KR20220007130A - Counterfeit face recognition method, device and non-transitory computer readable storage medium - Google Patents

Abstract

The present disclosure relates to a fake face recognition method, an apparatus, and a computer-readable storage medium related to the field of computer technology. The method of the present disclosure includes: acquiring a small sample set of fake face images of a new category; updating parameters of the pre-trained anti-spoofing face model using a small sample set of new categories of spoof face images, the parameters of the pre-trained anti-spoofing face model being determined based on a plurality of sets of parameters; , wherein the plurality of sets of parameters are a plurality of sets of anti-spoofing model parameters determined according to a plurality of tasks, each task comprising a small sample set of fake face images of each existing category of one or more existing categories and a real face Contains a small sample set of images - ; inputting a face image to be recognized into an anti-spoofing face model with updated parameters; and determining whether the face image to be recognized is a fake face image.

Description

Counterfeit face recognition method, device and computer readable storage medium

CROSS-REFERENCE TO RELATED APPLICATIONS

This application is based on, and claims priority to, Chinese Patent Application No. 201910387689.0, filed on May 10, 2019, the disclosure of which is incorporated herein by reference in its entirety.

TECHNICAL FIELD The present disclosure relates to the field of computer technology, and more particularly to a method, apparatus, and computer readable storage medium for spoofing face recognition.

Facial anti-spoofing recognition refers to effectively identifying forged facial features. For example, in a facial recognition system, some forged faces (eg, videos, photos, and masks) can be crafted to attack the facial recognition system. Face anti-spoofing recognition relates to the identification of forged facial features. In general, in a facial recognition system, a user is required to pass detection of anti-counterfeiting face recognition, which may also be referred to as biometric detection.

The fake face recognition method known to the present inventors is to perform deep learning training on a classification model based on a plurality of fake face images of a known category. The trained classification model can recognize fake faces belonging to the same category of fake face images for training.

The present inventors found that in the existing fake face recognition method, it is necessary to use a large number of samples to train the model so that the trained model can accurately recognize the fake face. When a new category of fake face image emerges, when there are very few fake face image samples, the model training cannot continue, so that the model cannot recognize the fake face of the new category in a timely manner, It becomes impossible to effectively prevent attacks from a new category of fake faces in a timely manner.

One technical problem to be solved by the present disclosure is a method for realizing timely and efficient recognition of fake face images of a new category, and improving the safety of a face recognition system when a small number of fake face images of a new category appear to be.

According to some embodiments of the present disclosure, a fake face recognition method is provided. The method comprises: acquiring a small sample set of fake face images of a new category; updating parameters of the pre-trained anti-counterfeiting model with a small sample set of new categories of spoofed face images, wherein the parameters of the pre-trained anti-spoofing model are a plurality of sets each determined based on a plurality of tasks determined based on the parameters of , each task including a small sample set of fake face images in each of one or more existing categories and a small sample set of real face images in each of one or more existing categories; and inputting the to-be-recognized face image into the face anti-counterfeiting model after parameter update, to determine whether the to-be-recognized face image is a forged face image.

According to some embodiments, based on a fake face image in a small sample set of fake face images of a new category and a classification label corresponding to each of the fake face images, updating a parameter of the anti-counterfeiting model, the when the anti-counterfeiting face model is a classification model, the classification label is a first classification label, configured to label a category of a forged face image corresponding to the classification label, wherein the category includes at least one of a forgery method or a model; or when the anti-counterfeiting model is a regression model, the classification label is a second classification label, configured to label a fake face image corresponding to the classification label as a positive sample or a negative sample.

According to some embodiments, on the basis of a fake face image in a small sample set of fake face images of a new category and a classification label corresponding to each of the fake face images, the step of updating the parameter of the anti-counterfeiting model comprises: inputting a fake face image in a small sample set of fake face images of a category into the anti-counterfeiting model as a supporting sample image for one task; determining a gradient of a parameter of the anti-counterfeiting model based on a classification label corresponding to each of the forged face images and an output value of the anti-counterfeiting model for each of the forged face images; and updating the parameter of the anti-counterfeiting model based on the gradient of the parameter of the anti-counterfeiting model.

According to some embodiments, when the anti-counterfeiting model is a classification model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a different category, and the forged face Based on the classification label corresponding to each image and the output value of the anti-counterfeiting model for each of the forged face images, the step of determining the inclination of the parameter of the anti-counterfeiting model includes: determining a first inclination of a parameter of the anti-counterfeiting model, based on a first classification label and a probability that each of the fake facial images belongs to a different category; Alternatively, when the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a positive sample, and corresponding to each of the forged face images. Based on the classification label and the output value of the anti-counterfeiting model for each of the forged face images, the step of determining the inclination of the parameter of the anti-counterfeiting model includes: a second classification label corresponding to each of the forged face images; and determining a second gradient of a parameter of the anti-spoofing model, based on a probability that each of the fake facial images belongs to a positive sample.

According to some embodiments, the recognition method comprises: dividing a fake face image of an existing category and a real face image of an existing category into a plurality of tasks, each task including a plurality of methods, each method comprising one including a plurality of real face images of the model or a plurality of fake face images of an existing category, each task including a first preset number of existing categories, each scheme comprising a small sample set consisting of a supporting sample image a second preset number of real face images or fake face images, and a third preset number of real face images or fake face images as query sample images; and a plurality of sets of parameters of the anti-counterfeiting model are respectively determined based on a plurality of tasks in a batch and parameters of the anti-counterfeiting model are re-determined based on a plurality of sets of parameters, each time selecting a fourth preset number of tasks to form a batch; and completing pre-training of the anti-counterfeiting model until the anti-counterfeiting model reaches a convergence state.

According to some embodiments, a plurality of sets of parameters of the anti-counterfeiting model are respectively determined based on a plurality of tasks of the batch, and the parameters of the anti-counterfeiting model are determined based on the plurality of sets of parameters. The re-determining is to determine, based on each support sample image and a classification label corresponding to each support sample image, a third gradient corresponding to the task, and based on the third gradient for each task in the batch, , determining a parameter of one set of the anti-counterfeiting model; determining a fourth slope corresponding to the batch based on a parameter set determined by each of a plurality of tasks in the batch, a query sample image corresponding to the plurality of tasks, and a classification label corresponding to each of the query sample images thing; and re-determining the parameter of the anti-counterfeiting model based on the fourth gradient corresponding to the arrangement.

According to some embodiments, when the face anti-spoofing prevention model is a classification model, each of the classification models is a first classification label, configured to label a category of a support sample image corresponding to the classification model, wherein the category is a forgery at least one of a method or a model. Alternatively, when the anti-counterfeiting model is a regression model, each of the classification labels is a second classification label, configured to label a supporting sample image corresponding to the classification label as a positive sample or a negative sample.

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

According to some embodiments, different pre-existing categories of fake face images are different in at least one of a forgery method or model, wherein the forged method includes a photo, a video or a mask, and wherein the model is a depth model, near-infrared model or red-green-blue (RGB) model; Real face images in different tasks have the same or different models; The fake face image of the new category is different from the existing image in at least one of a forgery method or a model, and the existing image includes a fake face image of the old category and a real face image of the old category.

According to another embodiment of the present disclosure, a fake face recognition apparatus is provided. The apparatus includes: a small sample acquisition module, configured to acquire a small sample set of a new category of fake face images; a model tuning module, configured to update parameters of a pre-trained anti-counterfeiting model with a small sample set of new categories of spoofed face images, wherein parameters of the pre-trained anti-spoofing model are each determined based on a plurality of tasks determined based on a plurality of sets of parameters, each task including a small sample set of fake face images in each of one or more existing categories and a small sample set of real facial images in each of one or more existing categories; and a face recognition module, configured to input the to-be-recognized face image into the anti-spoofing face model after the parameter update, to determine whether the to-be-recognized face image is a forged face image.

According to some embodiments, the model adjustment module updates the parameters of the anti-counterfeiting model, based on a fake face image in a small sample set of a new category of fake face images and a classification label corresponding to each of the fake face images wherein, when the anti-counterfeiting model is a classification model, the classification label is a first classification label, configured to label a category of a forged face image corresponding to the classification label, wherein the category is at least one of a forgery method or a model comprising; or when the anti-counterfeiting model is a regression model, the classification label is a second classification label, configured to label a fake face image corresponding to the classification label as a positive sample or a negative sample.

According to some embodiments, the model adjustment module inputs a fake face image in a small sample set of a new category of fake face images into the face anti-counterfeiting model as a support sample image for one task, each of the fake face images Based on the classification label corresponding to and the output value of the anti-counterfeiting model for each of the forged face images, determine the slope of the parameter of the anti-counterfeiting model, and based on the slope of the parameter of the anti-counterfeiting model , and update parameters of the anti-counterfeiting model.

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

Alternatively, when the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a positive sample, and the model adjustment module is configured to: and determine a second gradient of a parameter of the anti-spoofing model, based on a second classification label corresponding to each image and a probability that each of the forged facial images belongs to a positive sample.

According to some embodiments, the recognition device further comprises a pre-training module, wherein the pre-training module is configured to divide a fake face image of an existing category and a real face image of an existing category into a plurality of tasks - each task includes a plurality of schemes, each scheme includes a plurality of real face images of one model or a plurality of fake face images of an existing category, each task includes a first preset number of existing categories, Each manner includes a second preset number of real face images or fake face images as a small sample set consisting of support sample images, and a third preset number of real face images or fake face images as query sample images; ; forming a batch each time, such that a plurality of sets of parameters of the anti-counterfeiting model of the face are respectively determined based on a plurality of tasks of the batch and parameters of the anti-counterfeiting model are re-determined based on a plurality of sets of parameters. select a fourth preset number of tasks to do; and complete the pre-training of the anti-counterfeiting model until the anti-counterfeiting model reaches a convergence state.

According to some embodiments, the pre-training module is configured to: determine a third gradient corresponding to the task based on each support sample image and a classification label corresponding to each support sample image; determine a parameter of one set of the face forgery prevention model based on a third gradient for determine a fourth gradient corresponding to the placement based on a parameter set determined by a label, and re-determining the parameter of the anti-counterfeiting model based on the fourth gradient corresponding to the placement.

According to some embodiments, when the face anti-spoofing prevention model is a classification model, each of the classification models is a first classification label, configured to label a category of a support sample image corresponding to the classification model, wherein the category is a forgery at least one of a method or a model; or when the anti-counterfeiting face model is a regression model, each of the classification labels is a second classification label, configured to label a support sample image corresponding to the classification label as a positive sample or a negative sample.

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

According to some embodiments, different pre-existing categories of fake face images differ in at least one of a forgery method or model, wherein the forged method includes a photograph, a video or a mask, and wherein the model is a depth model, a near-infrared model or a red- green-blue (RGB) models; Real face images in different tasks have the same or different models; The fake face image of the new category is different from the existing image in at least one of a forgery method or a model, and the existing image includes a fake face image of the old category and a real face image of the old category.

According to another embodiment of the present disclosure, a fake face recognition apparatus is provided. The device includes a memory; and a processor coupled to the memory, wherein the processor is configured to implement the fake face recognition method according to any one of the preceding embodiments based on the instruction stored in the memory.

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

In the present disclosure, the parameters of the pre-trained anti-counterfeiting model are directly updated with a small sample set of fake face images of new categories, so that the updated anti-counterfeiting model recognizes various forged face images of new and existing categories. can do. This means that the pre-trained anti-counterfeiting model is trained with a small sample set of fake face images and a small sample set of real face images in the existing category, and the parameters of multiple sets of anti-counterfeiting models are each determined by multiple tasks. This is because the parameters of the anti-counterfeiting model are then re-determined based on the multiple sets of parameters during the training process. This training process consists of two steps: the model learns features from a small set of samples, or learns to distinguish fake faces from real faces. a first step of training each face anti-counterfeiting model with a set of samples; and learning how the model learns features of a small sample set or learning how to distinguish a fake face from a real face. and a second step of re-updating the parameters of .

Therefore, after training a number of tasks involving a small sample set of fake face images from an existing category and real face images from an existing category, the anti-counterfeiting model can master how to train a small sample set, or You have mastered how to learn to distinguish faces or learn to spot spoof clues and obtain favorable initialization parameters. When a new category of fake face images relates to the state of a small sample set, the face anti-counterfeiting model is updated. The face anti-counterfeiting model, which has benefited from the previously learned fast learning ability, also has fast convergence and fast learning of new categories of fake facial features, and timely and effective recognition of new categories of fake face images, and from new categories of fake faces. Timely and effective update of the anti-counterfeiting model can be performed to effectively and timely defend against the attacks of

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

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, described herein to provide a further understanding of the present disclosure, form a part of this application. The drawings of the present disclosure as well as the exemplary embodiments of the present disclosure to explain the present disclosure do not constitute an improper definition of the present disclosure.
In the accompanying drawings,
1 illustrates a schematic flowchart of a method for forged face recognition according to some embodiments of the present disclosure;
2 is a schematic flowchart of a fake face recognition method according to another embodiment of the present disclosure.
3 is a schematic structural diagram of a forged face recognition apparatus according to some embodiments of the present disclosure.
4 is a schematic structural diagram of a forged face recognition apparatus according to some other exemplary embodiments of the present disclosure.
5 is a schematic structural diagram of a forged face recognition apparatus according to another exemplary embodiment of the present disclosure.

The technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present disclosure. It is clear that the described embodiments are not all but only some of the embodiments of the present disclosure. The following description of at least one exemplary embodiment, merely illustrative in nature, should not serve as any limitation to the disclosure, as well as the application or use of the disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those of ordinary skill in the art to which this technology pertains, on the premise that progressive efforts are not involved, will fall within the protection scope of the present disclosure.

When the above model cannot timely and effectively recognize a new category of fake face, when a new category of fake face image emerges and there is a very small number of samples, it can effectively defend against an attack from a new category of fake face in a timely manner. To solve the missing problem, the present solution is provided. A fake face recognition method according to some embodiments of the present disclosure will be described below with reference to FIG. 1 .

1 illustrates a schematic flowchart of a method for forged face recognition according to some embodiments of the present disclosure; As shown in Fig. 1, the method of this embodiment includes steps S102 to S106.

In step S102, a small sample set of fake face images of a new category is acquired.

In the present disclosure, the number of images in a small sample set is less than a threshold. The threshold is such that the face anti-spoofing model is updated with a small sample set according to the deep learning training method, so that the model cannot converge. The threshold is, for example, 10 to 30, but is not limited to the examples listed. The new category of fake face images belongs to a different category from the fake face images used during the pre-training of the anti-counterfeiting model, and the category of the fake face images used during the pre-training of the anti-counterfeiting model can be referred to as an existing category. .

The different existing categories of fake face images differ in at least one of the fake methods or models. Counterfeiting methods include, for example, photographs, videos and masks, and models include, for example, Depth models, Near Infrared (NIR) models and Red-Green-Blue (RGB) models. For example, a fake face image with a photo as a forgery method and a depth model as a model falls into a different category than a fake face image with a photo as a forgery method and an RGB model as a model, and a photo with a forgery method. , and a fake face image with a depth model as a model also belongs to a different category from a fake face image with a video as a fake method and an RGB model as a model. A multi-model image refers to an image collected by different descriptive means to describe different properties of the same object. For example, for a particular object, a color image captured by a camera and a depth image scanned by a laser radar are referred to as two model images of the object. The fake face image of the new category is different from the existing image in at least one of a forgery method or a model, and the existing image includes a fake face image and a real face image of the old category.

In step S104, the parameters of the pre-trained anti-counterfeiting model are updated with a small sample set of a new category of fake face images. parameters of the pretrained anti-spoofing model are determined based on a plurality of sets of parameters, and parameters of the plurality of sets of anti-spoofing models are respectively determined based on a plurality of tasks; Each task includes a small sample set of fake face images and a small sample set of real face images in each of one or more existing categories.

In some embodiments, the parameters of the anti-spoofing model are updated based on a plurality of fake face images within a small sample set of fake face images of a new category and a classification label corresponding to each of the fake face images. The face anti-counterfeiting model may be a classification model or a regression model. When the anti-counterfeiting face model is a classification model, each classification label is a first classification label, configured to label a category of a fake face image corresponding to the classification label, wherein the category includes at least one of a forgery method or a model. . Alternatively, when the anti-spoofing face model is a regression model, each classification label is a second classification label, configured to label a fake face image corresponding to the classification label as a positive sample or a negative sample. The first classification label may label at least one of a model of the fake face image or a method of forgery. If it is necessary to learn features of images of different spoof methods and different models, both spoof methods and models are labeled.

Each of the fake face images in a small sample set of fake face images of a new category is input into a pre-trained anti-counterfeiting model, so that the slope of the parameters of the anti-counterfeiting model can be calculated using the output value, the classification label corresponding to each of the fake face images, and the pre-trained anti-counterfeiting model. It is determined based on the set loss function, and the parameters of the pre-trained anti-counterfeiting model are updated based on the gradient. When the face forgery prevention model is a classification model, the preset loss function may be a mean square error (MSE) loss function. When the face forgery prevention model is a regression model, the preset loss function may be a cross entropy loss function, but this is not limited to the listed examples. A method for updating the parameters of the pre-trained anti-counterfeiting model based on the gradient may be an existing gradient descent method if required, which is not described in detail herein.

In some embodiments, a plurality of fake face images in a small sample set of a new category of fake face images are input to the face anti-counterfeiting model as support sample images for one task; the gradient of the parameter of the anti-counterfeiting model is determined based on a classification label corresponding to each of the forged face images and an output value of the anti-counterfeiting model for each of the forged face images; The parameter of the anti-counterfeiting model is updated based on the gradient of the parameter of the anti-counterfeiting model.

A task may be constructed based on a small sample set of a new category of fake face images. The fake face images in the small sample set of the new category of fake face images may be the supporting sample images of the task. This is similar to the form of setting the task in the pre-training method of the fake face model, which will be described later. A constructed task may include one or more different new categories, each new category having a corresponding small set of samples. Tasks may also include small sample sets of real face images, and the like.

In some embodiments, the anti-counterfeiting model is updated based on a first classification label in which a parameter of the anti-counterfeiting model corresponds to each fake face image and each of the fake face images in a small sample set of a new category of fake face images. It may be a classification model as much as possible. For example, the first classification label of a real face image whose model is a depth model is 0, the fake method is a mask, the first classification label of a fake face image whose model is a depth model is 1, and the first classification label of the new category is 1 The label is 2.

The fake face image and the real face image may be labeled with a first classification label or a second classification label according to an 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 a loss function value, a gradient value, etc., which may improve the accuracy of the model. When the anti-counterfeiting model classifies images and outputs the results, it only needs to classify the images into two categories: real face images and fake face images. During the testing process after the anti-spoofing model is pre-trained, the second classification label may be used to test the accuracy of the anti-spoofing model in recognizing whether an image is fake or real. For example, the second classification label representing the fake face image is 0, and the second classification label representing the real face image is 1.

In addition, when the anti-counterfeiting model is a classification model, the output value of the anti-counterfeiting model for each forged face image is the probability that each forged face image belongs to a different category, and accordingly, the parameter of the anti-counterfeiting model is The first gradient is determined based on a first classification label corresponding to each of the fake face images and a probability that each of the fake face images belongs to a different category, respectively.

In some embodiments, the anti-counterfeiting model is updated based on a second classification label in which the parameter of the anti-counterfeiting model corresponds to each forged face image and each of the forged face images in a small sample set of a new category of forged face images. Preferably, it may be a regression model. The regression model may finally output a value for the input image, and a threshold value corresponding to the value may be set. If the value is greater than the threshold, it means that the image is a real face image, and if the value is less than the threshold, it means that the image is a fake face image. Thus, the regression model can apply a second classification label of the image.

In addition, when the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the forged face images is the probability that each forged face image belongs to the positive sample, and accordingly, the second corresponding to each of the forged face images A second gradient is determined based on the classification label and the probability that each fake face image belongs to a positive sample, and a parameter of the face anti-counterfeiting model is updated based on the second gradient.

The parameters of the pre-trained anti-counterfeiting model are trained by the following method: parameters of multiple sets of anti-counterfeiting models are respectively determined based on multiple tasks, and then parameters of the anti-counterfeiting model is re-determined based on the plurality of sets of parameters. Each task includes a small sample set of fake facial images in each of one or more existing categories and a small sample set of real facial images in each of one or more existing categories. The real face image may be divided into real face images of different models. Each task may include a small sample set of real face images of each of one or more models. Real face images in different tasks have the same or different models. That is, a small sample set consisting of a fake face image of an existing category and a real face image of an existing category is assigned to multiple tasks. Although, different small sample sets can be used to train the anti-counterfeiting model, it is necessary to recrystallize the parameters of the anti-counterfeiting model after training based on a large number of training results. The process of training a face anti-spoofing model with different small sample sets is a process of model learning, and the process of recrystallizing parameters based on multiple training results is a process of "learning how to learn" the model.

After the above two processes, the model not only learned the features of a small sample set, but also learned how to learn the features of a small sample set, or the face anti-counterfeiting model finally, simply identifies fake faces and real faces Instead of learning how to do it or simply learning fake clues, we learned "how to learn to identify fake faces and real faces" and "how to learn to find fake clues". Therefore, when the fake face image of a new category belongs to a small sample set, the face anti-counterfeiting model can also perform fast convergence and fast learning of the features of the fake face of the new category, making effective and timely defense.

In step S106, the face image to be recognized is input into the face anti-counterfeiting model after the parameter is updated, so that it is determined whether the face image to be recognized is a forgery face image.

The face image to be recognized may be a fake face image belonging to a new category or an existing category, both of which can be recognized by the face forgery prevention model.

In the method of the above-described embodiment, the parameters of the pre-trained anti-counterfeiting model are directly updated with a small sample set of spoof face images of a new category, so that the updated anti-counterfeiting model is configured with a variety of spoof face images of the old category and the new category. can recognize This means that the pre-trained anti-counterfeiting model is trained with a small sample set of fake face images and a small sample set of real face images in the existing category, and the parameters of multiple sets of anti-counterfeiting models are each determined by multiple tasks. This is because the parameters of the anti-counterfeiting model are then re-determined based on the multiple sets of parameters during the training process. This training process consists of two steps: the first step of training the anti-spoofing model on a large number of small sample sets of fake face images from an existing category and real face images from an existing category, respectively - this is where the model learns features from a small set of samples. Stage or stage - learning to identify fake and real faces; and a second step of performing aggregation and re-updating the parameters of the anti-counterfeiting model based on the plurality of training results, wherein the model learns how to learn features from a small sample set, or identifies fake and real faces. It is the stage of learning how to learn to do - including.

Therefore, after training a number of tasks involving a small sample set of fake face images from an existing category and real face images from an existing category, the anti-counterfeiting model can master how to train a small sample set, or You have mastered how to learn to identify faces or learn to spot spoofing cues and obtain favorable initialization parameters. When a new category of fake face images relates to the state of a small sample set, the face anti-counterfeiting model is updated. The face anti-counterfeiting model, which has benefited from the previously learned fast learning ability, also has fast convergence and fast learning of new categories of fake facial features, and timely and effective recognition of new categories of fake face images, and from new categories of fake faces. Timely and effective update of the anti-counterfeiting model can be performed to effectively and timely defend against the attacks of

A fake face recognition method according to another embodiment of the present disclosure will be described later with reference to FIG. 2 . This embodiment includes a pre-training method and a working method of a face anti-counterfeiting model.

2 is a schematic flowchart of a fake face recognition method according to another embodiment of the present disclosure. As shown in Fig. 2, the method of this embodiment includes steps S202 to S214.

In step S202, the anti-counterfeiting face model is initially trained with a fake face image of an existing category and a real face image of an existing category according to the deep learning training method, so that the initial parameters of the anti-counterfeiting model are determined before pre-training. .

First, the anti-counterfeiting face model is trained using the existing deep learning training method, and then proceeds to the subsequent pre-training process, so that the face anti-spoofing model after double training shows favorable identification and generalization. Because the anti-counterfeiting face model can more accurately identify images due to the deep learning training method, and the model can generalize different categories of fake face images according to the pre-training method of the present disclosure, the face anti-counterfeiting model after double training may indicate favorable identification for known or similar tasks, as well as rapid adaptability to new tasks.

In step S204, the fake face image of the existing category and the real face image of the existing category are divided into a plurality of tasks. Each task includes multiple schemes, each scheme includes multiple real face images of one model or multiple fake face images of an existing category, each scheme comprising a small set of samples consisting of supporting sample images a second preset number of real face images or fake face images and a third preset number of real face images or fake face images as query sample images. Each task includes a first preset number of existing categories. Each existing category represents a counterfeiting method and/or model. If the anti-counterfeiting face model is only configured to recognize the fake face image with different forgery methods, the existing category is only configured to indicate the forgery method, and all the fake face images and real face images belong to the same model. If the anti-spoofing face model is configured to recognize the fake face images of different forgery methods and different models, then each existing category represents a forgery method or model.

A task includes a number of ways. Each manner includes a second preset number of support sample images and a third preset number of query sample images. The supporting sample images of each scheme form a small sample set. That is, the supported sample image of each manner within each task is less than a threshold, which is, for example, 10 to 30. During the pre-training process, the number of images in the small sample set formed by the supporting sample images may be the same as or different from the number of images in the small sample set of the new category of fake face images. If the number of images in the small sample set formed by the supporting sample images is less than or equal to the number of images in the fake face images of the small sample set of the new category, the face anti-counterfeiting model is tested in the pre-training process for the small sample of the new category. After being subsequently updated on the basis of the set, the anti-counterfeiting model may exhibit 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 a large number of images of new categories at the time of update, so that the update of the parameters of the model can perform better. .

Each time, a first preset number of existing categories can be randomly selected, and then a second preset number of support sample images and a third preset number of query sample images can be randomly selected from the selected fake face images of existing categories. have. Real face images of one or more models may be randomly selected, and then a second preset number of support sample images and a third preset number of query sample images may be randomly selected from the selected real face images to create a task. Images in different tasks cannot be exactly the same.

In step S206, a fourth preset number of tasks is selected to form a batch, so that parameters of a plurality of sets of anti-counterfeiting models are respectively determined based on the plurality of tasks of the batch, The parameters of the prevention model are re-determined based on the plurality of sets of parameters.

A fourth preset number of tasks may be arbitrarily selected to form a batch. For example, 5 tasks are selected as batches. For each batch, the parameters of the anti-counterfeiting model can be updated twice. The first update is to update the parameters of the anti-counterfeiting model of the face with the supported sample images in each task. The second update is to update the parameters of the anti-counterfeiting model with the query sample image in each task. That is, the update consists of a learning process and the aforementioned "learning method of learning" process. This is different from traditional deep learning methods, where each batch is directly input into the model to update the parameters once.

In some embodiments, for each task in the batch, a third gradient (updated from the first gradient) corresponding to the task is determined based on each support sample image and a classification label corresponding to each support sample image; , a parameter set of the face anti-counterfeiting model is determined based on the third gradient. A plurality of sets of parameters belonging to the intermediate parameters and determined by a plurality of tasks that do not actually update the parameters of the anti-counterfeiting model may be stored for the next update. Further, a parameter set for which a fourth gradient (updated from the second gradient) corresponding to the batch is determined for each of a plurality of tasks in the batch, a query sample image corresponding to the plurality of tasks, and a classification label corresponding to each of the query sample images is determined based on The parameters of the anti-counterfeiting model are re-determined based on the fourth inclination corresponding to the arrangement. The parameters of the anti-counterfeiting model determined for the second time are used to update the parameters of the anti-counterfeiting model.

)에 대해, i는 태스크의 일련번호(serial number)를 나타내는 양의 정수이고, T는 배치의 태스크 세트이다. 각각의 태스크(

)에 대해, 태스크에 대응하는 얼굴 위조 방지 모델의 매개변수의 제3 경사도(

)가 계산되며, 여기서 θ는 얼굴 위조 방지 모델의 현재 매개변수를 나타내고,

는 태스크(

)에 대한 지원 샘플 이미지를 사용하여 현재 매개변수(θ)에 대해 계산된 손실 함수(

)의 경사도를 나타낸다. 태스크(

)에 대해 결정된 얼굴 위조 방지 모델의 매개변수는

로 표현될 수 있으며, 여기서 α는 제1 스텝값을 나타낸다. 배치 내 다수의 태스크는 (가중치 매개변수 행렬일 수 있는) 다수의

를 결정할 수 있으며, 이는 저장된다.For example, for each task in the batch (

), i is a positive integer representing the serial number of the task, and T is the set of tasks in the batch. Each task (

), the third slope (

) is calculated, where θ represents the current parameters of the face anti-spoofing model,

is the task (

) computed loss function for the current parameter (θ) using the supporting sample images for

) is the slope of the task(

), the parameters of the anti-counterfeiting model determined for

It can be expressed as , where α represents the first step value. Multiple tasks in a batch (which may be weight parameter matrices)

can be determined, which is stored.

로 표현될 수 있으며, 이는 배치 내 다수의 태스크의 손실이 먼저 합해지고, 그 후 경사도가 계산됨을 나타낸다. 손실 및 제4 경사도는 태스크(

)의 쿼리 샘플 이미지 및 대응하는

에 기초하여 계산된다. 이러한 배치에 대해 업데이트된 얼굴 위조 방지 모델의 매개변수는

이다. β는 제2 스텝값이다.

가 제4 경사도를 계산하기 위한 중간 변수이고, 모델의 매개변수가 배치의 시작 시 θ에 기초하여 각각의 배치에 대해 업데이트되는 것으로 볼 수 있다. 얼굴 위조 방지 모델이 쿼리 샘플 이미지로 업데이트되어, 태스크 상의 모델의 일반화 능력이 향상되고, 지원 샘플 이미지의 오버피팅을 방지하는 것이 가능하다.Also, for the arrangement, the fourth slope corresponding to the arrangement is

It can be expressed as , which indicates that the loss of multiple tasks in the batch is first summed, and then the gradient is calculated. The loss and the fourth slope are the tasks (

) query sample images and their corresponding

is calculated based on The parameters of the face anti-counterfeiting model updated for these deployments are:

to be. β is the second step value.

It can be seen that is an intermediate variable for calculating the fourth slope, and the parameters of the model are updated for each batch based on θ at the start of the batch. It is possible that the anti-forgery face model is updated with the query sample image, so that the generalization ability of the model on the task is improved, and overfitting of the supporting sample image is prevented.

When the anti-counterfeiting face model is a classification model, each of the classification labels is a first classification label, configured to label a category of the supporting sample image corresponding to the classification label, and the category includes at least one of a forgery method or a model. Alternatively, when the anti-counterfeiting model is a regression model, each of the classification labels is a second classification label, configured to label a supporting sample image corresponding to the classification label as a positive sample or a negative sample. For a real face image, a preset value may be labeled when labeling is performed using a forgery method. That is, when the anti-counterfeiting model is a classification model, for each task in this batch, the third gradient corresponding to this task is in each support sample image and a first classification label corresponding to each support sample image. is determined based on Alternatively, when the face anti-spoofing model is a regression model, for each task in this batch, the third gradient corresponding to this task is a second classification corresponding to each support sample image and each support sample image. It is determined based on the label.

In step S208, it is determined whether or not the anti-counterfeiting model has reached a convergence state. If yes, the pre-training of the anti-counterfeiting model is completed, and in the case of no, the step for restarting. It returns to (S206).

The convergence state consists, for example, that the loss function value reaches a minimum or preset loss function value. If the loss function value does not satisfy the convergence state, it is possible to 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 of the anti-counterfeiting model. Although the number of supporting sample images of each category within each task is very small and therefore belongs to a small sample set, during the whole pre-training process a large number of samples (including many sample images of the existing categories and real face categories) It should be recognized that the use Also, the larger the number of samples, the better. The only difference is that many tasks and many small sample sets are partitioned, allowing the model to "learn how to learn" during the pre-training process, and the accuracy of the model can be improved.

In step S210, a small sample set of fake face images of a new category is acquired.

A small sample set of a new category of fake face images can be used as supporting sample images in the task. A task may contain one or more new categories, each new category corresponding to a small set of samples.

In step S212, the parameters of the pre-trained anti-spoofing model are updated with a small sample set of a new category of fake face images.

)의 지원 샘플 이미지를 사용함으로써 얼굴 위조 방지 모델의 매개변수(

)를 결정하는 프로세스와 유사하다. 예를 들어, 태스크(

)의 지원 샘플 이미지로서 하나 이상의 새로운 범주의 작은 샘플 세트 내 위조 얼굴 이미지의 경우에, 이러한 태스크(

)에 대응하는 얼굴 위조 방지 모델의 매개변수의 제3 경사도(

)가 계산되고, 이러한 태스크(

)에 대해 결정된 얼굴 위조 방지 모델의 매개변수는

으로 표현될 수 있으며, 여기서 α는 제1 스텝값을 나타낸다. 즉, 업데이트된 얼굴 위조 방지 모델의 매개변수는

이다. 쿼리 샘플 이미지를 사용하는 업데이트 프로세스는 더 이상 수행되지 않는다.Method (

) parameters of the anti-counterfeiting model (

) is similar to the process of determining For example, task(

), in the case of fake face images within a small sample set of one or more new categories as supporting sample images of

), the third slope of the parameter of the anti-counterfeiting model (

) is calculated, and these tasks (

), the parameters of the anti-counterfeiting model determined for

It can be expressed as , where α represents the first step value. That is, the parameters of the updated anti-counterfeiting model are

to be. The update process using the query sample image is no longer performed.

In step S214, the face image to be recognized is input into the face anti-counterfeiting model, after parameter update, so that it is determined whether the face image to be recognized is a fake face image.

The solution of the present disclosure may not only be applied to recognizing a fake face image of a new forgery method, but may also be applied to recognize a fake face image of a new model. For example, if it is only necessary to recognize the fake face image of the new spoofing method, the fake face image and the real face image of the existing spoofing method can be used as training samples during pre-training, and the training sample can be the same model, , the fake face image of the new forgery method also belongs to these models. For another example, if it is only necessary to recognize a fake face image of a new model, the fake face image and the real face image of the old model can be used as training samples during pre-training, and these fake face images of the old model are identical It may be forged by a forgery method, and a fake face image of a new model may also belong to this forgery method. If the training sample contains, during pre-training, fake face images of both different fake methods and different models, and real face images of one or more models, then the new fake face image is different from any fake method or any model in the training sample. can do.

For example, in some embodiments, for the fake face image of the new forgery method, a small sample set of the fake face image of the new forgery method is obtained; The parameters of the anti-counterfeiting model are updated with a small sample set of fake face images of the new forgery method; parameters of the pre-trained anti-spoofing model are determined based on a plurality of sets of parameters, each of which is a parameter of a plurality of sets of anti-spoofing models of the face determined based on a plurality of tasks; each task includes a small sample set of real face images and a small sample set of fake face images in each of one or more existing spoofing methods; The face image to be recognized is input into the anti-counterfeiting model after parameter update so that it is determined whether the face image to be recognized is a forgery face image.

For example, in some embodiments, for the fake face image of the new model, a small sample set of the fake face image of the new model is obtained; The parameters of the anti-counterfeiting face model are updated with a small sample set of fake face images of the new model; parameters of the pre-trained anti-spoofing model are determined based on a plurality of sets of parameters, each of which is a parameter of a plurality of sets of anti-spoofing models of the face determined based on a plurality of tasks; each task includes a small sample set of a fake face image of each of the one or more existing models and a small sample set of a real face image of each of the one or more existing models; The face image to be recognized is input into the anti-counterfeiting model after parameter update so that it is determined whether the face image to be recognized is a forgery face image.

The pre-training process is similar to that in the previous embodiment. For an environment where the training sample contains different spoof methods and fake face images of different models, for example, if there are 3 known models and 3 known spoofing methods, up to 9 different existing categories of fake face images there may be Also, if there are also real face images of 3 models, the generated task can contain up to 12 small sample sets.

In the embodiment described above, the pre-trained anti-spoofing face model is trained on a large number of small sample sets consisting of one or more existing models or fake face images of one or more existing spoofing methods or real face images of multiple existing models, The parameters of the multiple groups of the anti-spoofing model are each determined by multiple tasks, and then the parameters of the face anti-spoofing model are re-determined based on the multiple sets of parameters during the training process. This training process consists of two steps: the first step of training a face anti-spoofing model with a large number of small sample sets of fake face images from an existing category and real face images from an existing category, respectively, where the model learns the features of the small sample set. the first step, which is a step of learning to identify a fake face and a real face of different models; and a second step of performing aggregation and re-updating the parameters of the anti-counterfeiting model based on a plurality of training results, wherein the model learns how to learn features of a small sample set, or counterfeit faces of different models and the second step, which is a step of learning how to learn to identify a real face.

Thus, a pre-trained anti-counterfeiting model masters how to learn a small sample set, or how to learn to identify fake and real faces of different models and obtain favorable initialization parameters. When the fake face image of the new model belongs to the state of the small sample set, the face anti-counterfeiting model is updated. The face anti-counterfeiting model, which has benefited from the previously learned fast learning ability, can also perform fast convergence and fast learning of fake facial features of a new model or new forgery method, and timely and effective update of the face anti-counterfeiting model, A fake face image of a new model or a new forgery method can be recognized effectively in a timely manner, and an attack from a forged image of a new model can be effectively defended in a timely manner.

The present disclosure also provides a fake face recognition device, which will be described later in conjunction with FIG. 3 .

3 is a schematic structural diagram of a forged face recognition apparatus according to some embodiments of the present disclosure. As shown in FIG. 3 , the apparatus 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 a new category of fake face images.

The model adjustment model 304 is configured to update the parameters of the pre-trained anti-spoofing model of the face with a small sample set of a new category of fake face images; The parameters of the pre-trained anti-counterfeiting model are determined based on a plurality of sets of parameters, wherein the plurality of sets of parameters are respectively determined based on a plurality of tasks. ego; Each task includes a small sample set of fake face images and a small sample set of real face images in each of one or more existing categories.

In some embodiments, the model adjustment model 304 determines the parameters of the anti-counterfeiting model based on a plurality of fake face images in a small sample set of fake face images of a new category and a classification label corresponding to each of the fake face images. configured to update; When the anti-counterfeiting face model is a classification model, each classification label is a first classification label, configured to label a category of a forged face image corresponding to the classification label, wherein the category includes at least one of a forgery method or a model. . Alternatively, when the anti-spoofing face model is a regression model, each classification label is a second classification label, configured to label a fake face image corresponding to the classification label as a positive sample or a negative sample.

In some embodiments, the model adjustment model 304 inputs a plurality of fake face images in a small sample set of a new category of fake face images into the face anti-counterfeiting model as support sample images for one task; determine a gradient of a parameter of the anti-counterfeiting model, based on the classification label corresponding to each of the forged face images and the output value of the anti-counterfeiting model for each of the forged face images; and update the parameter of the anti-counterfeiting model according to the gradient of the parameter of the anti-counterfeiting model.

In some embodiments, when the anti-counterfeiting face model is a classification model, the output value of the anti-counterfeiting model for each forged face image is the probability that each forged face image belongs to a different category, and the model adjustment model 304 is determine a first inclination of a parameter of the anti-counterfeiting model, based on the first classification label corresponding to each of the forged face images and the probability that each of the forged face images belong to different categories, respectively; When the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each forged face image is the probability that each forged face image belongs to a positive sample, and the model adjustment model 304 corresponds to each of the forged face images. and determine a second gradient of the parameter of the anti-spoofing model, based on the second classification label and the probability that each fake face image belongs to the positive sample.

The face recognition module 306 is configured to input the face image to be recognized into the anti-counterfeiting face model after parameter update, so that it is determined whether the face image to be recognized is a fake face image.

In some embodiments, apparatus 30 further includes a pre-training module 308 . The pre-training module 308 is configured to divide the fake face image and the real face image of an existing category into a plurality of tasks, each task including a plurality of schemes, each scheme being a plurality of real faces of one model image or a plurality of fake face images of an existing category, each task including a first preset number of existing categories, each manner comprising a second preset number of a small sample set consisting of supporting sample images a third preset number of real face images or fake face images as real face images or fake face images and query sample images, and a fourth preset number of tasks are selected to form a batch each time, whereby the anti-counterfeiting face model a plurality of sets of parameters are respectively determined based on a plurality of tasks of the batch, the parameters of the face anti-counterfeiting model, the face anti-counterfeiting model reaches a convergence state to complete the pre-training of the face anti-counterfeiting model. It is re-determined based on multiple sets of parameters until

In some embodiments, the pre-training module 308 is configured to determine, for each task in the batch, a third gradient corresponding to the task based on each support sample image and a classification label corresponding to each support sample image; a parameter set of the anti-counterfeiting model is determined based on the third gradient; a fourth gradient corresponding to the batch is determined based on each of the plurality of tasks in the batch, a query sample image corresponding to the plurality of tasks and a classification label corresponding to each of the query sample images; The parameters of the anti-counterfeiting model are re-determined based on the fourth gradient corresponding to the arrangement.

In some embodiments, when the anti-counterfeiting face model is a classification model, each of the classification labels is a first classification label configured to label a category of a supporting sample image corresponding to the classification label, wherein the category is at least one of a forgery method or model includes one Alternatively, when the anti-counterfeiting model is a regression model, each of the classification labels is a second classification label, configured to label a supporting sample image corresponding to the classification label as a positive sample or a negative sample.

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

In some embodiments, the different pre-existing categories of spoofed face images differ in at least one of a spoof method or model, wherein the spoofing method comprises a photograph, a video or a mask, and the model is a depth model, a near-infrared model, or a red-green-blue (RGB) models; Real face images in different tasks have the same or different models; The new category of fake face images is different from the existing images in at least one of the forgery methods and models, and the existing images include fake face images and real face images of the old categories.

The fake face recognition apparatus according to the embodiment of the present disclosure may be performed by various computing devices or computer systems to be described later in conjunction with FIGS. 4 and 5 .

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

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

5 is a schematic structural diagram of a fake face recognition apparatus 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 a memory 410 and a processor 420 , respectively. It may further include an I/O interface 530 , a network interface 540 , a storage interface 550 , and the like. Memory 510 and processor 520 , as well as these interfaces 530 , 540 , 550 may be coupled between them, for example, via bus 560 . Here, I/O interface 530 provides a connection interface for input and output devices such as display, mouse, keyboard, and touch identification. The network interface 540 provides a connection interface for various networked devices. For example, it may be connected to a database server or a cloud storage server. The storage interface 550 provides a connection interface for an external storage device such as an SD card or a USB flash disk.

One of ordinary skill in the art will recognize that embodiments of the present disclosure may be provided as a method, system, or computer program product. Accordingly, the present disclosure may take the form of an all hardware embodiment, an all software embodiment, or a combination of software and hardware aspects. The present disclosure also relates to a form of a computer program product embodied in one or more computer usable non-transitory storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) embodied therein computer usable program code. can take

The present disclosure is described in conjunction with flowcharts and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present disclosure. It will be understood that each step and/or block of the flowcharts and/or block diagrams, as well as combinations of steps and/or blocks in the flowcharts and/or block diagrams, may be executed by computer program instructions. These computer program instructions are provided to a processor of a general purpose computer, special purpose computer, embedded processing machine, or other programmable data processing device for producing the machine, such that the instructions to be executed by the processor of the computer or other programmable data processing device are flow diagrams. One or more steps of and/or one or more blocks in the block diagram may produce a device for realizing a specified function.

Additionally, these computer program instructions may be stored in a computer readable memory that may direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory create a product comprising the instruction device. will produce The instruction device implements functions specified in one or more steps in the flowchart or one or more blocks in the block diagram.

In addition, these computer program instructions may be loaded into a computer or other programmable data processing device, such that a series of operational steps are performed on the computer or other programmable device to produce computer-executed processing, such that the computer or other programmable device The instructions executed on the device provide steps for implementing the specified functions in one or more steps of the flowchart and/or one or more blocks in the block diagrams.

The foregoing description is merely a preferred embodiment of the present disclosure, and is not intended to limit the present disclosure. Any correction, equivalent substitution, improvement, etc. made within the spirit and principle of the present disclosure should all fall within the protection scope of the present disclosure.

Claims (20)

A fake face recognition method comprising:
acquiring a small sample set of fake face images of a new category;
updating parameters of a pre-trained anti-counterfeiting model with the small sample set of new categories of spoofed face images, the parameters of the pre-trained anti-spoofing model being determined based on a plurality of sets of parameters and a plurality of sets of pre-trained parameters are each determined based on a plurality of tasks, each task comprising a small sample set of fake face images in each of one or more existing categories and an actual set of each of one or more existing categories. Contains a small sample set of face images - ; and
inputting the face image to be recognized into the anti-counterfeiting model after parameter update to determine whether the face image to be recognized is a fake face image;
A fake face recognition method comprising: According to claim 1,
The step of updating the parameters of the anti-counterfeiting model pre-trained with a small sample set of a new category of fake face images comprises:
updating the parameter of the anti-counterfeiting model, based on a fake face image in a small sample set of fake face images of the new category and a classification label corresponding to each of the fake face images;
when the anti-counterfeiting model is a classification model, the classification label is a first classification label, configured to label a category of a forged face image corresponding to the classification label, the category comprising at least one of a forgery method or a model; ; or
and when the anti-spoofing face model is a regression model, the classification label is a second classification label, configured to label a fake face image corresponding to the classification label as a positive sample or a negative sample. 3. The method of claim 2,
Based on a fake face image in the small sample set of the new category of fake face images and a classification label corresponding to each of the fake face images, the step of updating the parameter of the anti-counterfeiting model comprises:
inputting a fake face image in the small sample set of the new category of fake face images into the face anti-counterfeiting model as a supporting sample image for one task;
determining a gradient of the parameter of the anti-counterfeiting model based on a classification label corresponding to each of the forged face images and an output value of the anti-counterfeiting model for each of the forged face images; and
updating the parameter of the anti-counterfeiting model according to the gradient of the parameter of the anti-counterfeiting model
A fake face recognition method comprising: 4. The method of claim 3,
When the anti-counterfeiting model is a classification model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a different category,
The step of determining a gradient of the parameter of the anti-counterfeiting model on the basis of a classification label corresponding to each of the forged face images and an output value of the anti-counterfeiting model for each of the forged face images,
determining a first inclination of the parameter of the anti-counterfeiting model based on the first classification label corresponding to each of the forged face images and the probability that each of the forged face images belong to different categories, respectively; do;
or, when the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a positive sample,
The step of determining a gradient of the parameter of the anti-counterfeiting model on the basis of a classification label corresponding to each of the forged face images and an output value of the anti-counterfeiting model for each of the forged face images,
determining a second slope of the parameter of the anti-counterfeiting model based on the second classification label corresponding to each of the fake face images and the probability that each of the fake face images belongs to a positive sample , a fake face recognition method. According to claim 1,
dividing the fake face image of the existing category and the real face image of the existing category into a plurality of tasks -
Each task includes a plurality of manners, each manner includes a plurality of real face images of one model or a plurality of fake face images of an existing category, and each task includes a first preset number of existing categories including,
Each manner includes a second preset number of real face images or fake face images as a small sample set consisting of support sample images, and a third preset number of real face images or fake face images as query sample images; ;
the plurality of sets of parameters of the anti-counterfeiting model are respectively determined based on the plurality of tasks in a batch and the parameters of the anti-counterfeiting model are re-determined based on the plurality of sets of parameters preferably selecting a fourth preset number of tasks to form the batch each time; and
Completing the pre-training of the anti-counterfeiting model until the anti-counterfeiting model reaches a convergence state.
Further comprising, a fake face recognition method. 6. The method of claim 5,
wherein the plurality of sets of parameters of the anti-counterfeiting model are respectively determined based on the plurality of tasks of the batch, and the parameters of the anti-counterfeiting model are re-determined based on the plurality of sets of parameters. thing is,
determine a third gradient corresponding to the task based on each support sample image and a classification label corresponding to each support sample image, and based on the third gradient for each task in the batch, the face determining one set of parameters of the anti-counterfeiting model;
a fourth gradient corresponding to the batch, based on a parameter set determined by each of the plurality of tasks in the batch, the query sample image corresponding to the plurality of tasks, and a classification label corresponding to each of the query sample images to determine; and
re-determining the parameter of the anti-counterfeiting model based on the fourth gradient corresponding to the arrangement;
A fake face recognition method comprising: 7. The method of claim 6,
when the face anti-counterfeiting model is a classification model, each of the classification labels is a first classification label, configured to label a category of the supported sample image corresponding to the classification label, wherein the category is at least one of a forgery method or model including one; or
when the anti-counterfeiting model is a regression model, each of the classification labels is a second classification label, configured to label the supporting sample image corresponding to the classification label as a positive sample or a negative sample. According to claim 1,
An initial parameter of the anti-spoofing model before pre-training is determined according to a deep learning training method. 9. The method according to any one of claims 1 to 8,
Different existing categories of fake face images differ in at least one of a forgery method or model, wherein the forged method includes a photograph, a video or a mask, the model being a depth model, a near infrared model or a red-green- blue (RGB) models;
Real face images in different tasks have the same or different models;
The fake face image of the new category is different from an existing image in at least one of a forgery method or a model, and the existing image includes a fake face image of the old category and a real face image of the existing category. . A counterfeit facial recognition device, comprising:
a small sample acquisition module, configured to acquire a small sample set of a new category of fake face images;
a model tuning module, configured to update parameters of a pre-trained anti-counterfeiting model with the small sample set of a new category of spoof face images, wherein the parameters of the pre-trained anti-spoofing model are a plurality of sets of parameters and a plurality of sets of pre-trained parameters are each determined based on a plurality of tasks, each task comprising a small sample set of fake face images in each of one or more existing categories and each of one or more existing categories. Contains a small sample set of real face images - ; and
a face recognition module, configured to input the face image to be recognized into the anti-counterfeiting model after parameter update, to determine whether the face image to be recognized is a forged face image
A counterfeit facial recognition device, comprising: 11. The method of claim 10,
the model adjustment module is configured to update the parameter of the anti-counterfeiting model according to a classification label corresponding to each of the fake face image and the fake face image in the small sample of the fake face image of the new category; ,
when the anti-counterfeiting model is a classification model, the classification label is a first classification label, configured to label a category of a forged face image corresponding to the classification label, the category comprising at least one of a forgery method or a model; ; or
when the anti-spoofing face model is a regression model, the classification label is a second classification label, configured to label a fake face image corresponding to the classification label as a positive sample or a negative sample. 12. The method of claim 11,
The model adjustment module is configured to input a fake face image in the small sample set of the new category of fake face image into the face anti-counterfeiting model as a support sample image for one task, and a classification corresponding to each of the fake face images determine a gradient of the parameter of the anti-counterfeiting model according to the output value of the anti-counterfeiting model for each of the label and the face image, and based on the gradient of the parameter of the anti-counterfeiting model, and update the parameter of the face anti-spoofing model. 13. The method of claim 12,
When the anti-counterfeiting model is a classification model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a different category, and the model adjustment module is configured to: determine a first inclination of the parameter of the anti-spoofing model, based on the first classification label corresponding to each of the face images and the probability that each of the forged face images respectively belong to different categories;
or, when the anti-counterfeiting model is a regression model, the output value of the anti-counterfeiting model for each of the forged face images is a probability that each of the forged face images belongs to a positive sample, and the model adjustment module is configured to: and determine a second slope of the parameter of the anti-spoofing model, based on the second classification label corresponding to each of the fake face images and the probability that each of the fake face images belongs to a positive sample. recognition device. 11. The method of claim 10,
additionally including a pre-training module,
The pre-training module is
Splitting the fake face image of the existing category and the real face image of the existing category into multiple tasks;
Each task includes a plurality of manners, each manner includes a plurality of real face images of one model or a plurality of fake face images of an existing category, and each task includes a first preset number of existing categories including,
Each manner includes a second preset number of real face images or fake face images as a small sample set consisting of support sample images, and a third preset number of real face images or fake face images as query sample images; ;
so that the plurality of sets of parameters of the anti-counterfeiting model are respectively determined based on the plurality of tasks of a batch and the parameters of the anti-counterfeiting model are re-determined based on the plurality of sets of parameters, each time select a fourth preset number of tasks to form the batch;
and complete the pre-training of the anti-counterfeiting model until the anti-counterfeiting model reaches a convergence state. 15. The method of claim 14,
The pre-training module is configured to determine a third gradient corresponding to the task based on each support sample image and a classification label corresponding to each support sample image, and to determine the third gradient for each task in the batch. determine one set of parameters of the anti-counterfeiting model based on each of the plurality of tasks in the batch, the query sample image corresponding to the plurality of tasks, and a classification label corresponding to each of the query sample images determine a fourth gradient corresponding to the arrangement based on a parameter set determined by recognition device. 16. The method of claim 15,
when the face anti-counterfeiting model is a classification model, each of the classification labels is a first classification label, configured to label a category of the supported sample image corresponding to the classification label, wherein the category is at least one of a forgery method or model including one; or when the anti-counterfeiting model is a regression model, each of the classification labels is a second classification label, configured to label the support sample image corresponding to the classification label as a positive sample or a negative sample. . 11. The method of claim 10,
The initial parameters of the anti-counterfeiting model before pre-training are determined according to a deep learning training method. 18. The method according to any one of claims 10 to 17,
Different existing categories of fake face images differ in at least one of a forgery method or model, wherein the forged method comprises a photograph, a video or a mask, the model being a depth model, a near infrared model or a red-green-blue (RGB) model. includes a model;
Real face images in different tasks have the same or different models;
The fake face image of the new category is different from an existing image in at least one of a forgery method or a model, and the existing image includes a fake face image of the old category and a real face image of the existing category. . A counterfeit facial recognition device, comprising:
Memory; and
a processor coupled to the memory
including,
and the processor is configured to implement the fake face recognition method according to any one of claims 1 to 9, based on the instructions stored in the memory. A computer readable storage medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

Images