CN115170565A - Image fraud detection method and device based on automatic neural network architecture search - Google Patents
Image fraud detection method and device based on automatic neural network architecture search Download PDFInfo
- Publication number
- CN115170565A CN115170565A CN202211083963.3A CN202211083963A CN115170565A CN 115170565 A CN115170565 A CN 115170565A CN 202211083963 A CN202211083963 A CN 202211083963A CN 115170565 A CN115170565 A CN 115170565A
- Authority
- CN
- China
- Prior art keywords
- neural network
- image
- convolutional neural
- node
- network architecture
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
- G06T2207/30201—Face
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image fraud detection method and a device based on automatic neural network architecture search, which realize the improvement of sample data quality by adopting an integrated automatic image preprocessing method, help the rapid convergence in the subsequent fraud image detection model training process and reduce the non-relevant data interference; by using a data federal training mode, on the premise of protecting data privacy, existing fraud image sample data of different organizations are safely and effectively fused to construct a more accurate deep neural network model, and different types of fraud images are rapidly identified; the architecture optimization of the federated model is automatically completed by using a neural network architecture search strategy, the expert labor cost in the design and training stage of the deep neural network model is reduced, and meanwhile, the communication overhead and the computing resource overhead of repeated training of multiple rounds are also reduced.
Description
Technical Field
The invention relates to the technical fields of automatic machine learning, convolutional neural network, computer vision, federal learning and the like, in particular to an image fraud detection method and device based on automatic neural network architecture search.
Background
With the continuous popularization and application of the digitization technology in the financial industry, more and more financial business processes are evolving towards the direction of online, remote and intelligent processing. In the business processes of account opening, loan application, insurance purchase and the like of the financial institution, a client can convert various application materials into an electronic image to upload. However, according to the combing discovery of industry risk cases, a plurality of application materials have fraudulent behaviors such as blurred images, copied images, deep synthetic images and the like, which provides higher technical requirements for the financial industry access auditing system.
An accurate and efficient fraud image detection model needs to rely on enough sufficient training sample data, but often, each financial institution holds fewer data samples, which is not enough to meet the training requirements of a deep neural network model, but local data of each institution cannot be directly shared due to the protection of data privacy and the industrial supervision requirements.
In addition, the quality of sample data held by various organizations in different regions or different fields is uneven, the distribution of the sample data has Non-independent and same-distribution characteristics (Non-IID), the direct use of a predefined model architecture may not be an optimal neural network selection, since the distribution of the data is black-box and abstract for developers, in order to find a model architecture with better effect, the developers must design or select a plurality of preselected architectures, and then adapt to the data of the Non-IID by remotely adjusting the hyper-parameters of the model, however, the computational overhead of the process becomes huge, and the development cycle of the whole project is also lengthened.
Disclosure of Invention
Aiming at the practical situation of data barriers among mechanisms, the distribution characteristics of samples, the quality level of the samples, and pain spots and difficulties in the prior art, the invention provides an image fraud detection method and device based on automatic neural network architecture search, which are used for breaking the data barriers of the mechanisms, reducing the labor participation cost of experts, automatically completing the preprocessing of images of financial admission materials, and quickly searching the distributed deep neural network architecture most suitable for the current sample data so as to improve the detection effect of fraud images.
The purpose of the invention is realized by the following technical scheme:
according to a first aspect of the present specification, there is provided an image fraud detection method based on automatic neural network architecture search, comprising the steps of:
s1, preprocessing local image sample data by each organization, and maintaining the preprocessed sample data to a federal node server deployed by each organization to obtain a training data set and a verification data set;
s2, each node uses an initial convolutional neural network architecture to perform feature extraction on the input image, and training of an initial local fraud image detection model is completed;
s3, defining a learning process of the federal convolutional neural network model;
s4, determining a search space, and describing the search space of the convolutional neural network architecture into a directed acyclic graph formed by a plurality of characterization points and a plurality of transformation operations through an ordered sequence;
s5, searching a convolutional neural network architecture by each node, searching for optimal convolutional neural network weight and convolutional neural network architecture, and sending to a central server;
s6, the central server side aggregates the optimal convolutional neural network weight and convolutional neural network architecture of each node to obtain the globally optimal convolutional neural network weight and convolutional neural network architecture, and returns the weights and architectures to each node;
s7, updating the convolutional neural network weight and convolutional neural network architecture of the convolutional neural network model of each node, performing multi-round search until the evaluation index of the current convolutional neural network model reaches the expectation, finishing the training of the fraud image detection model by using the optimal convolutional neural network architecture by each node, and encrypting and sending a training intermediate value to a central server;
s8, the central server receives the training intermediate values from the nodes, jointly constructs an optimal global fraud image detection model and distributes the optimal global fraud image detection model to the nodes;
and S9, after each node obtains the optimal global fraud image detection model, deploying the model to a service system for online detection.
Further, in step S1, the local image is an image of a financial admission material in a financial admission auditing system, and sample data of the local image has a labeled fraud type tag, including copied image fraud, deep synthetic image fraud, blurred image fraud, and achromatic original fraud.
Further, in step S1, each organization performs integrated automated preprocessing on the local image sample data, including sample deduplication, invalid sample deletion, image size unification, automatic data desensitization, and image data normalization processing.
Further, the unified image sizes are specifically: carrying out automatic unified scaling on an original image, and resampling a pixel area relation;
the automatic data desensitization is specifically: masking sensitive data in a local image, performing character recognition on a special number by adopting an OCR (optical character recognition), and automatically masking the sensitive data if the sensitive data is judged to be the sensitive data; firstly, face information is subjected to face detection by adopting a target recognition algorithm, and masking processing is automatically performed if sensitive data is judged;
the image data normalization specifically comprises: and uniformly converting the value range of each pixel value of the original image into the value range of 0,1 in batches.
Further, the initial convolutional neural network architecture comprises the following three network elements:
the first network unit comprises a convolutional layer, the convolutional layer consists of 32 convolutional kernels with the size of 7*7 and the step length of 1 and a pooling layer with the step length of 1, and then a unified feature map is obtained through twice maximum value combination;
the second network unit comprises three convolutional layers which are connected in sequence, and each convolutional layer consists of 16 convolutional cores with the size of 3*3 and the step size of 1 and a pooling layer with the step size of 4;
the third network unit comprises a Dropout layer, two full connection layers and an output layer which are sequentially connected, the activation function of the output layer is a Softmax function, and the identification result of the image category, including normal images and various suspected fraud images, is fed back through the full connection layers and the output layer.
Further, step S3 specifically includes:
if K federal nodes interact with a central server, an optimal global model is finally found
The optimizing process is as follows:
wherein the function
The loss expectation under the data distribution of node i is expressed as follows:
wherein the content of the first and second substances,
the weights of the convolutional neural network are represented,
it is shown that the convolutional neural network architecture,
and
corresponding to the current global model
，
Is sample data randomly extracted according to the data distribution of the node i,
is corresponding to sample data
And a global model
Is used to determine the loss function of (c),
is shown in
And
under the condition of (1), loss function
(iii) a desire;
finding an expected average for the losses of all federal nodes
Corresponds to the minimum value of
And
to obtain an optimal global model
。
Further, in step S4, each token point is calculated from the previous token point through a transformation operation, and is represented as follows:
wherein a represents the current token point, b represents the next token point,
for the state of the current token point,
for the state of the next token point to be characterized,
indicating a transformation operation between a and b,
is shown in
The result after the conversion operation is carried out under the state;
all candidate transformation operations are expressed as a group of vectors, and Softmax processing is carried out on the selected process of the candidate transformation operations, which is expressed as follows:
wherein Q is a set of all candidate operations, Q is one operation in the set Q, i is a node currently participating in architecture search among K federated nodes,
for the convolutional neural network architecture of node i,
for the convolutional neural network architecture of node i under operation q,
is a vector of node i between two token points a, b,
for a vector between two token points a, b for node i under operation q,
representing the mapped probability distribution of all candidate transformation operation vectors.
Further, in step S5, each node finds the optimal convolutional neural network weight
And convolutional neural network architecture
The optimizing process is as follows:
wherein the content of the first and second substances,
representing a loss of the local training data set,
representing a loss of the local validation data set,
respectively representing a representation of a convolutional neural networkWeight of
And convolutional neural network architecture
The computation of the gradient in each direction in space,
representing the learning rate in the optimization process,
representing the hyper-parameters used to equalize the loss of the training data set and the loss of the validation data set.
Further, in step S6, after receiving the optimal convolutional neural network weight and convolutional neural network architecture of each node, the central server aggregates the weights in a weighted average manner to obtain the globally optimal convolutional neural network weight and convolutional neural network architecture, where the weight is a federal influence factor of each node and is determined by a ratio of the local data volume of the node to the total data volume of all nodes.
According to a second aspect of the present specification, there is provided an image fraud detection apparatus based on automatic neural network architecture search, comprising a memory and one or more processors, wherein the memory stores executable code, and the processors execute the executable code to implement the image fraud detection method based on automatic neural network architecture search according to the first aspect.
The invention has the beneficial effects that: the invention provides an image fraud detection method and device based on automatic neural network architecture search, firstly, an integrated automatic financial access material image preprocessing method is adopted to realize the improvement of sample data quality, and the rapid convergence and the reduction of non-relevant data interference in the subsequent fraud image detection model training process are facilitated; secondly, by using a data federal training mode, on the premise of protecting data privacy, existing fraud image sample data of different organizations are safely and effectively fused to construct a more accurate deep neural network model, and different types of fraud images are rapidly identified; and thirdly, the architecture optimization of the federated model is automatically completed by using a neural network architecture search strategy, the expert labor cost in the deep neural network model design and training stage is reduced, and meanwhile, the communication overhead and the computing resource overhead of repeated training for multiple rounds are also reduced.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a flow chart of an exemplary embodiment of an integrated automated pre-processing of sample data.
FIG. 2 is a flow chart of search and joint training of an automatic neural network architecture provided by an exemplary embodiment.
Fig. 3 is a schematic diagram of characterization points and candidate operations of a network element according to an exemplary embodiment.
Fig. 4 is a schematic diagram of an interaction process between the central server and each federated node provided in an exemplary embodiment.
FIG. 5 is a schematic diagram of an optimal global fraud image detection model deployment according to an exemplary embodiment.
Fig. 6 is a block diagram of an image fraud detection apparatus based on an automatic neural network architecture search according to an exemplary embodiment.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below. It is to be understood that the described embodiments are merely exemplary of the invention, and not restrictive of the full scope of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The terminology used in the embodiments of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in the examples of this application and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
The embodiment of the invention provides an image fraud detection method based on automatic neural network architecture search, which comprises the following steps:
firstly, each mechanism carries out integrated automatic preprocessing on local image sample data, including sample duplicate removal, invalid sample deletion, image size unification, automatic data desensitization, image data normalization processing and the like, and the steps are shown in figure 1.
Specifically, the local image is a financial admission material image in a financial admission auditing system, and the sample data of the local image has a labeled fraud category label, such as copied image fraud, deep synthetic image fraud, blurred image fraud, achromatic original fraud and the like.
Specifically, the unified image size refers to the automatic unified scaling of the original image by using a resize function in the Opencv tool library for image processing, in this embodiment, the image size is uniformly converted into 512 × 384, and the INTER _ AREA interpolation method is used to perform pixel region relation resampling, so as to reduce distortion that may be caused by scaling.
Specifically, the automatic data desensitization refers to masking sensitive data in a local image, such as a special number and face information, and the masking refers to performing all-0 processing on pixel values of a sensitive area. Firstly, character recognition is carried out on the special number by adopting an OCR (optical character recognition), and masking processing is automatically carried out when sensitive data are judged; and firstly, carrying out face detection on the face information by adopting a target recognition algorithm, and automatically carrying out masking processing on the face information if the face information is judged to be sensitive data.
Specifically, the image data normalization processing refers to uniformly converting the value range of each pixel value of the original image from [0,255] to [0,1] in batches so as to accelerate the convergence rate during deep neural network training.
And secondly, maintaining the sample data after the pretreatment to a federal node server deployed by each organization, and dividing the sample data into a training data set and a verification data set according to the proportion of 8:2.
And thirdly, each node uses the initial convolutional neural network architecture to extract the characteristics of the input image, and training of an initial local fraud image detection model is completed.
The initial convolutional neural network architecture comprises the following three network elements:
the first network unit comprises a convolutional layer which is composed of 32 convolutional kernels with the size of 7*7 and the step size of 1 and a pooling layer with the step size of 1, and then the unified feature map is obtained through two times of maximum value combination.
The second network unit comprises three convolutional layers connected in sequence, and each convolutional layer consists of 16 convolutional cores with the size of 3*3 and the step size of 1 and a pooling layer with the step size of 4.
The third network unit comprises a Dropout layer, two full connection layers and an output layer which are sequentially connected, the activation function of the output layer is a Softmax function, and the identification result of the image category, including normal images and various suspected fraud images, is fed back through the full connection layers and the output layer.
And fourthly, defining a learning process of the federated convolutional neural network model.
If K federal nodes interact with a central server, an optimal global model is finally found
The optimizing process is as follows:
wherein the function
The loss expectation under the data distribution of node i is expressed as follows:
wherein the content of the first and second substances,
the weights of the convolutional neural network are represented,
it is shown that the convolutional neural network architecture,
and
corresponding to the current global model
，
Is sample data randomly extracted according to the data distribution of the node i,
is corresponding to the sample data
And a global model
Is used to determine the loss function of (c),
is shown in
And
under the condition of (2), a loss function
(iii) a desire;
finding an expected average for the losses of all federal nodes
Corresponds to the minimum value of
And
to obtain an optimal global model
。
Since a node may be an organization from different regions or different domains, its distribution of held sample data has the Non-IID characteristic. For sample data, pair
And
and meanwhile, automatic architecture search optimization is carried out to obtain an optimal convolutional neural network architecture, so that a fraud image detection model with higher accuracy and better robustness is trained. The main process of search and joint training of the automatic neural network architecture is shown in fig. 2.
And fifthly, determining a search space.
The convolutional neural network architecture comprises a plurality of network units, each network unit comprises a plurality of characterization points representing the current image feature extraction effect, and the network units can also be understood as feature maps obtained after different transformation operations act. The dashed lines between token points represent candidate transform operations for transforming token points, including convolution, pooling, dropout, max merging, link skipping, etc., and the dashed lines represent possible different candidate transform operations, as shown in fig. 3. Each token point is calculated from the previous token point by a transformation operation, and is represented as follows:
wherein a represents the current token, b represents the next token, S represents the state of the token,
for the state of the current token point,
for the state of the next token point, O denotes the transform operation,
indicating a transformation operation between a and b,
is then indicated at
The result of the transformation operation in the state is the dashed line in fig. 3.
The search space of the convolutional neural network architecture is described as a directed acyclic graph composed of a plurality of token points and a plurality of transform operations through an ordered sequence, see fig. 3. In order to make the search space continuously differentiable, all candidate transformation operations can be represented as a set of vectors, and the selected process of the candidate transformation operations is subjected to Softmax processing, i.e. the set of vectors is mapped into a probability distribution, which is represented as follows:
wherein Q is a set of all candidate operations, Q is one operation in the set Q, and i is K federal sectionsThe nodes of the points that are currently participating in the architectural search,
for the convolutional neural network architecture of node i,
then for the convolutional neural network architecture for node i under operation q,
is a vector of node i between two token points a, b,
then for node i a vector between the two token points a and b under operation q,
and the probability distribution after mapping of all candidate transformation operation vectors is shown, and exp represents exponential function operation with a natural constant e as a base.
From the above equation, the probability weight of the possible candidate transformation operation between the two characteristic points a and b of the current node i can be represented by a vector
As a parametric representation, further, the optimization task of neural network architecture search can be simplified to describe parameters that look for a set of continuous variables, i.e.
。
And sixthly, searching the convolutional neural network architecture of the node.
After the search space is determined, the search work of the convolutional neural network architecture of each node is carried out, namely, the optimal convolutional neural network weight is found
And convolutional neural network architecture
The optimizing process is as follows:
wherein the content of the first and second substances,
representing a loss of the local training data set,
representing a loss of the local validation data set,
respectively represent a pair
And
the computation of the gradient in each direction in space,
represents the learning rate in the optimization process, and is usually in [0,1]]The value of medium, in this embodiment 0.01,
representing a hyperparameter for equalizing the loss of the training data set and the loss of the validation data set, is typically found in [0,1]]The value is 0.5 in this example.
Seventhly, all nodes optimize the weights of the convolutional neural network
And convolutional neural network architecture
Sending to the central server。
Eighthly, the central server receives the optimal convolutional neural network weight of each node
And convolutional neural network architecture
Then, aggregating in a weighted average mode to obtain the aggregated global convolutional neural network weight
And convolutional neural network architecture
Expressed as follows:
wherein the content of the first and second substances,
to
The federate influence factor is determined by the local data volume of the node, the local data quality, the federate contribution degree of the previous round and the like, and must meet the requirement
. In the embodiment, the federal influence factor is determined by adopting the proportion of the local data volume of the node to the total data volume of all the nodes.
Ninth step, the central server side will update
And
sent back to the nodes, each node updating its respective convolutional neural network model
And
and preparing to perform the next round of convolutional neural network architecture search, namely repeating the sixth step, the seventh step and the eighth step until the evaluation indexes of the current convolutional neural network model reach expectations, wherein the evaluation indexes comprise precision rate, recall rate, F1 value, AUC value and the like. The interaction process of the central server and each federal node is shown in fig. 4.
And step ten, completing automatic searching of the federal convolutional neural network architecture by each node, completing respective fraud image detection model training by using the optimal convolutional neural network architecture, and encrypting and sending a training intermediate value to the central server.
And eleventh, the central server receives the training intermediate values from the nodes, jointly constructs an optimal global fraud image detection model of an optimal convolutional neural network architecture, and distributes the optimal global fraud image detection model to the nodes.
And step twelve, after each node obtains the optimal global fraud image detection model, deploying the model to a service system for online detection, and carrying out automatic auditing work of the financial access material image, as shown in figure 5.
Corresponding to the embodiment of the image fraud detection method based on the automatic neural network architecture search, the invention also provides an embodiment of an image fraud detection device based on the automatic neural network architecture search.
Referring to fig. 6, an image fraud detection apparatus based on automatic neural network architecture search according to an embodiment of the present invention includes a memory and one or more processors, where the memory stores executable codes, and the processors execute the executable codes to implement the image fraud detection method based on automatic neural network architecture search in the foregoing embodiment.
The embodiment of the image fraud detection apparatus based on the automatic neural network architecture search in the invention can be applied to any device with data processing capability, such as a computer or other devices or apparatuses. The device embodiments may be implemented by software, or by hardware, or by a combination of hardware and software. The software implementation is taken as an example, and as a logical device, the device is formed by reading corresponding computer program instructions in the nonvolatile memory into the memory for running through the processor of any device with data processing capability. From a hardware aspect, as shown in fig. 6, a hardware structure diagram of any device with data processing capability where the image fraud detection apparatus based on automatic neural network architecture search is located in the present invention is shown, except for the processor, the memory, the network interface, and the nonvolatile memory shown in fig. 6, any device with data processing capability where the apparatus is located in the embodiment may also include other hardware according to the actual function of the any device with data processing capability, which is not described again.
The implementation process of the functions and actions of each unit in the above device is specifically described in the implementation process of the corresponding step in the above method, and is not described herein again.
For the device embodiments, since they substantially correspond to the method embodiments, reference may be made to the partial description of the method embodiments for relevant points. The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of the invention. One of ordinary skill in the art can understand and implement without inventive effort.
An embodiment of the present invention further provides a computer-readable storage medium, on which a program is stored, where the program, when executed by a processor, implements the image fraud detection method based on automatic neural network architecture search in the foregoing embodiments.
The computer readable storage medium may be an internal storage unit, such as a hard disk or a memory, of any data processing device described in any previous embodiment. The computer readable storage medium may also be any external storage device of a device with data processing capabilities, such as a plug-in hard disk, a Smart Media Card (SMC), an SD Card, a Flash memory Card (Flash Card), etc. provided on the device. Further, the computer readable storage medium may include both an internal storage unit and an external storage device of any data processing capable device. The computer-readable storage medium is used for storing the computer program and other programs and data required by the arbitrary data processing-capable device, and may also be used for temporarily storing data that has been output or is to be output.
It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a … …" does not exclude the presence of another identical element in a process, method, article, or apparatus that comprises the element.
The foregoing description has been directed to specific embodiments of this disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The terminology used in the description of the one or more embodiments is for the purpose of describing the particular embodiments only and is not intended to be limiting of the description of the one or more embodiments. As used in one or more embodiments of the present specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.
It should be understood that although the terms first, second, third, etc. may be used in one or more embodiments of the present description to describe various information, such information should not be limited to these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of one or more embodiments herein. The word "if," as used herein, may be interpreted as "at … …" or "at … …" or "in response to a determination," depending on the context.
The above description is only for the purpose of illustrating the preferred embodiments of the one or more embodiments of the present disclosure, and is not intended to limit the scope of the one or more embodiments of the present disclosure, and any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the one or more embodiments of the present disclosure should be included in the scope of the one or more embodiments of the present disclosure.
Claims (10)
1. An image fraud detection method based on automatic neural network architecture search is characterized by comprising the following steps:
s1, preprocessing local image sample data by each organization, and maintaining the preprocessed sample data to a federal node server deployed by each organization to obtain a training data set and a verification data set;
s2, each node uses an initial convolutional neural network architecture to perform feature extraction on the input image, and training of an initial local fraud image detection model is completed;
s3, defining a learning process of the federal convolutional neural network model;
s4, determining a search space, and describing the search space of the convolutional neural network architecture into a directed acyclic graph formed by a plurality of characterization points and a plurality of transformation operations through an ordered sequence;
s5, searching a convolutional neural network architecture by each node, searching for optimal convolutional neural network weight and convolutional neural network architecture, and sending to a central server;
s6, the central server side aggregates the optimal convolutional neural network weight and convolutional neural network architecture of each node to obtain the globally optimal convolutional neural network weight and convolutional neural network architecture, and returns the weights and architectures to each node;
s7, updating the convolutional neural network weight and convolutional neural network architecture of the convolutional neural network model of each node, performing multi-round search until the evaluation index of the current convolutional neural network model reaches the expectation, finishing the training of the fraud image detection model by using the optimal convolutional neural network architecture by each node, and encrypting and sending a training intermediate value to a central server;
s8, the central server receives the training intermediate values from the nodes, jointly constructs an optimal global fraud image detection model and distributes the optimal global fraud image detection model to the nodes;
and S9, after each node obtains the optimal global fraud image detection model, deploying the model to a service system for online detection.
2. The method according to claim 1, wherein in step S1, the local image is an image of financial admission material in a financial admission auditing system, and the local image sample data has labeled fraud category labels, including copied image fraud, deep synthetic image fraud, blurred image fraud, and achromatic original fraud.
3. The method according to claim 1, wherein in step S1, each institution performs integrated automated pre-processing on the local image sample data, including sample deduplication, invalid sample deletion, image size unification, automatic data desensitization, and image data normalization.
4. The method according to claim 3, wherein the image sizes are unified by: carrying out automatic unified zooming on an original image, and resampling the pixel area relation;
the automatic data desensitization is specifically: masking sensitive data in a local image, performing character recognition on a special number by adopting OCR (optical character recognition), and automatically masking the sensitive data if the sensitive data is judged to be the sensitive data; firstly, carrying out face detection on face information by adopting a target recognition algorithm, and automatically carrying out mask processing if the face information is judged to be sensitive data;
the image data normalization specifically comprises: and uniformly converting the value range of each pixel value of the original image into the value range of 0,1 in batches.
5. The method of claim 1, wherein the initial convolutional neural network architecture comprises three network elements:
the first network unit comprises a convolutional layer, the convolutional layer consists of 32 convolutional kernels with the size of 7*7 and the step length of 1 and a pooling layer with the step length of 1, and then a unified feature map is obtained through twice maximum value combination;
the second network unit comprises three convolutional layers which are connected in sequence, and each convolutional layer consists of 16 convolutional cores with the size of 3*3 and the step size of 1 and a pooling layer with the step size of 4;
the third network unit comprises a Dropout layer, two full connection layers and an output layer which are sequentially connected, the activation function of the output layer is a Softmax function, and the identification result of the image category, including normal images and various suspected fraud images, is fed back through the full connection layers and the output layer.
6. The method according to claim 1, wherein step S3 is specifically:
if K federal nodes interact with a central server, an optimal full node is finally foundOffice model
The optimizing process is as follows:
wherein the function
The loss expectation under the data distribution of node i is expressed as follows:
wherein the content of the first and second substances,
the weights of the convolutional neural network are represented,
it is shown that the convolutional neural network architecture,
and
corresponding to the current global model
，
Is sample data randomly extracted according to the data distribution of the node i,
is corresponding to the sample data
And a global model
Is used to determine the loss function of (c),
is shown in
And
under the condition of (1), loss function
(iii) a desire;
finding an expected average for the losses of all federal nodes
Corresponds to the minimum value of
And
to obtain an optimal global model
。
7. The method according to claim 1, wherein in step S4, each token is calculated from the previous token by a transformation operation, and is represented as follows:
wherein a represents the current token point, b represents the next token point,
for the state of the current token point,
for the state of the next token point,
indicating a transformation operation between a and b,
is shown in
The result after the conversion operation is carried out under the state;
representing all candidate transformation operations as a group of vectors, and performing Softmax processing on the selected process of the candidate transformation operations, which is represented as follows:
wherein Q is a set of all candidate operations, Q is one operation in the set Q, i is a node currently participating in architecture search among K federated nodes,
is a convolutional neural network architecture of the node i,
for the convolutional neural network architecture of node i under operation q,
is a vector of node i between two token points a, b,
for a vector between two token points a, b for node i under operation q,
representing the probability distribution after mapping of all candidate transformation operation vectors.
8. The method of claim 1, wherein in step S5, each node finds an optimal convolutional neural network weight
And convolutional neural network architecture
The optimizing process is as follows:
wherein the content of the first and second substances,
representing a loss of the local training data set,
representing a loss of the local validation data set,
respectively represent weights for representing convolutional neural networks
And convolutional neural network architecture
The computation of the gradient in each direction in space,
represents the learning rate in the optimization process,
representing the hyper-parameters used to equalize the loss of the training data set and the loss of the validation data set.
9. The method according to claim 1, wherein in step S6, after receiving the optimal convolutional neural network weight and convolutional neural network architecture of each node, the central server performs aggregation in a weighted average manner to obtain the globally optimal convolutional neural network weight and convolutional neural network architecture, wherein the weight is a federal influence factor of each node and is determined by a ratio of the local data volume of the node to the total data volume of all nodes.
10. An image fraud detection apparatus based on automatic neural network architecture search, comprising a memory and one or more processors, the memory having stored therein executable code, wherein the processors, when executing the executable code, are configured to implement the image fraud detection method based on automatic neural network architecture search according to any one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211083963.3A CN115170565B (en) | 2022-09-06 | 2022-09-06 | Image fraud detection method and device based on automatic neural network architecture search |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211083963.3A CN115170565B (en) | 2022-09-06 | 2022-09-06 | Image fraud detection method and device based on automatic neural network architecture search |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115170565A true CN115170565A (en) | 2022-10-11 |
| CN115170565B CN115170565B (en) | 2022-12-27 |
Family
ID=83480390
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211083963.3A Active CN115170565B (en) | 2022-09-06 | 2022-09-06 | Image fraud detection method and device based on automatic neural network architecture search |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115170565B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116758390A (en) * | 2023-08-14 | 2023-09-15 | 腾讯科技（深圳）有限公司 | Image data processing method, device, computer equipment and medium |
| CN117373066A (en) * | 2023-12-07 | 2024-01-09 | 华侨大学 | Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325619A (en) * | 2018-12-15 | 2020-06-23 | 深圳先进技术研究院 | Credit card fraud detection model updating method and device based on joint learning |
| CN111814966A (en) * | 2020-08-24 | 2020-10-23 | 国网浙江省电力有限公司 | Neural network architecture searching method, neural network application method, device and storage medium |
| CN113239972A (en) * | 2021-04-19 | 2021-08-10 | 温州医科大学 | Artificial intelligence auxiliary diagnosis model construction system for medical images |
| CN113361680A (en) * | 2020-03-05 | 2021-09-07 | 华为技术有限公司 | Neural network architecture searching method, device, equipment and medium |
| US20210374502A1 (en) * | 2020-06-01 | 2021-12-02 | Nvidia Corporation | Technique to perform neural network architecture search with federated learning |
| CN113962359A (en) * | 2021-09-30 | 2022-01-21 | 华东师范大学 | Self-balancing model training method based on federal learning |
| US20220092416A1 (en) * | 2018-12-27 | 2022-03-24 | Google Llc | Neural architecture search through a graph search space |
| CN114358912A (en) * | 2021-11-17 | 2022-04-15 | 北京交通大学 | Risk weight fusion anomaly detection method based on federal learning |
-
2022
- 2022-09-06 CN CN202211083963.3A patent/CN115170565B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325619A (en) * | 2018-12-15 | 2020-06-23 | 深圳先进技术研究院 | Credit card fraud detection model updating method and device based on joint learning |
| US20220092416A1 (en) * | 2018-12-27 | 2022-03-24 | Google Llc | Neural architecture search through a graph search space |
| CN113361680A (en) * | 2020-03-05 | 2021-09-07 | 华为技术有限公司 | Neural network architecture searching method, device, equipment and medium |
| US20210374502A1 (en) * | 2020-06-01 | 2021-12-02 | Nvidia Corporation | Technique to perform neural network architecture search with federated learning |
| CN111814966A (en) * | 2020-08-24 | 2020-10-23 | 国网浙江省电力有限公司 | Neural network architecture searching method, neural network application method, device and storage medium |
| CN113239972A (en) * | 2021-04-19 | 2021-08-10 | 温州医科大学 | Artificial intelligence auxiliary diagnosis model construction system for medical images |
| CN113962359A (en) * | 2021-09-30 | 2022-01-21 | 华东师范大学 | Self-balancing model training method based on federal learning |
| CN114358912A (en) * | 2021-11-17 | 2022-04-15 | 北京交通大学 | Risk weight fusion anomaly detection method based on federal learning |
Non-Patent Citations (2)
| Title |
|---|
| BARRET ZOPH等: "NEURAL ARCHITECTURE SEARCH WITH REINFORCEMENT LEARNING", 《ARXIV》 * |
| 易东义等: "基于图卷积神经网络的医保欺诈检测算法", 《计算机应用》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116758390A (en) * | 2023-08-14 | 2023-09-15 | 腾讯科技（深圳）有限公司 | Image data processing method, device, computer equipment and medium |
| CN116758390B (en) * | 2023-08-14 | 2023-10-20 | 腾讯科技（深圳）有限公司 | Image data processing method, device, computer equipment and medium |
| CN117373066A (en) * | 2023-12-07 | 2024-01-09 | 华侨大学 | Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method |
| CN117373066B (en) * | 2023-12-07 | 2024-03-12 | 华侨大学 | Pedestrian re-identification method and system based on Yun Bian searching federal deep learning method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115170565B (en) | 2022-12-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115170565B (en) | Image fraud detection method and device based on automatic neural network architecture search | |
| CN111291183B (en) | Method and device for carrying out classification prediction by using text classification model | |
| Kulhánek et al. | Viewformer: Nerf-free neural rendering from few images using transformers | |
| CN113420729B (en) | Multi-scale target detection method, model, electronic equipment and application thereof | |
| EP3029606A2 (en) | Method and apparatus for image classification with joint feature adaptation and classifier learning | |
| CN112541458B (en) | Domain self-adaptive face recognition method, system and device based on meta learning | |
| CN110852881B (en) | Risk account identification method and device, electronic equipment and medium | |
| CN113128478B (en) | Model training method, pedestrian analysis method, device, equipment and storage medium | |
| CN111882426B (en) | Method, device, equipment and storage medium for training business risk classifier | |
| KR20220126614A (en) | Method for processing image for registration | |
| CN116580257A (en) | Feature fusion model training and sample retrieval method and device and computer equipment | |
| CN111626311B (en) | Heterogeneous graph data processing method and device | |
| CN112749737A (en) | Image classification method and device, electronic equipment and storage medium | |
| CN112989182A (en) | Information processing method, information processing apparatus, information processing device, and storage medium | |
| CN112183303A (en) | Transformer equipment image classification method and device, computer equipment and medium | |
| Ma et al. | Multi-view clustering based on view-attention driven | |
| CN116740422A (en) | Remote sensing image classification method and device based on multi-mode attention fusion technology | |
| CN110991247A (en) | Electronic component identification method based on deep learning and NCA fusion | |
| CN115565115A (en) | Outfitting intelligent identification method and computer equipment | |
| CN112712695B (en) | Traffic flow prediction method, device and storage medium | |
| CN116090538A (en) | Model weight acquisition method and related system | |
| JP2022175851A (en) | Information processing apparatus, information processing method, and program | |
| Chen et al. | Multi-view robust discriminative feature learning for remote sensing image with noisy labels | |
| Okada et al. | CycleGAN using Semi-Supervised Learning. | |
| CN112417260A (en) | Localized recommendation method and device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |