preset number of real or forged face images as a small sample set consisting of support sample images, and a third preset number as query sample images. It has a real face image or a forged face image, and a fourth preset number of tasks is to be selected to form a batch each time, and the parameters of the face anti-counterfeiting model. Multiple sets are each determined based on multiple tasks in the batch, face anti-counterfeiting model parameters are redetermined based on multiple sets of parameters, and the face anti-counterfeiting model is a convergence condition. Further equipped with a pre-training module configured to complete and perform pre-training of the face anti-counterfeiting model until.

In some embodiments, the pre-training module determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and a third for each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on the gradient of, and a set of parameters determined by each of the multiple tasks in the batch, query sample images corresponding to multiple tasks, and queries. Determining the fourth gradient for the batch based on the classification label corresponding to each of the sample images, and redetermining the parameters of the anti-counterfeiting model based on the fourth gradient corresponding to the batch. Is configured to do.

In some embodiments, if the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label. Yes, if the category contains at least one of the counterfeiting methods or models, or if the anti-counterfeiting model of the face is a regression model, then each of the classification labels is a positive sample or a negative sample of the support sample image corresponding to the classification label. A second classification label configured to be labeled as a sample.

In some embodiments, the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method.

In some embodiments, different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises a photo, video, or mask, and the model is a depth model, near. Actual facial images of different tasks, including infrared models, or red-green-blue (RGB) models, have the same model or different models, and a new category of counterfeit facial images is within at least one of the counterfeit methods or models. Unlike existing images in, existing images include forged facial images in existing categories and actual facial images in existing categories.

Yet another embodiment of the present disclosure provides a forged face recognition device. The device comprises a memory and a processor coupled to the memory, the processor configured to perform the forged face recognition method according to any of the above embodiments based on the instructions stored in the memory. Will be done.

In yet another embodiment of the present disclosure, a computer-readable storage medium is provided. The storage medium stores a computer program that implements the method according to any of the above embodiments when executed by the processor.

In this disclosure, the parameters of the pre-trained facial anti-counterfeiting model are updated directly with a small sample set of forged facial images in a new category, resulting in the updated facial anti-counterfeiting model. It can recognize various forged facial images in existing and new categories. This is a parameter of the face anti-counterfeiting model, where a pre-trained facial anti-counterfeiting model is trained by a small sample set of forged facial images in the existing category and a small sample set of real facial images. This is because multiple sets of each are determined by multiple tasks, and then the parameters of the face anti-counterfeiting model are redetermined based on the multiple sets of parameters during the training process. This training process consists of two stages: the first stage of training a facial anti-counterfeiting model using multiple small sample sets of forged facial images in an existing category and real facial images in an existing category. And with the first stage, where the model learns the features of a small sample set, or the stage where it learns to distinguish between a fake face and a real face, as well as performing aggregation. The second step is to re-update the parameters of the face anti-counterfeiting model based on multiple training results, learning how the model learns a small sample set of features, or the forged face and the actual It includes a second stage, which is a stage for learning how to learn to distinguish from the face.

Therefore, after learning multiple tasks with a small sample set of forged face images in an existing category and a real face image in an existing category, the face anti-counterfeiting model learns how to learn a small sample set. Learn how to learn to distinguish between a forged face and a real face, or learn to find forged clues, and get the appropriate initialization parameters. The anti-counterfeiting model of the face is updated if the new category of counterfeit facial images is related to the conditions of a small sample set. As an advantage of the fast learning ability previously learned, the face anti-counterfeiting model also performs fast convergence and fast learning of a new category of counterfeit facial features, making the face anti-counterfeiting model timely and effective. Updates can be performed, the forged face image of the new category can be recognized in a timely and effective manner, and the attack from the forged face of the new category can be effectively protected in a timely manner.

Other features and advantages of the present disclosure will become apparent from the subsequent detailed description of the exemplary embodiments of the present disclosure with reference to the accompanying drawings.

The accompanying drawings set forth herein for the purpose of providing further understanding of the present disclosure form part of this application. The embodiments and illustrations presented in this disclosure are intended to illustrate this disclosure, but do not constitute an inappropriate definition of this disclosure.

It is a schematic flowchart of the method of recognizing a forged face according to some embodiments of this disclosure. It is a schematic flowchart of the forged face recognition method by another embodiment of this disclosure. FIG. 3 is a schematic structural diagram of a forged face recognition device according to some embodiments of the present disclosure. FIG. 3 is a schematic structural diagram of a forged face recognition device according to some other embodiment of the present disclosure. FIG. 3 is a schematic diagram of the structure of a forged face recognition device according to still another embodiment of the present disclosure.

The technical solutions of the embodiments of the present disclosure will be described expressly and fully herein, along with the accompanying drawings of the embodiments of the present disclosure. Obviously, the embodiments described are not all of the embodiments, but only some of the embodiments of the present disclosure. The following description of at least one embodiment of the description, which is merely exemplary in practice, shall by no means serve as any limitation with respect to the present disclosure and its application or use. All other embodiments obtained by one of ordinary skill in the art on the premise that the efforts of the invention are not included in accordance with the embodiments of the present disclosure shall be included in the scope of protection of the present disclosure.

From a new category of forged faces when the model cannot timely and effectively recognize the new category of forged faces, new categories of forged face images appear, and very few samples. The solution is provided to solve the problem of not being able to prevent attacks in a timely and effective manner. A method for recognizing a forged face according to some embodiments of the present disclosure is described below together with FIG.

FIG. 1 shows a schematic flow chart of a forged face recognition method according to some embodiments of the present disclosure. As shown in FIG. 1, the method of this embodiment comprises steps S102-S106.

In step S102, a small sample set of forged facial images in a new category is obtained.

In the present disclosure, the number of images in a small sample set is less than the threshold. This threshold acknowledges that the anti-counterfeiting model is updated with a small sample set according to deep learning training methods, making it impossible for the model to converge. The threshold is, for example, 10 to 30, but is not limited to the examples listed. The new category of forged face images belongs to a different category than the forged face image category used when the face anti-counterfeiting model was pre-trained, and the face anti-counterfeiting model is pre-trained. The forged face category used when used can be referred to as an existing category.

Disguised facial images in different existing categories differ in at least one of the counterfeiting methods or models. Counterfeiting methods include, for example, photographs, videos, and masks, and models include, for example, depth (Depth) models, near-infrared (NIR) models, and red-green-blue (RGB) models. For example, a photo forgery method and a forged face image with a depth model belong to a different category than a photo forgery method and an RGB model forged face image, and a photo forgery method and a forged face image with a depth model It also belongs to a different category from video forgery methods and forged facial images with RGB models. Multi-model (multi-model) images refer to images collected by different technical means for describing different attributes of the same object. For example, for an object, a color image captured by a camera and a depth image scanned by a laser radar are called two model images of the object. Forged facial images in the new category differ from at least one existing image of the forged method or model, and existing images include forged facial images and actual facial images in the existing category.

In step S104, the parameters of the pre-trained facial anti-counterfeiting model are updated with a small sample set of forged facial images in a new category. Pre-trained facial anti-counterfeiting model parameters are determined based on multiple sets of parameters, each of which is a multi-tasking-determined set of facial anti-counterfeiting model parameters. Includes a small sample set of forged facial images for each of one or more existing categories, and a small sample set of real facial images.

In some embodiments, the parameters of the face anti-counterfeiting model are multiple forged face images in a small sample set of forged face images in a new category, and a classification label corresponding to each of the forged face images. And will be updated based on. The anti-counterfeiting model of the face may be a classification model or a regression model. If the anti-counterfeiting model of the face is a classification model, each classification label is the first classification label configured to label the category of forged face images corresponding to the classification label, and the category is counterfeit. Includes at least one of the methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each classification label is configured to label the forged face image corresponding to the classification label as a positive or negative sample. It is a classification label of 2. The first classification label may label at least one of the counterfeit methods or models of the counterfeit facial image. If it is necessary to learn different forgery methods and image features of different models, both the forgery method and the model are labeled.

Each of the counterfeit facial images in a small sample set of counterfeit facial images in the new category is entered into a pre-trained anti-counterfeiting model of the face, and the gradient of the parameters of the anti-counterfeiting model of the face is the output value, The parameters of the pre-trained anti-counterfeiting model, determined based on the classification label corresponding to each of the forged facial images and the preset loss function, are updated based on the gradient. If the face anti-counterfeiting model is a classification model, the preset loss function may be an MSE (Mean Squared Error) loss function. If the face anti-counterfeiting model is a regression model, the preset loss function may be a cross entropy loss function, but these are not limited to the examples taken. Methods of updating the parameters of a pre-trained anti-counterfeiting model based on the gradient are not detailed herein, but existing gradient descent methods may be used as desired.

In some embodiments, multiple forged facial images in a small sample set of forged facial images in a new category are entered into the facial anti-counterfeiting model as support sample images for one task and of the face. The gradient of the anti-counterfeiting model parameters is determined based on the output value of the face anti-counterfeiting model for each of the forged face images and the classification label corresponding to each of the forged face images to prevent face anti-counterfeiting. The model parameters are updated based on the gradient of the face anti-counterfeiting model parameters.

The task can be built on a small sample set of forged facial images in a new category. The forged face image in a small sample set of forged face images in the new category may be a support sample image for the task. This is similar to the format in which tasks are set by the pre-training method for forged facial images, which will be described later. The constructed task contains one or more different new categories, each of which has a corresponding small sample set. The task also has a small sample set of real facial images and the like.

In some embodiments, the face anti-counterfeiting model is a classification model, and the parameters of the face anti-counterfeiting model are each of the forged face images in a small sample set of forged face images in a new category. Updated based on a first classification label corresponding to each of the forged facial images. For example, the first classification label for the actual face image by the depth model is 0, the first classification label for the forged mask method and the forged face image by the depth model is 1, and the first in the new category. The classification label is 2.

The forged facial image and the actual facial image can be labeled with a first classification label and a second classification label depending on the application scenario. For example, during the process of updating the parameters of the face anti-counterfeiting model, the first classification label may be used to calculate the value of the loss function, the value of the gradient, etc. to improve the accuracy of the model. Can be done. When the face anti-counterfeiting model classifies images and outputs the results, it only needs to distinguish the images into two categories: real face images and forged face images. During the testing process after pre-training the anti-counterfeiting model of the image, a second classification label is used to recognize whether the image is forged or real, so that the anti-counterfeiting model of the face You can test the accuracy. For example, the second classification label representing a forged face image is 0 and the second classification label representing an actual face image is 1.

Further, when the face anti-counterfeiting model is a classification model, the output value of the face anti-counterfeiting model for each of the forged face images is the probability that each of the forged face images belongs to a different category, and the face. The first gradient of the anti-counterfeiting model parameters is based on the probability that each of the forged facial images belongs to a different category and the first classification label that corresponds to each of the forged facial images. To.

In some embodiments, the face anti-counterfeiting model may be a regression model, and the parameters of the face anti-counterfeiting model are each of the forged face images in a small sample set of forged face images in a new category. And a second classification label corresponding to each of the forged facial images. The regression model can finally output a value for the input image and set a threshold corresponding to the value. If the value is higher than the threshold, it means that the image is a real face image, and if the value is lower than the threshold, it means that the image is a forged face image. Therefore, the regression model may apply a second classification label for the image.

Further, when the anti-counterfeiting model of the image is a regression model, the output value of the anti-counterfeiting model of the face for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a positive sample. The gradient of 2 is determined based on the probability that each of the forged facial images belongs to a positive sample and the second classification label that corresponds to each of the forged facial images, and is a parameter of the anti-counterfeiting model of the face. Is updated based on the second gradient.

Pre-trained facial anti-counterfeiting model parameters are trained in the following way: multiple sets of facial anti-counterfeiting model parameters are determined based on multiple tasks, respectively, and then face counterfeiting. The parameters of the prevention model are redetermined based on multiple sets of parameters. Each task includes a small sample set of each forged facial image in one or more existing categories and a small sample set of actual facial images in each of one or more existing categories. The actual facial image is also divided into actual facial images of different models. Each task may include a small sample set of each actual facial image of one or more models. Actual facial images in different tasks have the same model or different models. That is, a small sample set composed of a forged face image of an existing category and an actual face image of an existing category is assigned to a plurality of tasks. A different small sample set should be used to train the anti-counterfeiting model, and after training the parameters of the anti-counterfeiting model should be redetermined based on multiple training results. The process of training a face anti-counterfeiting model with different small sample sets is the process of model learning, and the process of redetermining parameters based on multiple training results is the process of a "learning how to learn" model. ..

After the two processes described above, the model not only learns the features of the small sample set, but also learns how to learn the features of the small sample set, or distinguishes between a forged face and a real face. Rather than simply learning how, or simply learning forged clues, the anti-counterfeiting model of the face ultimately "how to learn to distinguish between a disguised face and a real face" and " Learn how to learn to find forged clues. Therefore, if the new category of counterfeit facial images belong to a small sample set, the face anti-counterfeiting model also performs fast convergence and fast learning of the new category of counterfeit facial features, effectively and timely. Defend Lee.

In step S106, the face image to be recognized is input to the face anti-counterfeiting model after the parameters are updated, and it is determined whether the face image to be recognized is a forged face image.

The facial image to be recognized may be a forged facial image that belongs to a new category or an existing category that can be recognized by the anti-counterfeiting model of the face.

In the method of the above embodiment, the parameters of the pre-trained face anti-counterfeiting model are updated directly with a small sample set of forged face images in a new category, resulting in the updated face forgery. The prevention model can recognize various forged facial images in existing and new categories. This is a parameter of the face anti-counterfeiting model, where a pre-trained facial anti-counterfeiting model is trained by a small sample set of forged facial images in the existing category and a small sample set of real facial images. This is because multiple sets of each are determined by multiple tasks, and then the parameters of the face anti-counterfeiting model are redetermined based on the multiple sets of parameters during the training process. This training process consists of two stages: the first stage of training a facial anti-counterfeiting model using multiple small sample sets of forged facial images in an existing category and real facial images in an existing category. And with the first stage, where the model learns the features of a small sample set, or the stage where it learns to distinguish between a fake face and a real face, as well as performing aggregation. The second step is to re-update the parameters of the face anti-counterfeiting model based on multiple training results, learning how the model learns a small sample set of features, or the forged face and the actual It includes a second stage, which is a stage for learning how to learn to distinguish from the face.

Therefore, after learning multiple tasks with a small sample set of forged face images in an existing category and a real face image in an existing category, the face anti-counterfeiting model learns how to learn a small sample set. Learn how to learn to distinguish between a forged face and a real face, or learn to find forged clues, and get the appropriate initialization parameters. The anti-counterfeiting model of the face is updated if the new category of counterfeit facial images is related to the conditions of a small sample set. As an advantage of the fast learning ability previously learned, the face anti-counterfeiting model also performs fast convergence and fast learning of a new category of counterfeit facial features, making the face anti-counterfeiting model timely and effective. Updates can be performed, the forged face image of the new category can be recognized in a timely and effective manner, and the attack from the forged face of the new category can be effectively protected in a timely manner.

According to another embodiment of the present disclosure, a forged face recognition method will be described below together with FIG. This embodiment includes a pre-training method and a method of operation of a face anti-counterfeiting model.

FIG. 2 shows a schematic flow chart of a forged face recognition method according to another embodiment of the present disclosure. As shown in FIG. 2, the method of the present embodiment includes steps S202 to S214.

In step S202, the face anti-counterfeiting model is first trained with a forged face image of an existing category and an actual face image of an existing category according to a deep learning training method, and the face before pre-training. The initial parameters of the anti-counterfeit model are determined.

First, the face anti-counterfeiting model is trained by using existing deep learning training methods, then proceed to the subsequent pre-training process, so that the two post-training facial anti-counterfeiting models are suitable. Provides distinction and generalization. Deep learning training methods allow facial anti-counterfeiting models to more accurately distinguish images, and pre-training methods of the present disclosure allow models to generalize different categories of forged facial images. The post-training facial anti-counterfeiting model not only provides quick adaptability to new tasks, but can also provide appropriate distinctions for known or similar tasks.

In step S204, the forged face image of the existing category and the actual face image of the existing category are divided into a plurality of tasks. Each task has multiple methods, each method has multiple real facial images of one model or multiple counterfeit facial images of existing categories, and each method consists of support sample images. A second preset number of real or forged face images as a small set of samples, and a third preset number of real or forged face images as a query sample image. Has a face image. Each task has a first preset number of existing categories. Each existing category represents a counterfeit method and / or model. Face anti-counterfeiting models are simply configured to recognize counterfeit facial images using different counterfeiting methods, existing categories are simply configured to represent counterfeiting methods, and all counterfeit facial images and The actual face image belongs to the same model. If the face anti-counterfeiting model is configured to recognize different counterfeiting methods and counterfeit facial images of different models, each existing category represents a counterfeiting method or model.

The task has multiple methods. Each method includes a second preset number of support sample images and a third preset number of query sample images. The support sample images for each method form a small sample set, and the support sample images for each method in each task are less than the threshold, eg 10-30. During the pre-training process, the number of images in the small sample set formed by the support sample images may be the same as or different from the number of images in the small sample set of counterfeit facial images in the new category. If the number of images in the small sample set formed by the support sample images is less than or equal to the number of images in the small sample set of forged face images in the new category, the face anti-counterfeiting model is a new category in the pre-training process. Subsequently updated based on a small sample set of, the face anti-counterfeiting model can present a better recognition effect. This is because the face anti-counterfeiting model learns how to learn a smaller number of samples during the pre-training process, and there are numerous images in the new category at the time of the update, so the model parameters are more updated. Can be achieved well.

Each time, a first preset number of existing categories is randomly selected, followed by a second preset number of support sample images and a third preset number of query sample images in the existing category. Can be randomly selected from selected forged facial images. The actual facial images of one or more models were randomly selected, followed by a second preset number of support sample images and a third preset number of query sample images. Randomly selected from the face image of the to generate a task. Images of different tasks cannot be exactly the same.

In step S206, a fourth preset number of tasks are selected to form a batch, and multiple sets of face anti-counterfeiting model parameters are determined, respectively, based on multiple tasks in the batch. The parameters of the anti-counterfeit model are redetermined based on multiple sets of parameters.

A fourth preset number of tasks are randomly selected to form a batch. For example, 5 tasks are selected as a batch. For each batch, the face anti-counterfeiting model parameters can be updated twice. The first update is to update the parameters of the face anti-counterfeiting model with support sample images in each task. The second update is to update the parameters of the face anti-counterfeiting model with the query sample image in each task. That is, the update is composed of the above-mentioned learning process and the process of "learning the learning method". This is different from existing deep learning methods where each batch is entered directly into the model once to update the parameters.

In some embodiments, for each task in the batch, the third gradient (updated from the first gradient) corresponding to the task is in each support sample image and the classification label corresponding to each support sample image. Determined on the basis of, the set of parameters for the anti-counterfeiting model of the face is determined on the basis of the third gradient. Multiple sets of parameters that belong to intermediate parameters and are determined by multiple tasks that do not accurately update the parameters of the face anti-counterfeiting model may be stored for the next update. In addition, the fourth gradient corresponding to the batch (updated from the second gradient) is a set of parameters determined for each of the multiple tasks in the batch, a query sample image corresponding to the multiple tasks, and a query sample image. It is determined based on the classification label corresponding to each of the query sample images. The parameters of the anti-counterfeiting model of the face are redetermined based on the fourth gradient corresponding to the batch. Update the face anti-counterfeiting model parameters using the face anti-counterfeiting model parameters determined during the second time.

が計算され、ここで、θは、顔の偽造防止モデルの現在のパラメータを表し、

は、タスクτ_iのためのサポートサンプル画像を使用して現在のパラメータθに関して計算された損失関数の勾配

を表す。タスクτ_iについて決定された顔の偽造防止モデルのパラメータは、

と表され、ここで、αは、第1のステップ値を表す。バッチ内の複数のタスクは、格納される複数の

(重みパラメータ行列であってよい)を決定する。 For example, for each task τ _i ∈ T in a batch, i is a positive integer representing the serial number of the task and T is the task set for the batch. For each task τ _i , the third gradient of the parameters of the face anti-counterfeiting model corresponding to the task

Is calculated, where θ represents the current parameters of the anti-counterfeiting model of the face.

Is the gradient of the loss function calculated for the current parameter θ using the support sample image for task τ _i .

Represents. The parameters of the face anti-counterfeiting model determined for task _τi are:

Here, α represents the first step value. Multiple tasks in a batch are stored in multiple

Determine (may be a weight parameter matrix).

と表され、これは、バッチ内の複数のタスクの損失がまず加算され、次いで勾配が計算されることを表している。損失および第4の勾配は、タスクτ_iのクエリサンプル画像と、対応する

に基づいて計算される。このバッチに対して更新される顔の偽造防止モデルのパラメータは、

である。βは、第2のステップ値である。

は、第4の勾配を計算するための中間変数であり、モデルのパラメータはバッチの開始時にθに基づいて各バッチについて更新されることが分かる。顔の偽造防止モデルは、クエリサンプル画像によって更新され、タスクについてのモデルの一般化能力を向上させ、サポートサンプル画像の過剰適合を回避できる。 In addition, for the batch, the fourth gradient corresponding to the batch,

This means that the losses of multiple tasks in the batch are added first, and then the gradient is calculated. The loss and the fourth gradient correspond to the query sample image of task _τi .

It is calculated based on. The face anti-counterfeiting model parameters updated for this batch are:

Is. β is the second step value.

Is an intermediate variable for calculating the fourth gradient, and it can be seen that the model parameters are updated for each batch based on θ at the start of the batch. The face anti-counterfeiting model can be updated with query sample images to improve the model's generalization ability for tasks and avoid overfitting of support sample images.

If the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label, and the category is the counterfeiting method. Or include at least one of the models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. It is a classification label. For the actual facial image, the preset values can be labeled when performing the labeling using the impersonation method. That is, if the face anti-counterfeiting model is a classification model, for each task in this batch, the third gradient corresponding to this task is the first classification label corresponding to each support sample image and each support sample image. Is determined based on. Alternatively, if the face anti-counterfeiting model is a regression model, for each task in the batch, the third gradient corresponding to this task is the second classification corresponding to each support sample image and each support sample image. Determined based on the label.

In step S208, it is determined whether the anti-counterfeiting model has reached the convergence condition. If YES, the pre-training of the anti-counterfeiting model is completed, and if NO, the process returns to step S206 for restart. ..

The convergence condition corresponds, for example, to minimizing the value of the loss function or reaching a preset value of the loss function. If the value of the loss function does not reach the convergence condition, return to step S206 to select another batch for training the face anti-counterfeiting model.

The steps described above belong to the pre-training process for the anti-counterfeiting model of the face. The number of supported sample images in each category of each task is very small and therefore belongs to a small sample set, but a large number of samples (including many sample images in existing categories and actual image categories) are pre-existing. Note that it is used during the entire training process. Also, the larger the number of samples, the better. The only difference is that many tasks and many small sample sets are split so that the model can "learn how to learn" during the pre-training process, increasing the accuracy of the model. be able to.

In step S210, a small sample set of forged facial images in a new category is obtained.

A small sample set of forged facial images in the new category can serve as support sample images in the task. The task contains one or more new categories, with each new category corresponding to a small sample set.

In step S212, the parameters of the pre-trained facial anti-counterfeiting model are updated with a small sample set of counterfeit facial images in a new category.

を決定するプロセスに類似している。例えば、タスクτ_jのサポートサンプル画像として1つまたは複数の新しいカテゴリの小さなサンプルセット内の偽造された顔画像の場合、このタスクτ_jに対応する顔の偽造防止モデルのパラメータの第3の勾配

が計算され、タスクτ_j対して決定された顔の偽造防止モデルのパラメータは、

と表され、ここで、αは第1のステップ値を表す。すなわち、更新された顔の偽造防止モデルのパラメータは、

である。クエリサンプル画像を使用した更新プロセスは、実行されない。 How to update the parameters of the pre-trained face anti-counterfeiting model with a small sample set of forged face images in the new category is by using the support sample images of task _τi in step S206. Anti-counterfeit model parameters

It is similar to the process of determining. For example, for a forged face image in a small sample set in one or more new categories as a support sample image for task τ _j , the third gradient of the parameters of the face anti-counterfeiting model corresponding to this task τ _j .

Is calculated and the parameters of the face anti-counterfeiting model determined for the task _τj are:

Here, α represents the first step value. That is, the parameters of the updated face anti-counterfeiting model are:

Is. The update process using the query sample image is not performed.

In step S214, the face image to be recognized is input to the face anti-counterfeiting model after the parameters are updated, and it is determined whether the face image to be recognized is a forged face image.

The solutions of the present disclosure can be applied not only to recognize forged facial images of new counterfeit methods, but also to recognize forged facial images of new models. For example, if you only need to recognize the forged face image of the new counterfeit method, the counterfeit face image and the actual face image of the existing counterfeit method will act as a training sample during the pre-training, and the training sample will Forged facial images of the same model and new forgery methods also belong to such models. In another example, if only the forged face image needs to be recognized, the forged face image and the actual face image of the existing model act as training samples during the pre-training and these of the existing model. The forged face image of is forged by the same forgery method, and the forged face image of the new forged model also belongs to such a forgery method. If the training sample contains forged facial images of both different counterfeit methods and different models, and the actual facial images of one or more models during pre-training, the new counterfeit facial images will be trained. May differ from any counterfeiting method or any model in the sample.

For example, for a forged face image of a new counterfeit method, in some embodiments, a small sample set of the forged face image of the new counterfeit method is obtained, and the parameters of the face anti-counterfeiting model are the parameters of the new counterfeit method. Updated with a small sample set of forged facial images, the parameters of the pre-trained facial anti-counterfeiting model are multiple sets of facial anti-counterfeiting model parameters, each determined based on multiple tasks. Determined on the basis of a set of parameters, each task includes a small sample set of forged facial images and a small sample set of actual facial images for each of one or more existing counterfeit methods. The face image to be recognized is input to the face anti-counterfeiting model after the parameters are updated, and it is determined whether the face image to be recognized is a forged face image.

For example, for a forged face image of a new model, in some embodiments a small sample set of the forged face image of the new model was obtained and the parameters of the anti-face anti-counterfeit model were forged in the new model. Updated with a small sample set of facial images, the pre-trained facial anti-counterfeiting model parameters are multiple sets of facial anti-counterfeiting model parameters, each determined based on multiple tasks. Determined based on a set of parameters, each task is a small sample set of each forged facial image of one or more existing models and each actual facial image of one or more existing models. A face image to be recognized, including a small sample set, is entered into the face anti-counterfeiting model after the parameters have been updated to determine if the face image to be recognized is a forged face image.

The pre-training process is similar to that of the embodiments described above. In situations where the training sample contains different forgery methods and forged facial images of different models, for example, if there are 3 known models and 3 known forgery methods, up to 9 different existing categories of forgery. There can be a facial image. In addition, if there are actual facial images of the three models, the generated task can contain up to 12 small sample sets.

In the embodiments described above, the pre-trained facial anti-counterfeiting model is a set of small samples, one or more existing anti-counterfeiting, consisting of forged facial images of one or more existing models. Trained by the method, or the actual facial images of multiple existing models, multiple groups of facial anti-counterfeiting model parameters are determined by multiple tasks, respectively, and then facial anti-counterfeiting model parameters are determined during the training process. Redetermined based on multiple sets of parameters during. This training process consists of two stages: the first stage of training a facial anti-counterfeiting model using multiple small sample sets of forged facial images in an existing category and real facial images in an existing category. And with the first stage, where the model learns the features of a small sample set, or the stage where it learns to distinguish between a fake face and a real face, as well as performing aggregation. The second step is to re-update the parameters of the face anti-counterfeiting model based on multiple training results, learning how the model learns a small sample set of features, or the forged face and the actual It includes a second stage, which is a stage for learning how to learn to distinguish from the face.

Therefore, a pre-trained face anti-counterfeiting model learns how to learn a small sample set or how to distinguish between a forged face and a real face, and proper initialization. Get the parameters. The face anti-counterfeiting model is updated when the forged face image of the new model belongs to the conditions of a small sample set. As an advantage of the fast learning ability previously learned, the face anti-counterfeiting model also performs fast convergence and fast learning of a new category of counterfeit facial features, making the face anti-counterfeiting model timely and effective. Updates can be performed, the forged face image of the new category can be recognized in a timely and effective manner, and the attack from the forged face of the new category can be effectively protected in a timely manner.

The disclosure also provides a forged face recognition device, described below with FIG.

FIG. 3 shows a schematic structural diagram of a forged face recognition device according to some embodiments of the present disclosure. As shown in FIG. 3, the device 30 of this embodiment includes a small sample acquisition module 302, a model adjustment module 304, and a face recognition module 306.

The small sample acquisition module 302 is configured to acquire a small sample set of forged facial images in a new category.

The model adjustment module 304 is configured to update the parameters of a pre-trained facial anti-counterfeiting model with a small sample set of pre-trained facial anti-counterfeiting models in a new category. Parameters are determined based on multiple sets of parameters, multiple sets of pre-trained parameters are determined respectively based on multiple tasks, and each task is each in one or more existing categories. Includes a small sample set of forged facial images and a small sample set of actual facial images for each of one or more existing categories.

In some embodiments, the model adjustment module 304 is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is configured to update the parameters of the anti-counterfeiting model and the anti-counterfeiting model is a classification model, the classification label will now label the category of forged face images corresponding to the classification label. The first classified label constructed, the category comprising at least one of the counterfeiting methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, the classification label is a second classification label configured to label the forged face image corresponding to the classification label as a positive sample or a negative sample. Is.

In some embodiments, the model adjustment module 304 turns the forged face image in a small sample set of forged face images in a new category into a face anti-counterfeiting model as a support sample image for one task. Enter and determine the gradient of the face anti-counterfeiting model parameters based on the output value of the face anti-counterfeiting model for each of the forged facial images and the classification label corresponding to each of the forged facial images. , Based on the gradient of the face anti-counterfeiting model parameters, it is configured to update the face anti-counterfeiting model parameters.

In some embodiments, if the anti-counterfeit model of the face is a classification model, the output value of the anti-counterfeit model for each of the counterfeit face images is the probability that each of the counterfeit face images belongs to a different category. And the model adjustment module 304 is an anti-counterfeiting model of the face based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is configured to determine the first gradient of the parameters of the counterfeited face. Each of the images is the probability of belonging to the positive sample, and the model adjustment module 304 has the probability that each of the forged face images belongs to the positive sample and the second classification label corresponding to each of the forged face images. Based on, it is configured to determine the second gradient of the parameters of the anti-counterfeiting model.

The face recognition module 306 is configured to input a face image to be recognized in the face anti-counterfeiting model after updating the parameters and determine whether the face image to be recognized is a forged face image. To.

In some embodiments, the device 30 further comprises a pre-training module 308. The pre-training module 308 is to divide the forged face image of the existing category and the actual face image of the existing category into a plurality of tasks, and each task has a plurality of methods, and each method has a plurality of methods. Each task has a first preset number of existing categories, each method supports, with multiple real facial images of one model or multiple forged facial images of existing categories. A second preset number of real or forged face images as a small sample set consisting of sample images, and a third preset number of real face images as a query sample image. Or with a forged facial image, and to select a fourth preset number of tasks to form a batch each time, with multiple sets of facial anti-counterfeiting model parameters. Determined based on multiple tasks in the batch, and the parameters of the face anti-counterfeiting model are redetermined based on multiple sets of parameters, and until the face anti-counterfeiting model reaches convergence conditions. It is configured to complete the pre-training of the anti-counterfeiting model.

In some embodiments, the pre-training module 308 determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and the first for each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on a gradient of 3 and a set of parameters determined by each of the multiple tasks in the batch, query sample images corresponding to the multiple tasks, and Determine the fourth gradient corresponding to the batch based on the classification label corresponding to each of the query sample images, and redetermine the parameters of the anti-counterfeit model based on the fourth gradient corresponding to the batch. It is configured to do things.

In some embodiments, when the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label. Yes, the category contains at least one of the counterfeit methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. Is.

In some embodiments, the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method.

In some embodiments, different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises a photo, video, or mask, and the model is a depth model, near. Actual facial images of different tasks, including infrared models, or red-green-blue (RGB) models, have the same model or different models, and a new category of counterfeit facial images is within at least one of the counterfeit methods or models. Unlike existing images in, existing images include forged facial images in existing categories and actual facial images in existing categories.

The forged face recognition device of the embodiments of the present disclosure is implemented by various computing devices or computer systems and is described below with FIGS. 4 and 5.

FIG. 4 shows a schematic structural diagram of a forged face recognition device according to some embodiments of the present disclosure. As shown in FIG. 4, the device 40 of the embodiment comprises a memory 410 and a processor 420 coupled to the memory 410, wherein the processor 420 is any of the present disclosures based on instructions stored in the memory 410. It is configured to implement the forged face recognition method described in the embodiment.

The memory 410 may include, for example, a system memory, a predetermined non-volatile storage medium, and the like. System memory stores, for example, operating systems, applications, boot loaders, databases, and other programs.

FIG. 5 shows a schematic structural diagram of a forged face recognition device according to another embodiment of the present disclosure. As shown in FIG. 5, the device 50 of this embodiment includes a memory 510 and a processor 520, which are similar to the memory 410 and the processor 420. In addition, there are I / O interface 530, network interface 540, storage interface 550, etc. These interfaces 530, 540, 550, as well as the memory 510 and processor 520 in between, may be connected, for example, via bus 560. The I / O interface 530 provides a connectivity interface for input and output devices such as displays, mice, keyboards, and touch classifiers. The network interface 540 provides a connection interface for various network devices. For example, it can be connected to a database server or cloud storage server. The storage interface 550 provides a connectivity interface for external storage devices such as SD cards or USB flash disks.

It will be apparent to those skilled in the art that embodiments of the present disclosure may be provided as methods, systems, or computer program products. Accordingly, the present disclosure may take the form of a fully hardware embodiment, a fully software embodiment, or a combination of software and hardware aspects. Further, the present disclosure is embodied in one or more computer-usable non-temporary storage media including, but not limited to, disk memory, CD-ROM, optical memory, etc., including computer-usable program code. It can take the form of a computer program product.

The present disclosure is described with flow charts and / or block diagrams of the method, devices (systems), and computer program products according to embodiments of the present disclosure. It can be seen that the flow chart and / or block diagram steps and / or blocks, as well as the flow chart and / or block diagram steps and / or block combinations, are implemented by computer program instructions. These computer program instructions are provided to the processor of a general purpose computer, dedicated computer, embedded processor, or other programmable data processing device to create a machine and are executed by the computer or the processor of another programmable data processing device. The instructions given give rise to a device for achieving the function specified in one or more steps in the flowchart and / or one or more blocks in the block diagram.

These computer program instructions may also be stored in computer-readable memory that can guide the computer or other programmable data processing device to operate in a particular way, and as a result, are stored in computer-readable memory. Instructions produce a product that includes an instruction device. The instruction device implements the function specified in one or more steps in the flowchart, or in one or more blocks in the block diagram.

These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operating steps is performed on the computer or other programmable device and is performed on the computer. An instruction that produces a computer-implemented process, such as an instruction, that is executed on a computer or other programmable device is specified in one or more steps in a flowchart and / or one or more blocks in a block view. It provides the steps to realize the function.

The above description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. within the principles of this disclosure are included within the scope of this disclosure.

302 Small sample acquisition module
304 Model Adjustment Module
306 Face recognition module
308 Pre-training module
410 memory
420 processor
510 memory
520 processor
530 I / O interface
540 network interface
550 storage interface
560 bus

Cross-reference to related applications This application claims priority in Chinese Patent Application No. 201910387689.0 filed May 10, 2019, which is incorporated herein by reference in its entirety.

The present disclosure relates to the field of computer technology, in particular to counterfeit face recognition methods, devices, and non-temporary computer-readable storage media.

Face anti-counterfeiting recognition refers to the effective identification of counterfeit facial features. For example, a facial recognition system may produce several forged faces (videos, photos, masks, etc.) to attack the facial recognition system. Face anti-counterfeiting recognition aims to identify counterfeit facial features. Generally, in a face recognition system, a user needs to pass the detection of recognition for preventing forgery of a face, which is also called biological detection.

A method of recognizing counterfeit faces known to the inventor is to perform deep learning training on the classification model based on a large number of counterfeit facial images in known categories. The trained classification model may recognize that the forged face belongs to the same category as the forged face image for training.

Some embodiments of the present disclosure provide forged face recognition methods. The method is to get a small sample set of forged facial images in a new category, where the new category is different from the existing category used when the facial anti-counterfeiting model was pre-trained. , A step to update the parameters of a pre-trained facial anti-counterfeiting model with a small sample set of pre-trained facial anti-counterfeiting models in a new category, with multiple pre-trained facial anti-counterfeiting model parameters. Determined based on multiple sets of pre-trained parameters, each determined on the basis of each task, each task with a small sample set of each forged facial image in one or more existing categories. Recognized by entering a step and a face image that should be recognized by the face anti-counterfeiting model after parameter updates, including a small sample set of each actual face image in one or more existing categories. It includes a step of determining whether the face image to be forged is a forged face image.

In some embodiments, the method is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. If the anti-counterfeit model contains a step to update the parameters of the anti-counterfeit model and the anti-counterfeit model is a classification model, the classification label is configured to label the category of forged face images corresponding to the classification label. If the classification label is 1 and the category contains at least one of the counterfeiting methods or models, or the anti-counterfeiting model of the face is a regression model, the classification label is a forged face image corresponding to the classification label. A second classification label configured to be labeled as a positive sample or a negative sample.

In some embodiments, a face anti-counterfeiting model is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. The step to update the parameters of is to enter the forged face image in a small sample set of forged face images in a new category into the face anti-counterfeiting model as a support sample image for one task. Steps to determine the gradient of facial anti-counterfeiting model parameters based on the output value of the face anti-counterfeiting model for each of the forged facial images and the classification label corresponding to each of the forged facial images. Includes a step to update the face anti-counterfeiting model parameters based on the facial anti-counterfeiting model parameter gradient.

In some embodiments, if the anti-counterfeit model of the face is a classification model, the output value of the anti-counterfeit model for each of the counterfeit face images is the probability that each of the counterfeit face images belongs to a different category. The gradient of the parameters of the anti-counterfeiting model of the face is determined based on the output value of the anti-counterfeiting model of the face for each of the counterfeited face images and the classification label corresponding to each of the counterfeited face images. The first of the parameters of the anti-counterfeiting model of the face is based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is a regression model, or if the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is a positive sample for each of the counterfeited face images. The gradient of the parameters of the anti-counterfeit model based on the output value of the anti-counterfeit model of the face for each of the counterfeit face images and the classification label corresponding to each of the counterfeit face images. The steps to determine the parameters of the anti-counterfeiting model of the face are based on the probability that each of the forged face images belongs to a positive sample and the second classification label that corresponds to each of the forged face images. Provide a step to determine the second gradient.

In some embodiments, the recognition method is further a step of dividing the forged face image of the existing category and the actual face image of the existing category into a plurality of tasks, and each task has a plurality of methods. And each method has multiple real facial images of one model or multiple counterfeit facial images of existing categories, and each task has a first preset number of existing categories. Each method has a second preset number of real or forged face images as a small sample set consisting of support sample images, and a third preset as a query sample image. A step with a number of real or forged facial images, and a step of selecting a fourth preset number of tasks to form a batch each time, in advance of the face anti-counterfeiting model. Multiple sets of trained parameters are determined respectively based on multiple tasks in the batch, and parameters of the face anti-counterfeiting model are redetermined based on multiple sets of pre-trained parameters. It further comprises a step and a step to complete the pre-training of the face anti-counterfeiting model until the face anti-counterfeiting model reaches the convergence condition.

In some embodiments, a plurality of sets of pre- trained parameters for the anti-counterfeiting model are each determined based on multiple tasks in a batch, and the parameters for the anti-counterfeiting model are pre-trained. Redetermination based on multiple sets of parameters determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on a third gradient of about, and a set of parameters determined by each of the multiple tasks in the batch, a query sample corresponding to multiple tasks. Determine the fourth gradient corresponding to the batch based on the classification label corresponding to each of the image and the query sample image, and the parameters of the anti-counterfeiting model based on the fourth gradient corresponding to the batch. Includes deciding.

In some embodiments, when the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label. Yes, the category contains at least one of the counterfeit methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. Is.

In some embodiments, the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method.

In some embodiments, different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises multiple photographs, videos, or masks, and the model is a depth model. The actual facial images of different tasks, including multiple near-infrared models, or red-green-blue (RGB) models, have the same model or different models, and new categories of counterfeit facial images are of the counterfeit method or model. Unlike existing images in at least one, the existing image includes a forged facial image in an existing category and an actual facial image in an existing category.

In another embodiment of the present disclosure, a forged face recognition device is provided. The device is a small sample acquisition module configured to acquire a small sample set of forged facial images in a new category, which was used when the new category was pre-trained for facial anti-counterfeiting models. A model tuning module configured to update the parameters of a pre-trained facial anti-counterfeiting model with a small sample acquisition module that differs from the existing category and a small sample set of forged facial images in the new category. And the parameters of the pre-trained facial anti-counterfeiting model are determined based on multiple sets of pre-trained parameters, each determined on the basis of multiple tasks, with each task being one or more. Model tuning modules and parameter updates, including a small sample set of forged facial images for each of multiple existing categories and a small sample set of actual facial images for each of one or more existing categories. It includes a face recognition module configured to later input a face image to be recognized into a face anti-counterfeiting model and determine whether the face image to be recognized is a forged face image.

In some embodiments, the model adjustment module is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is configured to update the parameters and the anti-counterfeiting model of the face is a classification model, the classification label is configured to label the category of forged face images corresponding to the classification label. If it is the first classification label given and the category contains at least one of the counterfeiting methods or models, or if the anti-counterfeiting model of the face is a regression model, the classification label is counterfeited corresponding to the classification label. A second classification label configured to label the facial image as a positive sample or a negative sample.

In some embodiments, the model adjustment module inputs the forged face image in a small sample set of forged face images in a new category into the face anti-counterfeiting model as a support sample image for one task. Then, the gradient of the parameters of the anti-counterfeiting model of the face is determined based on the output value of the anti-counterfeiting model of the face for each of the forged facial images and the classification label corresponding to each of the forged facial images. It is configured to update the parameters of the face anti-counterfeiting model based on the gradient of the face anti-counterfeiting model parameters.

In some embodiments, when the anti-counterfeiting model of the face is a classification model, the output value of the anti-counterfeiting model for each of the counterfeit face images is the probability that each of the counterfeit face images belongs to a different category. And the model adjustment module is based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images of the anti-counterfeiting model of the face. It is configured to determine the first gradient of the parameter.

Alternatively, if the anti-counterfeiting model of the face is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a positive sample, and the model adjustment. A second of the parameters of the anti-counterfeiting model of the face, based on the probability that each of the forged facial images belongs to a positive sample and the second classification label that corresponds to each of the forged facial images. It is configured to determine the gradient.

In some embodiments, the recognition device is to divide the forged face image of an existing category from the actual face image of an existing category into a plurality of tasks, each task having a plurality of methods. Each method has multiple real facial images of one model or multiple forged facial images of existing categories, and each task has a first preset number of existing categories. Each method has a second preset number of real or forged face images as a small sample set consisting of support sample images, and a third preset number as query sample images. To have a real or forged facial image of the face, and to select a fourth preset number of tasks to form a batch each time, in advance of the facial anti-counterfeiting model. Multiple sets of trained parameters are each determined based on multiple tasks in the batch, and parameters of the face anti-counterfeiting model are redetermined based on multiple sets of pre-trained parameters. And further includes a pre-training module configured to complete the pre-training of the face anti-counterfeiting model until the face anti-counterfeiting model reaches the convergence condition.

In some embodiments, the pre-training module determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and a third for each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on the gradient of, and a set of parameters determined by each of the multiple tasks in the batch, query sample images corresponding to multiple tasks, and queries. Determining the fourth gradient for the batch based on the classification label corresponding to each of the sample images, and redetermining the parameters of the anti-counterfeiting model based on the fourth gradient corresponding to the batch. Is configured to do.

In some embodiments, if the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label. Yes, if the category contains at least one of the counterfeiting methods or models, or if the anti-counterfeiting model of the face is a regression model, then each of the classification labels is a positive sample or a negative sample of the support sample image corresponding to the classification label. A second classification label configured to be labeled as a sample.

In some embodiments, the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method.

In some embodiments, different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises multiple photographs, videos, or masks, and the model is a depth model. The actual facial images of different tasks, including multiple near-infrared models, or red-green-blue (RGB) models, have the same model or different models, and new categories of counterfeit facial images are of the counterfeit method or model. Unlike existing images in at least one, the existing image includes a forged facial image in an existing category and an actual facial image in an existing category.

Yet another embodiment of the present disclosure provides a forged face recognition device. The device comprises a memory and a processor coupled to the memory, the processor configured to perform the forged face recognition method according to any of the above embodiments based on the instructions stored in the memory. Will be done.

Yet another embodiment of the present disclosure provides a non-temporary computer-readable storage medium. The storage medium stores a computer program that implements the method described in any of the above embodiments or the training method of the anti-counterfeiting model when executed by the processor.

Yet another embodiment of the present disclosure provides a training method for a face anti-counterfeiting model. The method of training the face anti-counterfeiting model is the step of dividing the forged face image of the existing category and the actual face image of the existing category into multiple tasks, and each task is one or more existing ones. Faces using steps and multiple tasks, including a small sample set of forged facial images for each category and a small sample set of actual facial images for each of one or more existing categories. Pre-trained based on the steps of pre-training the anti-counterfeit model to determine each of the multiple sets of pre-trained parameters for the face anti-counterfeit model and the multiple sets of pre-trained parameters. A step to determine the parameters of the face anti-counterfeiting model and a step to get a small sample set of forged facial images in a new category, when the new category is pre-trained for the face anti-counterfeiting model. It comprises steps that differ from the existing categories used and steps that update the parameters of the pre-trained facial anti-counterfeiting model with a small sample set of forged facial images in the new category.

In some embodiments, each task comprises multiple methods, each method having multiple real facial images of one model or multiple counterfeit facial images of an existing category, and each task has multiple methods. A first preset number of existing categories, each method as a small sample set composed of support sample images, a second preset number of real face images or forged faces. Having an image, having a third preset number of real or forged facial images as query sample images, and pre-training a facial anti-counterfeiting model using multiple tasks, Determine the parameters of the pre-trained face anti-counterfeiting model based on the steps to determine each of the multiple sets of pre-trained face anti-counterfeiting models and the multiple sets of pre-trained parameters. The step to do is to select a fourth preset number of tasks to form a batch each time, with multiple sets of pre-trained parameters for the face anti-counterfeiting model, multiple batches. Each step is determined based on the task, and the parameters of the face anti-counterfeiting model are redetermined based on multiple sets of pre-trained parameters, until the steps and the face anti-counterfeiting model reach convergence conditions. It includes steps to complete pre-training of the face anti-counterfeiting model.

In some embodiments, a plurality of sets of pre-trained parameters for the anti-counterfeiting model are each determined based on multiple tasks in a batch, and the parameters for the anti-counterfeiting model are pre-trained. Redetermination based on multiple sets of parameters determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on a third gradient against, a set of parameters determined by each of the multiple tasks in the batch, and a query sample corresponding to the multiple tasks. Determine the fourth gradient corresponding to the batch based on the image and the classification label corresponding to each of the query sample images, and the parameters of the anti-counterfeiting model based on the fourth gradient corresponding to the batch. Be prepared to make a decision again.

Yet another embodiment of the present disclosure provides a training device for a face anti-counterfeiting model. The device comprises a memory and a processor coupled to the memory, and the processor is configured to perform a training method of a face anti-counterfeiting model according to any of the above embodiments based on instructions stored in the memory. Will be done.

In yet another embodiment of the present disclosure, a computer program is provided. The computer program, upon execution by the processor, comprises an instruction to cause the processor to execute a forged face recognition method according to a training method of a face anti-counterfeiting model or a forged face recognition method according to any of the above embodiments.

Other features and advantages of the present disclosure will become apparent from the subsequent detailed description of the exemplary embodiments of the present disclosure with reference to the accompanying drawings.

The accompanying drawings set forth herein for the purpose of providing further understanding of the present disclosure form part of this application. The embodiments and illustrations presented in this disclosure are intended to illustrate this disclosure, but do not constitute an inappropriate definition of this disclosure.

It is a schematic flowchart of the method of recognizing a forged face according to some embodiments of this disclosure. It is a schematic flowchart of the forged face recognition method by another embodiment of this disclosure. FIG. 3 is a schematic structural diagram of a forged face recognition device according to some embodiments of the present disclosure. FIG. 3 is a schematic structural diagram of a forged face recognition device according to some other embodiment of the present disclosure. FIG. 3 is a schematic diagram of the structure of a forged face recognition device according to still another embodiment of the present disclosure.

The technical solutions of the embodiments of the present disclosure will be described expressly and fully herein, along with the accompanying drawings of the embodiments of the present disclosure. Obviously, the embodiments described are not all of the embodiments, but only some of the embodiments of the present disclosure. The following description of at least one embodiment of the description, which is merely exemplary in practice, shall by no means serve as any limitation with respect to the present disclosure and its application or use. All other embodiments obtained by one of ordinary skill in the art on the premise that the efforts of the invention are not included in accordance with the embodiments of the present disclosure shall be included in the scope of protection of the present disclosure.

The inventor found that in existing counterfeit face recognition methods, it was necessary to use a large number of samples to train the model so that the trained model could accurately recognize the counterfeit face. There is. With the advent of new categories of forged facial images, it is not possible to achieve reopened model training if the sample size of the forged facial images is very small, and as a result, the model is forged in the new category. The face is not recognized in a timely manner, making it impossible to effectively prevent a new category of counterfeit face attacks in a timely manner.

One technical challenge to be resolved by this disclosure is to achieve timely and effective recognition of new categories of forged facial images, and for faces when a small number of new categories of forged facial images appear. This is a way to improve the security of the recognition system.

From a new category of forged faces when the model cannot timely and effectively recognize the new category of forged faces, new categories of forged face images appear, and very few samples. The solution is provided to solve the problem of not being able to prevent attacks in a timely and effective manner. A method for recognizing a forged face according to some embodiments of the present disclosure is described below together with FIG.

FIG. 1 shows a schematic flow chart of a forged face recognition method according to some embodiments of the present disclosure. As shown in FIG. 1, the method of this embodiment comprises steps S102-S106.

In step S102, a small sample set of forged facial images in a new category is obtained.

In the present disclosure, the number of images in a small sample set is less than the threshold. This threshold acknowledges that the anti-counterfeiting model is updated with a small sample set according to deep learning training methods, making it impossible for the model to converge. The threshold is, for example, 10 to 30, but is not limited to the examples listed. The new category of forged face images belongs to a different category than the forged face image category used when the face anti-counterfeiting model was pre-trained, and the face anti-counterfeiting model is pre-trained. The forged face category used when used can be referred to as an existing category.

Disguised facial images in different existing categories differ in at least one of the counterfeiting methods or models. Counterfeiting methods include, for example, photographs, videos, and masks, and models include, for example, depth (Depth) models, near-infrared (NIR) models, and red-green-blue (RGB) models. For example, a photo forgery method and a forged face image with a depth model belong to a different category than a photo forgery method and an RGB model forged face image, and a photo forgery method and a forged face image with a depth model It also belongs to a different category from video forgery methods and forged facial images with RGB models. Multi-model (multi-model) images refer to images collected by different technical means for describing different attributes of the same object. For example, for an object, a color image captured by a camera and a depth image scanned by a laser radar are called two model images of the object. Forged facial images in the new category differ from at least one existing image of the forged method or model, and existing images include forged facial images and actual facial images in the existing category.

In step S104, the parameters of the pre-trained facial anti-counterfeiting model are updated with a small sample set of forged facial images in a new category. The parameters of the pre-trained face anti-counterfeiting model are multiple sets of pre-trained parameters of the face anti-counterfeiting model, each determined based on multitasking. Determined on the basis of a set, each task includes a small sample set of forged facial images for each of one or more existing categories and a small sample set of actual facial images.

In some embodiments, the parameters of the face anti-counterfeiting model are multiple forged face images in a small sample set of forged face images in a new category, and a classification label corresponding to each of the forged face images. And will be updated based on. The anti-counterfeiting model of the face may be a classification model or a regression model. If the anti-counterfeiting model of the face is a classification model, each classification label is the first classification label configured to label the category of forged face images corresponding to the classification label, and the category is counterfeit. Includes at least one of the methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each classification label is configured to label the forged face image corresponding to the classification label as a positive or negative sample. It is a classification label of 2. The first classification label may label at least one of the counterfeit methods or models of the counterfeit facial image. If it is necessary to learn different forgery methods and image features of different models, both the forgery method and the model are labeled.

Each of the counterfeit facial images in a small sample set of counterfeit facial images in the new category is entered into a pre-trained anti-counterfeiting model of the face, and the gradient of the parameters of the anti-counterfeiting model of the face is the output value, The parameters of the pre-trained anti-counterfeiting model, determined based on the classification label corresponding to each of the forged facial images and the preset loss function, are updated based on the gradient. If the face anti-counterfeiting model is a classification model, the preset loss function may be an MSE (Mean Squared Error) loss function. If the face anti-counterfeiting model is a regression model, the preset loss function may be a cross entropy loss function, but these are not limited to the examples taken. Methods of updating the parameters of a pre-trained anti-counterfeiting model based on the gradient are not detailed herein, but existing gradient descent methods may be used as desired.

In some embodiments, multiple forged facial images in a small sample set of forged facial images in a new category are entered into the facial anti-counterfeiting model as support sample images for one task and of the face. The gradient of the anti-counterfeiting model parameters is determined based on the output value of the face anti-counterfeiting model for each of the forged face images and the classification label corresponding to each of the forged face images to prevent face anti-counterfeiting. The model parameters are updated based on the gradient of the face anti-counterfeiting model parameters.

The task can be built on a small sample set of forged facial images in a new category. The forged face image in a small sample set of forged face images in the new category may be a support sample image for the task. This is similar to the format in which tasks are set by the pre-training method for forged facial images, which will be described later. The constructed task contains one or more different new categories, each of which has a corresponding small sample set. The task also has a small sample set of real facial images and the like.

In some embodiments, the face anti-counterfeiting model is a classification model, and the parameters of the face anti-counterfeiting model are each of the forged face images in a small sample set of forged face images in a new category. Updated based on a first classification label corresponding to each of the forged facial images. For example, the first classification label for the actual face image by the depth model is 0, the first classification label for the forged mask method and the forged face image by the depth model is 1, and the first in the new category. The classification label is 2.

The forged facial image and the actual facial image can be labeled with a first classification label and a second classification label depending on the application scenario. For example, during the process of updating the parameters of the face anti-counterfeiting model, the first classification label may be used to calculate the value of the loss function, the value of the gradient, etc. to improve the accuracy of the model. Can be done. When the face anti-counterfeiting model classifies images and outputs the results, it only needs to distinguish the images into two categories: real face images and forged face images. During the testing process after pre-training the anti-counterfeiting model of the image, a second classification label is used to recognize whether the image is forged or real, so that the anti-counterfeiting model of the face You can test the accuracy. For example, the second classification label representing a forged face image is 0 and the second classification label representing an actual face image is 1.

Further, when the face anti-counterfeiting model is a classification model, the output value of the face anti-counterfeiting model for each of the forged face images is the probability that each of the forged face images belongs to a different category, and the face. The first gradient of the anti-counterfeiting model parameters is based on the probability that each of the forged facial images belongs to a different category and the first classification label that corresponds to each of the forged facial images. To.

In some embodiments, the face anti-counterfeiting model may be a regression model, and the parameters of the face anti-counterfeiting model are each of the forged face images in a small sample set of forged face images in a new category. And a second classification label corresponding to each of the forged facial images. The regression model can finally output a value for the input image and set a threshold corresponding to the value. If the value is higher than the threshold, it means that the image is a real face image, and if the value is lower than the threshold, it means that the image is a forged face image. Therefore, the regression model may apply a second classification label for the image.

Further, when the anti-counterfeiting model of the image is a regression model, the output value of the anti-counterfeiting model of the face for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a positive sample. The gradient of 2 is determined based on the probability that each of the forged facial images belongs to a positive sample and the second classification label that corresponds to each of the forged facial images, and is a parameter of the anti-counterfeiting model of the face. Is updated based on the second gradient.

The parameters of the pre-trained anti-counterfeiting model are trained in the following way: multiple sets of pre-trained anti-counterfeiting models of the face are determined based on multiple tasks, respectively. The parameters of the anti-counterfeiting model of the face are then redetermined based on multiple sets of pre-trained parameters. Each task includes a small sample set of each forged facial image in one or more existing categories and a small sample set of actual facial images in each of one or more existing categories. The actual facial image is also divided into actual facial images of different models. Each task may include a small sample set of each actual facial image of one or more models. Actual facial images in different tasks have the same model or different models. That is, a small sample set composed of a forged face image of an existing category and an actual face image of an existing category is assigned to a plurality of tasks. A different small sample set should be used to train the anti-counterfeiting model, and after training the parameters of the anti-counterfeiting model should be redetermined based on multiple training results. The process of training a face anti-counterfeiting model with different small sample sets is the process of model learning, and the process of redetermining parameters based on multiple training results is the process of a "learning how to learn" model. ..

After the two processes described above, the model not only learns the features of the small sample set, but also learns how to learn the features of the small sample set, or distinguishes between a forged face and a real face. Rather than simply learning how, or simply learning forged clues, the anti-counterfeiting model of the face ultimately "how to learn to distinguish between a disguised face and a real face" and " Learn how to learn to find forged clues. Therefore, if the new category of counterfeit facial images belong to a small sample set, the face anti-counterfeiting model also performs fast convergence and fast learning of the new category of counterfeit facial features, effectively and timely. Defend Lee.

In step S106, the face image to be recognized is input to the face anti-counterfeiting model after the parameters are updated, and it is determined whether the face image to be recognized is a forged face image.

The facial image to be recognized may be a forged facial image that belongs to a new category or an existing category that can be recognized by the anti-counterfeiting model of the face.

In the method of the above embodiment, the parameters of the pre-trained face anti-counterfeiting model are updated directly with a small sample set of forged face images in a new category, resulting in the updated face forgery. The prevention model can recognize various forged facial images in existing and new categories. This is a pre-trained facial anti-counterfeiting model that is trained by a small sample set of forged facial images in the existing category and a small sample set of real facial images in advance of the facial anti-counterfeiting model. Multiple sets of trained parameters are determined by multiple tasks each, and then the parameters of the face anti-counterfeiting model are redetermined based on the multiple sets of pre- trained parameters during the training process. Because it is done. This training process consists of two stages: the first stage of training a facial anti-counterfeiting model using multiple small sample sets of forged facial images in an existing category and real facial images in an existing category. And with the first stage, where the model learns the features of a small sample set, or the stage where it learns to distinguish between a fake face and a real face, as well as performing aggregation. The second step is to re-update the parameters of the face anti-counterfeiting model based on multiple training results, learning how the model learns a small sample set of features, or the forged face and the actual It includes a second stage, which is a stage for learning how to learn to distinguish from the face.

Therefore, after learning multiple tasks with a small sample set of forged face images in an existing category and a real face image in an existing category, the face anti-counterfeiting model learns how to learn a small sample set. Learn how to learn to distinguish between a forged face and a real face, or learn to find forged clues, and get the appropriate initialization parameters. The anti-counterfeiting model of the face is updated if the new category of counterfeit facial images is related to the conditions of a small sample set. As an advantage of the fast learning ability previously learned, the face anti-counterfeiting model also performs fast convergence and fast learning of a new category of counterfeit facial features, making the face anti-counterfeiting model timely and effective. Updates can be performed, the forged face image of the new category can be recognized in a timely and effective manner, and the attack from the forged face of the new category can be effectively protected in a timely manner.

According to another embodiment of the present disclosure, a forged face recognition method will be described below together with FIG. This embodiment includes a pre-training method and a method of operation of a face anti-counterfeiting model.

FIG. 2 shows a schematic flow chart of a forged face recognition method according to another embodiment of the present disclosure. As shown in FIG. 2, the method of the present embodiment includes steps S202 to S214.

In step S202, the face anti-counterfeiting model is first trained with a forged face image of an existing category and an actual face image of an existing category according to a deep learning training method, and the face before pre-training. The initial parameters of the anti-counterfeit model are determined.

First, the face anti-counterfeiting model is trained by using existing deep learning training methods, then proceed to the subsequent pre-training process, so that the two post-training facial anti-counterfeiting models are suitable. Provides distinction and generalization. Deep learning training methods allow facial anti-counterfeiting models to more accurately distinguish images, and pre-training methods of the present disclosure allow models to generalize different categories of forged facial images. The post-training facial anti-counterfeiting model not only provides quick adaptability to new tasks, but can also provide appropriate distinctions for known or similar tasks.

In step S204, the forged face image of the existing category and the actual face image of the existing category are divided into a plurality of tasks. Each task has multiple methods, each method has multiple real facial images of one model or multiple counterfeit facial images of existing categories, and each method consists of support sample images. A second preset number of real or forged face images as a small set of samples, and a third preset number of real or forged face images as a query sample image. Has a face image. Each task has a first preset number of existing categories. Each existing category represents a counterfeit method and / or model. Face anti-counterfeiting models are simply configured to recognize counterfeit facial images using different counterfeiting methods, existing categories are simply configured to represent counterfeiting methods, and all counterfeit facial images and The actual face image belongs to the same model. If the face anti-counterfeiting model is configured to recognize different counterfeiting methods and counterfeit facial images of different models, each existing category represents a counterfeiting method or model.

The task has multiple methods. Each method includes a second preset number of support sample images and a third preset number of query sample images. The support sample images for each method form a small sample set, and the support sample images for each method in each task are less than the threshold, eg 10-30. During the pre-training process, the number of images in the small sample set formed by the support sample images may be the same as or different from the number of images in the small sample set of counterfeit facial images in the new category. If the number of images in the small sample set formed by the support sample images is less than or equal to the number of images in the small sample set of forged face images in the new category, the face anti-counterfeiting model is a new category in the pre-training process. Subsequently updated based on a small sample set of, the face anti-counterfeiting model can present a better recognition effect. This is because the face anti-counterfeiting model learns how to learn a smaller number of samples during the pre-training process, and there are numerous images in the new category at the time of the update, so the model parameters are more updated. Can be achieved well.

Each time, a first preset number of existing categories is randomly selected, followed by a second preset number of support sample images and a third preset number of query sample images in the existing category. Can be randomly selected from selected forged facial images. The actual facial images of one or more models were randomly selected, followed by a second preset number of support sample images and a third preset number of query sample images. Randomly selected from the face image of the to generate a task. Images of different tasks cannot be exactly the same.

In step S206, a fourth preset number of tasks are selected to form a batch, and multiple sets of pre-trained parameters of the face anti-counterfeiting model are based on multiple tasks in the batch. Each is determined and the parameters of the anti-counterfeiting model are redetermined based on multiple sets of pre-trained parameters.

A fourth preset number of tasks are randomly selected to form a batch. For example, 5 tasks are selected as a batch. For each batch, the face anti-counterfeiting model parameters can be updated twice. The first update is to update the parameters of the face anti-counterfeiting model with support sample images in each task. The second update is to update the parameters of the face anti-counterfeiting model with the query sample image in each task. That is, the update is composed of the above-mentioned learning process and the process of "learning the learning method". This is different from existing deep learning methods where each batch is entered directly into the model once to update the parameters.

In some embodiments, for each task in the batch, the third gradient (updated from the first gradient) corresponding to the task is in each support sample image and the classification label corresponding to each support sample image. Determined on the basis of, the set of parameters for the anti-counterfeiting model of the face is determined on the basis of the third gradient. Multiple sets of pre-trained parameters that belong to intermediate parameters and are determined by multiple tasks that do not accurately update the parameters of the face anti-counterfeiting model may be stored for the next update. In addition, the fourth gradient corresponding to the batch (updated from the second gradient) is a set of parameters determined for each of the multiple tasks in the batch, a query sample image corresponding to the multiple tasks, and a query sample image. It is determined based on the classification label corresponding to each of the query sample images. The parameters of the anti-counterfeiting model of the face are redetermined based on the fourth gradient corresponding to the batch. Update the face anti-counterfeiting model parameters using the face anti-counterfeiting model parameters determined during the second time.

が計算され、ここで、θは、顔の偽造防止モデルの現在のパラメータを表し、

は、タスクτ_iのためのサポートサンプル画像を使用して現在のパラメータθに関して計算された損失関数の勾配

を表す。タスクτ_iについて決定された顔の偽造防止モデルのパラメータは、

と表され、ここで、αは、第1のステップ値を表す。バッチ内の複数のタスクは、格納される複数の

(重みパラメータ行列であってよい)を決定する。 For example, for each task τ _i ∈ T in a batch, i is a positive integer representing the serial number of the task and T is the task set for the batch. For each task τ _i , the third gradient of the parameters of the face anti-counterfeiting model corresponding to the task

Is calculated, where θ represents the current parameters of the anti-counterfeiting model of the face.

Is the gradient of the loss function calculated for the current parameter θ using the support sample image for task τ _i .

Represents. The parameters of the face anti-counterfeiting model determined for task _τi are:

Here, α represents the first step value. Multiple tasks in a batch are stored in multiple

Determine (may be a weight parameter matrix).

と表され、これは、バッチ内の複数のタスクの損失がまず加算され、次いで勾配が計算されることを表している。損失および第4の勾配は、タスクτ_iのクエリサンプル画像と、対応する

に基づいて計算される。このバッチに対して更新される顔の偽造防止モデルのパラメータは、

である。βは、第2のステップ値である。

は、第4の勾配を計算するための中間変数であり、モデルのパラメータはバッチの開始時にθに基づいて各バッチについて更新されることが分かる。顔の偽造防止モデルは、クエリサンプル画像によって更新され、タスクについてのモデルの一般化能力を向上させ、サポートサンプル画像の過剰適合を回避できる。 In addition, for the batch, the fourth gradient corresponding to the batch,

This means that the losses of multiple tasks in the batch are added first, and then the gradient is calculated. The loss and the fourth gradient correspond to the query sample image of task _τi .

It is calculated based on. The face anti-counterfeiting model parameters updated for this batch are:

Is. β is the second step value.

Is an intermediate variable for calculating the fourth gradient, and it can be seen that the model parameters are updated for each batch based on θ at the start of the batch. The face anti-counterfeiting model can be updated with query sample images to improve the model's generalization ability for tasks and avoid overfitting of support sample images.

If the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label, and the category is the counterfeiting method. Or include at least one of the models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. It is a classification label. For the actual facial image, the preset values can be labeled when performing the labeling using the impersonation method. That is, if the face anti-counterfeiting model is a classification model, for each task in this batch, the third gradient corresponding to this task is the first classification label corresponding to each support sample image and each support sample image. Is determined based on. Alternatively, if the face anti-counterfeiting model is a regression model, for each task in the batch, the third gradient corresponding to this task is the second classification corresponding to each support sample image and each support sample image. Determined based on the label.

In step S208, it is determined whether the anti-counterfeiting model has reached the convergence condition. If YES, the pre-training of the anti-counterfeiting model is completed, and if NO, the process returns to step S206 for restart. ..

The convergence condition corresponds, for example, to minimizing the value of the loss function or reaching a preset value of the loss function. If the value of the loss function does not reach the convergence condition, return to step S206 to select another batch for training the face anti-counterfeiting model.

The steps described above belong to the pre-training process for the anti-counterfeiting model of the face. The number of supported sample images in each category of each task is very small and therefore belongs to a small sample set, but a large number of samples (including many sample images in existing categories and actual image categories) are pre-existing. Note that it is used during the entire training process. Also, the larger the number of samples, the better. The only difference is that many tasks and many small sample sets are split so that the model can "learn how to learn" during the pre-training process, increasing the accuracy of the model. be able to.

In step S210, a small sample set of forged facial images in a new category is obtained.

A small sample set of forged facial images in the new category can serve as support sample images in the task. The task contains one or more new categories, with each new category corresponding to a small sample set.

In step S212, the parameters of the pre-trained facial anti-counterfeiting model are updated with a small sample set of counterfeit facial images in a new category.

を決定するプロセスに類似している。例えば、タスクτ_jのサポートサンプル画像として1つまたは複数の新しいカテゴリの小さなサンプルセット内の偽造された顔画像の場合、このタスクτ_jに対応する顔の偽造防止モデルのパラメータの第3の勾配

が計算され、タスクτ_j対して決定された顔の偽造防止モデルのパラメータは、

と表され、ここで、αは第1のステップ値を表す。すなわち、更新された顔の偽造防止モデルのパラメータは、

である。クエリサンプル画像を使用した更新プロセスは、実行されない。 How to update the parameters of the pre-trained face anti-counterfeiting model with a small sample set of forged face images in the new category is by using the support sample images of task _τi in step S206. Anti-counterfeit model parameters

It is similar to the process of determining. For example, for a forged face image in a small sample set in one or more new categories as a support sample image for task τ _j , the third gradient of the parameters of the face anti-counterfeiting model corresponding to this task τ _j .

Is calculated and the parameters of the face anti-counterfeiting model determined for the task _τj are:

Here, α represents the first step value. That is, the parameters of the updated face anti-counterfeiting model are:

Is. The update process using the query sample image is not performed.

In step S214, the face image to be recognized is input to the face anti-counterfeiting model after the parameters are updated, and it is determined whether the face image to be recognized is a forged face image.

The solutions of the present disclosure can be applied not only to recognize forged facial images of new counterfeit methods, but also to recognize forged facial images of new models. For example, if you only need to recognize the forged face image of the new counterfeit method, the counterfeit face image and the actual face image of the existing counterfeit method will act as a training sample during the pre-training, and the training sample will Forged facial images of the same model and new forgery methods also belong to such models. In another example, if only the forged face image needs to be recognized, the forged face image and the actual face image of the existing model act as training samples during the pre-training and these of the existing model. The forged face image of is forged by the same forgery method, and the forged face image of the new forged model also belongs to such a forgery method. If the training sample contains forged facial images of both different counterfeit methods and different models, and the actual facial images of one or more models during pre-training, the new counterfeit facial images will be trained. May differ from any counterfeiting method or any model in the sample.

For example, for a forged face image of a new counterfeit method, in some embodiments, a small sample set of the forged face image of the new counterfeit method is obtained, and the parameters of the face anti-counterfeiting model are the parameters of the new counterfeit method. Updated and pre-trained facial anti-counterfeiting model parameters with a small sample set of forged facial images, each pre- trained facial anti-counterfeiting model parameter was determined based on multiple tasks. Determined on the basis of multiple sets of pre-trained parameters, which are multiple sets of, each task is a small sample set of each forged facial image of one or more existing counterfeit methods and the actual A face image to be recognized, including a small sample set of face images, is entered into the face anti-counterfeiting model after the parameters have been updated to determine if the face image to be recognized is a forged face image. Will be done.

For example, for a forged face image of a new model, in some embodiments a small sample set of the forged face image of the new model was obtained and the parameters of the anti-face anti-counterfeit model were forged in the new model. Updated with a small sample set of facial images, the parameters of the pre-trained facial anti-counterfeiting model are multiple of the pre -trained parameters of the pre-trained facial anti-counterfeiting model, each determined based on multiple tasks. Determined on the basis of multiple sets of pre-trained parameters that are a set, each task is a small sample set of each forged facial image of one or more existing models and one or more existing ones. The face image to be recognized was entered into the face anti-counterfeiting model after the parameters were updated, including a small sample set of each actual face image of the model, and the face image to be recognized was forged. It is determined whether it is a face image.

The pre-training process is similar to that of the embodiments described above. In situations where the training sample contains different forgery methods and forged facial images of different models, for example, if there are 3 known models and 3 known forgery methods, up to 9 different existing categories of forgery. There can be a facial image. In addition, if there are actual facial images of the three models, the generated task can contain up to 12 small sample sets.

In the embodiments described above, the pre-trained facial anti-counterfeiting model is a set of small samples, one or more existing anti-counterfeiting, consisting of forged facial images of one or more existing models. Trained by the method, or the actual facial images of multiple existing models, multiple groups of facial anti-counterfeiting model parameters are determined by multiple tasks, respectively, and then facial anti-counterfeiting model parameters are determined during the training process. During the redetermination based on multiple sets of pre-trained parameters. This training process consists of two stages: the first stage of training a facial anti-counterfeiting model using multiple small sample sets of forged facial images in an existing category and real facial images in an existing category. And with the first stage, where the model learns the features of a small sample set, or the stage where it learns to distinguish between a fake face and a real face, as well as performing aggregation. The second step is to re-update the parameters of the face anti-counterfeiting model based on multiple training results, learning how the model learns a small sample set of features, or the forged face and the actual It includes a second stage, which is a stage for learning how to learn to distinguish from the face.

Therefore, a pre-trained face anti-counterfeiting model learns how to learn a small sample set or how to distinguish between a forged face and a real face, and proper initialization. Get the parameters. The face anti-counterfeiting model is updated when the forged face image of the new model belongs to the conditions of a small sample set. As an advantage of the fast learning ability previously learned, the face anti-counterfeiting model also performs fast convergence and fast learning of a new category of counterfeit facial features, making the face anti-counterfeiting model timely and effective. Updates can be performed, the forged face image of the new category can be recognized in a timely and effective manner, and the attack from the forged face of the new category can be effectively protected in a timely manner.

The disclosure also provides a forged face recognition device, described below with FIG.

FIG. 3 shows a schematic structural diagram of a forged face recognition device according to some embodiments of the present disclosure. As shown in FIG. 3, the device 30 of this embodiment includes a small sample acquisition module 302, a model adjustment module 304, and a face recognition module 306.

The small sample acquisition module 302 is configured to acquire a small sample set of forged facial images in a new category.

The model adjustment module 304 is configured to update the parameters of a pre-trained facial anti-counterfeiting model with a small sample set of pre-trained facial anti-counterfeiting models in a new category. Pre-trained parameters are determined based on multiple sets of pre -trained parameters, multiple sets of pre-trained parameters are determined based on multiple tasks, and each task is determined. Includes a small sample set of forged facial images for each of one or more existing categories and a small sample set of actual facial images for each of one or more existing categories.

In some embodiments, the model adjustment module 304 is based on a forged face image in a small sample set of forged face images in a new category and a classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is configured to update the parameters of the anti-counterfeiting model and the anti-counterfeiting model is a classification model, the classification label will now label the category of forged face images corresponding to the classification label. The first classified label constructed, the category comprising at least one of the counterfeiting methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, the classification label is a second classification label configured to label the forged face image corresponding to the classification label as a positive sample or a negative sample. Is.

In some embodiments, the model adjustment module 304 turns the forged face image in a small sample set of forged face images in a new category into a face anti-counterfeiting model as a support sample image for one task. Enter and determine the gradient of the face anti-counterfeiting model parameters based on the output value of the face anti-counterfeiting model for each of the forged facial images and the classification label corresponding to each of the forged facial images. , Based on the gradient of the face anti-counterfeiting model parameters, it is configured to update the face anti-counterfeiting model parameters.

In some embodiments, if the anti-counterfeit model of the face is a classification model, the output value of the anti-counterfeit model for each of the counterfeit face images is the probability that each of the counterfeit face images belongs to a different category. And the model adjustment module 304 is an anti-counterfeiting model of the face based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images. If the anti-counterfeiting model of the face is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is configured to determine the first gradient of the parameters of the counterfeited face. Each of the images is the probability of belonging to the positive sample, and the model adjustment module 304 has the probability that each of the forged face images belongs to the positive sample and the second classification label corresponding to each of the forged face images. Based on, it is configured to determine the second gradient of the parameters of the anti-counterfeiting model.

The face recognition module 306 is configured to input a face image to be recognized in the face anti-counterfeiting model after updating the parameters and determine whether the face image to be recognized is a forged face image. To.

In some embodiments, the device 30 further comprises a pre-training module 308. The pre-training module 308 is to divide the forged face image of the existing category and the actual face image of the existing category into a plurality of tasks, and each task has a plurality of methods, and each method has a plurality of methods. Each task has a first preset number of existing categories, each method supports, with multiple real facial images of one model or multiple forged facial images of existing categories. A second preset number of real or forged face images as a small sample set consisting of sample images, and a third preset number of real face images as a query sample image. Or with a forged facial image, and to select a fourth preset number of tasks to form a batch each time, of the pre-trained parameters of the facial anti-counterfeiting model. Multiple sets are each determined based on multiple tasks in the batch, and face anti-counterfeiting model parameters are redetermined based on multiple sets of pre-trained parameters, and face counterfeiting. It is configured to complete pre-training of the face anti-counterfeiting model until the prevention model reaches the convergence condition.

In some embodiments, the pre-training module 308 determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and the first for each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on a gradient of 3 and a set of parameters determined by each of the multiple tasks in the batch, query sample images corresponding to the multiple tasks, and Determine the fourth gradient corresponding to the batch based on the classification label corresponding to each of the query sample images, and redetermine the parameters of the anti-counterfeit model based on the fourth gradient corresponding to the batch. It is configured to do things.

In some embodiments, when the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of support sample images corresponding to the classification label. Yes, the category contains at least one of the counterfeit methods or models. Alternatively, if the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. Is.

In some embodiments, the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method.

In some embodiments, different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises multiple photographs, videos, or masks, and the model is a depth model. The actual facial images of different tasks, including multiple near-infrared models, or red-green-blue (RGB) models, have the same model or different models, and new categories of counterfeit facial images are of the counterfeit method or model. Unlike existing images in at least one, the existing image includes a forged facial image in an existing category and an actual facial image in an existing category.

The forged face recognition device of the embodiments of the present disclosure is implemented by various computing devices or computer systems and is described below with FIGS. 4 and 5.

FIG. 4 shows a schematic structural diagram of a forged face recognition device according to some embodiments of the present disclosure. As shown in FIG. 4, the device 40 of the embodiment comprises a memory 410 and a processor 420 coupled to the memory 410, wherein the processor 420 is any of the present disclosures based on instructions stored in the memory 410. It is configured to implement the forged face recognition method described in the embodiment.

The memory 410 may include, for example, a system memory, a predetermined non-volatile storage medium, and the like. System memory stores, for example, operating systems, applications, boot loaders, databases, and other programs.

FIG. 5 shows a schematic structural diagram of a forged face recognition device according to another embodiment of the present disclosure. As shown in FIG. 5, the device 50 of this embodiment includes a memory 510 and a processor 520, which are similar to the memory 410 and the processor 420. In addition, there are I / O interface 530, network interface 540, storage interface 550, etc. These interfaces 530, 540, 550, as well as the memory 510 and processor 520 in between, may be connected, for example, via bus 560. The I / O interface 530 provides a connectivity interface for input and output devices such as displays, mice, keyboards, and touch classifiers. The network interface 540 provides a connection interface for various network devices. For example, it can be connected to a database server or cloud storage server. The storage interface 550 provides a connectivity interface for external storage devices such as SD cards or USB flash disks.

Yet another embodiment of the present disclosure provides a training method for a face anti-counterfeiting model. The method of training the face anti-counterfeiting model is the step of dividing the forged face image of the existing category and the actual face image of the existing category into multiple tasks, and each task is one or more existing ones. Faces using steps and multiple tasks, including a small sample set of forged facial images for each category and a small sample set of actual facial images for each of one or more existing categories. Pre-trained based on the steps of pre-training the anti-counterfeit model to determine each of the multiple sets of pre-trained parameters for the anti-counterfeiting model and the multiple sets of pre-trained parameters. A step to determine the parameters of the face anti-counterfeiting model and a step to get a small sample set of forged facial images in a new category, when the new category is pre-trained for the face anti-counterfeiting model. It comprises steps that differ from the existing categories used and steps that update the parameters of the pre-trained facial anti-counterfeiting model with a small sample set of forged facial images in the new category.

In some embodiments, each task comprises multiple methods, each method having multiple real facial images of one model or multiple counterfeit facial images of an existing category, and each task has multiple methods. A first preset number of existing categories, each method as a small sample set composed of support sample images, a second preset number of real face images or forged faces. Having an image, having a third preset number of real or forged facial images as query sample images, and pre-training a facial anti-counterfeiting model using multiple tasks, Determine the parameters of the pre-trained face anti-counterfeiting model based on the steps to determine each of the multiple sets of pre-trained face anti-counterfeiting models and the multiple sets of pre-trained parameters. The step to do is to select a fourth preset number of tasks to form a batch each time, with multiple sets of pre-trained parameters for the face anti-counterfeiting model, multiple batches. Each step is determined based on the task, and the parameters of the face anti-counterfeiting model are redetermined based on multiple sets of pre-trained parameters, until the steps and the facial anti-counterfeiting model reach convergence conditions. It includes steps to complete pre-training of the face anti-counterfeiting model.

In some embodiments, a plurality of sets of pre-trained parameters for the anti-counterfeiting model are each determined based on multiple tasks in a batch, and the parameters for the anti-counterfeiting model are pre-trained. Redetermination based on multiple sets of parameters determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and each task in the batch. Determining one set of parameters for the anti-counterfeiting model based on a third gradient against, a set of parameters determined by each of the multiple tasks in the batch, and a query sample corresponding to the multiple tasks. Determine the fourth gradient corresponding to the batch based on the image and the classification label corresponding to each of the query sample images, and the parameters of the anti-counterfeiting model based on the fourth gradient corresponding to the batch. Be prepared to make a decision again.

Yet another embodiment of the present disclosure provides a training device for a face anti-counterfeiting model. The device comprises a memory and a processor coupled to the memory, and the processor is configured to perform a training method of a face anti-counterfeiting model according to any of the above embodiments based on instructions stored in the memory. Will be done.

In yet another embodiment of the present disclosure, a computer program is provided. The computer program, upon execution by the processor, comprises an instruction to cause the processor to execute a forged face recognition method according to a training method of a face anti-counterfeiting model or a forged face recognition method according to any of the above embodiments.

It will be apparent to those skilled in the art that embodiments of the present disclosure may be provided as methods, systems, or computer program products. Accordingly, the present disclosure may take the form of a fully hardware embodiment, a fully software embodiment, or a combination of software and hardware aspects. Further, the present disclosure is embodied in one or more computer-usable non-temporary storage media including, but not limited to, disk memory, CD-ROM, optical memory, etc., including computer-usable program code. It can take the form of a computer program product.

The present disclosure is described with flow charts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present disclosure. It can be seen that the flow chart and / or block diagram steps and / or blocks, as well as the flow chart and / or block diagram steps and / or block combinations, are implemented by computer program instructions. These computer program instructions are provided to the processor of a general purpose computer, dedicated computer, embedded processor, or other programmable data processing device to create a machine and are executed by the computer or the processor of another programmable data processing device. The instructions given give rise to a device for achieving the functions specified in one or more steps in the flowchart and / or one or more blocks in the block diagram.

These computer program instructions may also be stored in computer-readable memory that can guide the computer or other programmable data processing device to operate in a particular way, and as a result, are stored in computer-readable memory. Instructions produce a product that includes an instruction device. The instruction device implements the function specified in one or more steps in the flowchart, or in one or more blocks in the block diagram.

These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operating steps is performed on the computer or other programmable device and is performed on the computer. An instruction that produces a computer-implemented process, such as an instruction, that is executed on a computer or other programmable device is specified in one or more steps in a flowchart and / or one or more blocks in a block view. It provides the steps to realize the function.

The above description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. within the principles of this disclosure are included within the scope of this disclosure.

302 Small sample acquisition module
304 Model Adjustment Module
306 Face recognition module
308 Pre-training module
410 memory
420 processor
510 memory
520 processor
530 I / O interface
540 network interface
550 storage interface
560 bus

Claims (20)

It ’s a forged face recognition method.
Steps to get a small sample set of forged facial images in a new category,
A step of updating the parameters of a pre-trained face anti-counterfeiting model using the small sample set of pre-trained face anti-counterfeiting models in a new category. , The multiple sets of pre-trained parameters are determined based on multiple sets of parameters, respectively, and each task is a forgery of each of one or more existing categories. With steps, including a small sample set of face images that have been made and a small sample set of actual facial images for each of one or more existing categories.
A recognition method comprising a step of inputting a face image to be recognized into the face anti-counterfeiting model after updating parameters and determining whether the face image to be recognized is a forged face image. The step of updating the parameters of the anti-counterfeiting model of the face pre-trained with the small sample set of counterfeit facial images in the new category
The parameters of the anti-counterfeiting model of the face based on the forged face image in the small sample set of the forged face image of the new category and the classification label corresponding to each of the forged face images. Including the step to update
When the anti-counterfeiting model of the face is a classification model, the classification label is a first classification label configured to label a category of forged face images corresponding to the classification label. , Includes at least one counterfeit method or model, or
If the anti-counterfeiting model of the face is a regression model, the classification label is a second classification label configured to label the forged face image corresponding to the classification label as a positive sample or a negative sample. The recognition method according to claim 1. The parameters of the anti-counterfeiting model of the face based on the forged face image in the small sample set of the forged face image of the new category and the classification label corresponding to each of the forged face images. The step to update is
The step of inputting the forged face image in the small sample set of the forged face image of the new category into the anti-counterfeiting model of the face as a support sample image for one task.
The gradient of the parameters of the face anti-counterfeiting model is determined based on the output value of the face anti-counterfeiting model for each of the forged face images and the classification label corresponding to each of the forged face images. Steps to do and
The recognition method according to claim 2, comprising the step of updating the parameters of the face anti-counterfeiting model based on the gradient of the parameters of the face anti-counterfeiting model. When the anti-counterfeiting model of the face is a classification model, the output value of the anti-counterfeiting model for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a different category. ,
The gradient of the parameters of the face anti-counterfeiting model is determined based on the output value of the face anti-counterfeiting model for each of the forged face images and the classification label corresponding to each of the forged face images. Steps to do
The parameters of the anti-counterfeiting model of the face are based on the probability that each of the forged face images belongs to a different category and the first classification label corresponding to each of the forged face images. Equipped with steps to determine the first gradient, or
When the anti-counterfeiting model of the face is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a positive sample. ,
The gradient of the parameters of the face anti-counterfeiting model is determined based on the output value of the face anti-counterfeiting model for each of the forged face images and the classification label corresponding to each of the forged face images. Steps to do
The parameters of the anti-counterfeiting model of the face are based on the probability that each of the forged face images belongs to a positive sample and the second classification label corresponding to each of the forged face images. The recognition method according to claim 3, comprising a step of determining a second gradient. It is a step of dividing the forged face image of the existing category and the actual face image of the existing category into a plurality of tasks.
Each task has multiple methods, each method has multiple real facial images of one model or multiple counterfeit facial images of existing categories, and each task has a first preset. Has a number of existing categories,
Each method has a second preset number of real or forged face images as a small sample set consisting of support sample images, and a third preset number as query sample images. With a step and, with a real face image or a forged face image of
A fourth step of selecting a preset number of tasks to form a batch each time, wherein the plurality of sets of parameters of the face anti-counterfeiting model are based on the plurality of tasks of the batch, respectively. A step and a step in which the parameters of the face anti-counterfeiting model are determined and redetermined based on the plurality of sets of parameters.
The recognition method according to claim 1, further comprising a step of completing the pre-training of the face anti-counterfeiting model until the face anti-counterfeiting model reaches the convergence condition. The plurality of sets of parameters of the face anti-counterfeiting model are each determined based on the plurality of tasks of the batch, and the parameters of the face anti-counterfeiting model are re-based on the plurality of sets of parameters. To be decided
A third gradient corresponding to the task is determined based on each support sample image and the classification label corresponding to each support sample image, and the face counterfeiting is based on the third gradient for each task in the batch. Determining one set of prevention model parameters,
Corresponds to the batch based on a set of parameters determined by each of the plurality of tasks in the batch, the query sample image corresponding to the plurality of tasks, and the classification label corresponding to each of the query sample images. To determine the fourth gradient to be
5. The recognition method of claim 5, comprising redetermining the parameters of the face anti-counterfeiting model based on the fourth gradient corresponding to the batch. When the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of the support sample image corresponding to the classification label. The category contains at least one of the counterfeit methods or models, or
If the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. The recognition method according to claim 6, which is a classification label. The recognition method according to claim 1, wherein the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to the deep learning training method. Different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises a photo, video, or mask, and the model is a depth model, near infrared model, or red. Including green-blue (RGB) model,
The actual facial images of the different tasks have the same model or different models and
The forged face image in the new category is different from the existing image in at least one of the forgery methods or models, and the existing image is the forged face image in the existing category and the existing category. The recognition method according to any one of claims 1 to 8, which includes an actual face image. A forged face recognition device,
With a small sample acquisition module configured to acquire a small sample set of forged facial images in a new category,
A model adjustment module configured to update the parameters of a pre-trained anti-counterfeiting model of a pre-trained face using the small sample set of pre-trained face images in a new category of the pre-trained face. The parameters of the anti-counterfeit model are determined based on a plurality of sets of parameters, the plurality of sets of pre-trained parameters are determined based on a plurality of tasks, and each task is one or more. A model tuning module, including a small sample set of forged facial images for each existing category and a small sample set of actual facial images for each of one or more existing categories.
A face recognition module configured to input a face image to be recognized in the face anti-counterfeiting model after updating the parameters and determine whether the face image to be recognized is a forged face image. And with a recognition device. The model adjustment module is based on the forged face image in the small sample set of the forged face image of the new category and the classification label corresponding to each of the forged face images of the face. It is configured to update the above parameters of the anti-counterfeit model,
When the anti-counterfeiting model of the face is a classification model, the classification label is a first classification label configured to label a category of forged face images corresponding to the classification label. , Includes at least one counterfeit method or model, or
If the anti-counterfeiting model of the face is a regression model, the classification label is a second classification label configured to label the forged face image corresponding to the classification label as a positive sample or a negative sample. The recognition device according to claim 10. The model adjustment module inputs the forged face image in the small sample set of the forged face image of the new category into the face anti-counterfeiting model as a support sample image for one task. The gradient of the parameters of the face anti-counterfeiting model is determined based on the output value of the face anti-counterfeiting model for each of the forged face images and the classification label corresponding to each of the forged face images. 11. The recognition device of claim 11, configured to update the parameters of the face anti-counterfeiting model based on the gradient of the parameters of the face anti-counterfeiting model. When the anti-counterfeiting model of the face is a classification model, the output value of the anti-counterfeiting model for each of the forged face images is the probability that each of the forged face images belongs to a different category. , The model adjustment module is based on the probability that each of the forged facial images belongs to a different category and the first classification label corresponding to each of the forged facial images. It is configured to determine the first gradient of said parameter in the anti-counterfeit model, or
When the anti-counterfeiting model of the face is a regression model, the output value of the anti-counterfeiting model for each of the counterfeited face images is the probability that each of the counterfeited face images belongs to a positive sample. , The model adjustment module is based on the probability that each of the forged facial images belongs to a positive sample and the second classification label corresponding to each of the forged facial images. 12. The recognition device of claim 12, configured to determine a second gradient of said parameter of the anti-counterfeit model. It is to divide the forged face image of the existing category and the actual face image of the existing category into a plurality of tasks.
Each task has multiple methods, each method has multiple real facial images of one model or multiple counterfeit facial images of existing categories, and each task has a first preset. Has a number of existing categories,
Each method has a second preset number of real or forged face images as a small sample set consisting of support sample images, and a third preset number as query sample images. With a real face image or a forged face image,
A fourth preset number of tasks is to be selected to form a batch each time, wherein the plurality of sets of parameters of the anti-counterfeiting model of the face are based on the plurality of tasks of the batch, respectively. The parameters of the face anti-counterfeiting model are determined and redetermined based on the plurality of sets of parameters.
The recognition according to claim 10, further comprising a pre-training module configured to complete said pre-training of the face anti-counterfeiting model until the face anti-counterfeiting model reaches a convergence condition. device. The pre-training module determines a third gradient corresponding to the task based on each support sample image and the classification label corresponding to each support sample image, and the third gradient for each task in the batch. To determine one set of parameters for the anti-counterfeiting model of the face, a set of parameters determined by each of the plurality of tasks in the batch, and the query sample image corresponding to the plurality of tasks. , And the fourth gradient corresponding to the batch based on the classification label corresponding to each of the query sample images, and the forgery of the face based on the fourth gradient corresponding to the batch. The recognition device according to claim 14, wherein the parameters of the prevention model are configured to be redetermined. When the anti-counterfeiting model of the face is a classification model, each of the classification labels is a first classification label configured to label the category of the support sample image corresponding to the classification label. The category contains at least one of the counterfeit methods or models, or
If the anti-counterfeiting model of the face is a regression model, each of the classification labels is configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. The recognition device according to claim 15, which is a classification label. The recognition device according to claim 10, wherein the initial parameters of the anti-counterfeiting model of the face before pre-training are determined according to a deep learning training method. Different existing categories of counterfeit facial images differ in at least one of the counterfeit methods or models, the counterfeit method comprises a photo, video, or mask, and the model is a depth model, near infrared model, or red. Including green-blue (RGB) model,
The actual facial images of the different tasks have the same model or different models and
The forged face image in the new category is different from the existing image in at least one of the forgery methods or models, and the existing image is the forged face image in the existing category and the existing category. The recognition device according to any one of claims 10 to 16, including an actual facial image. With memory
A processor coupled to the memory is provided, and the processor is configured to implement the forged face recognition method according to any one of claims 1 to 9 based on an instruction stored in the memory. A forged facial recognition device. A computer-readable storage medium containing a computer program that implements the method according to any one of claims 1 to 9 when executed by a processor.

Images