Face anti-counterfeiting generalization method based on distributed robustness learning and related device

Abstract

The invention discloses a face anti-counterfeiting generalization method based on distributed robustness learning and a related device. The method comprises the following steps: acquiring a to-be-detected face image; inputting a to-be-detected face image into the face anti-counterfeiting network model, and outputting a face result by the face anti-counterfeiting network model to realize face anti-counterfeiting generalization; wherein the face anti-counterfeiting network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module. According to the method, training distribution is enhanced on all dimensions of main conditions (such as illumination, facial appearance and camera quality) and attack types (such as emerging 3D masks made of new materials), so that the enhanced distribution covers part of test distribution as much as possible, and the problem of distribution deviation is greatly relieved; and the model has good generalization ability in unknown attack types and unknown scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of face anti-counterfeiting, and in particular relates to a face anti-counterfeiting generalization method based on distributional robustness learning and a related device. Background Art

[0002] In recent years, face recognition systems have been widely used in various identity authentication scenarios. However, it is vulnerable to face rendering attacks such as printed photos, video replays, and 3D face masks, posing potential security threats. To address this problem, researchers have proposed various face anti-counterfeiting methods based on deep neural networks. Existing face anti-counterfeiting methods perform well on specific datasets and protocols, but perform poorly on unseen scenarios and unknown attacks. However, when deploying face anti-counterfeiting systems in the real world, it is often necessary to deal with scene changes and new types of attacks. Therefore, the key is to build a robust face anti-counterfeiting model that can handle unknown scenarios and new types of attacks well.

[0003] Domain adaptation and domain generalization are two commonly used techniques to generalize face anti-counterfeiting models to unknown domains. Domain adaptation technology uses the knowledge of the target domain to bridge the gap between the source domain and the target domain. Face anti-counterfeiting methods based on domain adaptation mainly include two categories: 1) Domain adaptation methods based on domain mapping; 2) Domain adaptation methods based on domain adversary. Domain generalization technology can directly learn generalized feature representations from multiple source domains without accessing target data. Face anti-counterfeiting methods based on domain generalization mainly include three categories: 1) Methods based on domain adversary; 2) Methods based on domain decoupling; 3) Methods based on domain fusion. In addition, for the generalization of unknown attack types, there are mainly two types of methods: 1) Methods based on zero / small sample learning; 2) Methods based on anomaly detection.

[0004] In summary, the key to solving the generalization problem of face anti-counterfeiting is to alleviate the distribution shift caused by unseen scenes or new attacks, while existing methods mainly focus on learning domain-invariant or generalized features, and cannot alleviate the larger distribution shift caused by different lighting, camera quality, background, facial appearance, attack type and other conditions. Summary of the invention

[0005] In view of the problem of distribution bias in existing face anti-counterfeiting generalization methods, the purpose of the present invention is to provide a face anti-counterfeiting generalization method and related devices based on distributional robustness learning, which has good generalization ability in unknown attack types and unknown scenarios.

[0006] To achieve the above object, the present invention adopts the following technical solution: A face anti-counterfeiting generalization method based on distributional robustness learning includes the following steps: Get the face image to be detected; The face image to be detected is input into the face anti-counterfeiting network model, and the face anti-counterfeiting network model outputs the face result to achieve face anti-counterfeiting generalization; Among them, the face anti-counterfeiting network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

[0007] A further improvement of the present invention is that the face anti-counterfeiting network model is established through the following process: Obtain a face anti-counterfeiting dataset, pre-process the dataset using the MTCNN algorithm, and then divide it into a training set and a test set; Extract features from images in the training set and calculate pseudo labels; The joint distribution of facial samples is estimated based on pseudo labels to obtain the joint distribution of facial samples; Enhance the joint distribution of face samples estimated in each domain, and obtain the distribution enhancement loss; The joint distribution of each pair of facial samples from different domains is mapped to a common space through a Kantorovich potential network, and the dual optimal transfer distance between each pair of distributions is calculated using the dual optimal transfer distance. According to the dual optimal transfer distance between each pair of distributions, the distribution fusion loss is calculated; The total loss function of the face anti-counterfeiting network model is obtained based on the distribution enhancement loss and distribution fusion loss.

[0008] A further improvement of the present invention is that the pseudo-label is obtained by the following process: The training set includes standard training data and unlabeled training data; Each facial sample in the labeled training data is mapped to the facial sample feature space through the feature extraction function to obtain the sample features, and then the sample features are mapped to the sample feature label space through the classifier to obtain the pseudo label.

[0009] A further improvement of the present invention is that the total loss function of the face anti-counterfeiting network model is The calculation formula is:

[0010] In the formula, is the cross entropy loss, is the inter-domain distribution fusion loss, is the in-domain distribution fusion loss, is the first weight parameter, is the second weight parameter.

[0011] A further improvement of the present invention is that the cross entropy loss is calculated by the following formula:

[0012] In the formula, are network parameters, is a sample random variable, is the number of domains, For the The number of samples in the domain, For label The corresponding one-hot vector, For the The first samples, is the domain number, Number the sample; The inter-domain distribution fusion loss is calculated as follows:

[0013] In the formula, is the balance parameter, is the KL divergence of the sample labels, For adversarial samples, It is triple marginal loss; The intra-domain distribution fusion loss is calculated as follows:

[0014] In the formula, is the Kantorovich potential network, For the A random variable with samples from a domain, For the A random variable with samples from a domain, is the mean function, For the The probability distribution of a domain, For the The probability distribution of a domain, is a constant.

[0015] A further improvement of the present invention is that the distribution fusion loss is specifically calculated through the following process: using the dual optimal transmission distance to calculate the dual optimal transmission distance between each pair of distributions, and calculating the distribution fusion loss based on the dual optimal transmission distance between each pair of distributions.

[0016] A further improvement of the present invention is that the dual optimal transmission distance between each pair of distributions is Calculated by the following formula:

[0017] In the formula, For the A random variable with samples from a domain, For the A random variable with samples from a domain, is the mean function, For the The probability distribution of a domain, For the The probability distribution of a domain, is a constant.

[0018] A face anti-counterfeiting generalization system based on distributional robustness learning includes: A face image acquisition module is used to acquire a face image to be detected; The face anti-counterfeiting network model calculation module is used to input the face image to be detected into the face anti-counterfeiting network model for calculation. The face anti-counterfeiting network model outputs the face result to achieve face anti-counterfeiting generalization; Among them, the face anti-counterfeiting network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

[0019] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the face anti-counterfeiting generalization method based on distributional robustness learning is implemented.

[0020] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the face anti-counterfeiting generalization method based on distributional robustness learning is implemented.

[0021] Compared with the prior art, the present invention has the following beneficial effects: On the one hand, the present invention uses the in-domain distribution enhancement module to enhance the training distribution in various dimensions of major conditions (such as lighting, facial appearance, and camera quality) and attack types (such as emerging 3D masks made of new materials), and tries to make the enhanced distribution cover part of the test distribution, greatly alleviating the problem of distribution shift, so that the model can have good generalization ability in unknown attack types and unknown scenarios. On the other hand, the distribution fusion module is used to aggregate information from each training distribution to minimize the total cost of transferring each training distribution to the common product space, and the model trained using the fused distribution has better generalization ability for the test distribution. Through the joint action of these two modules, the present invention can obtain a model with good generalization ability when the test distribution is significantly different from the training distribution.

[0022] Furthermore, both the intra-domain distribution enhancement module and the inter-domain distribution fusion module are designed based on the joint distribution of sample features and labels (pseudo-labels), which enables the model to transfer information from samples within the training distribution class to the test distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the process of the face anti-counterfeiting generalization method based on distributional robustness learning in the present invention; Figure 2 This is a demonstration of the detection effect of the present invention on an unknown attack type in an unknown scenario; Figure 3 Schematic diagram of the face anti-counterfeiting generalization system based on distributional robustness learning of the present invention. DETAILED DESCRIPTION

[0024] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thoroughly and comprehensively understood.

[0025] In order to solve the distribution deviation problem of face anti-counterfeiting model caused by samples of unknown attack types in unknown scenarios in the test distribution, the present invention proposes a face anti-counterfeiting generalization method based on distributional robustness learning, including the following steps: First, given real and spoofed face samples from different domains, we first use a pre-trained feature extractor and a randomly initialized classifier to estimate the joint distribution of real and spoofed face samples from different domains.

[0026] Then, the joint distribution of each domain estimate is augmented by learning the worst-case joint distribution in a Wasserstein sphere around each training distribution using the Adversarial Stein Variational Gradient Descent (A-SVGD) algorithm.

[0027] Meanwhile, each pair of joint distributions from different domains is mapped into a common space by learning a Kantorovich potential network related to the dual optimal transport problem.

[0028] Finally, the parameters of the feature extractor and classifier will be optimized under distribution enhancement loss, distribution fusion loss and cross entropy classification loss. These three modules are learned and updated alternately. Extensive experimental results on five widely used face anti-counterfeiting datasets, namely MSU-MFSD, IDIAP REPLAY-ATTACK, CASIA-FASD, OULU-NPU and WMCA, show that compared with the most advanced domain adversarial-based methods, domain decoupling-based methods and domain fusion-based methods, the face anti-counterfeiting generalization method based on distribution robustness learning proposed in this invention can achieve the best performance in both unseen scenes and unknown attacks, showing its excellent generalization ability.

[0029] The present invention is specifically achieved by the following steps: Step 1: Collect the face anti-counterfeiting dataset, pre-process the dataset using the MTCNN algorithm, and then divide it into a training set and a test set.

[0030] Step 2: Extract features from the images in the training set and calculate pseudo labels.

[0031] Specifically, the training set includes standard training data and unlabeled training data; Training data labeled in the training set , is the number of domains in the training data, where Indicates The number of samples in a domain, The label of the domain. represents a face sample, represents the label, i represents the sample number in the kth domain, Representation Tags The corresponding one-hot vector.

[0032] Through feature extraction function For labeled training data Each face sample in Mapping to face samples Feature space, get the sample features, where is the network parameter of the feature extractor; then through the classifier Map the sample features to the sample feature label space to obtain pseudo labels.

[0033] No. In the domain The characteristics of a sample can be expressed as ,Right now, , the pseudo label can be expressed as ,Right now, .

[0034] Step 3: Construct an intra-domain distribution enhancement module for learning robust distribution and an inter-domain distribution fusion module for fusing multiple training distributions; like Figure 1 As shown, the training process includes three steps, namely, joint distribution estimation, intra-domain distribution enhancement, and inter-domain distribution fusion.

[0035] Specifically, for face samples from K domains, we first use a weight-sharing pre-trained feature extractor and a randomly initialized classifier to estimate the joint distribution of face samples, which consists of sample features and pseudo labels.

[0036] Then, the joint distribution of face samples estimated in each domain is enhanced. The most challenging (i.e., worst-case) distribution in the uncertain set around the joint distribution of each face sample is learned using the Adversarial Stein Variational Gradient Descent (A-SVGD) algorithm, resulting in a distribution enhancement loss.

[0037] At the same time, the sample features of the joint distribution of face samples in each domain are first passed into the Kantorovich potential network , then the dual optimal transmission distance is used to measure the difference between each pair of training distributions, and finally the Kantorovich potential network is updated according to the sum of the dual optimal transmission distances of the difference between each pair of training distributions, thereby obtaining the distribution fusion loss. This is achieved by learning the Kantorovich potential network related to the dual optimal transmission problem.

[0038] Finally, the parameters of the feature extractor and classifier will be optimized under distribution enhancement loss, distribution fusion loss and cross entropy classification loss to obtain the total loss.

[0039] Once the feature extractor and classifier are trained, they can be used during the testing phase to estimate the labels of samples from the test distribution.

[0040] An in-domain distribution enhancement module is constructed for learning robust distribution, and the Bayesian inference algorithm A-SVGD based on adversarial training is adopted to generate adversarial samples and approximate the most challenging distribution.

[0041] The details of the intra-domain distribution enhancement module and inter-domain distribution fusion are as follows: (1) Enhanced distribution within the domain For The first Samples , first generate in a Euclidean sphere of adversarial examples, and then use the multi-particle search strategy to update these examples. adversarial examples of all samples in the domain to approximate the most challenging distribution.

[0042] Specifically, we first add uniform noise to n adversarial examples Initialize and then project these adversarial samples onto the ball , ,Right now

[0043] In the formula, is a point in the sphere, is the radius of the sphere, is the initialized adversarial sample, j is the adversarial sample number, n is the number of adversarial samples, is the initial noise factor, is random noise, It's projection, projection The role of is to constrain all generated adversarial examples to be inside or on the boundary of the sphere.

[0044] According to the multi-particle search strategy of SVGD, there are the following iterations:

[0045] In the formula, represents the number of iterations, represents the step length, Indicates the update direction. yes Iteration of adversarial examples.

[0046] According to Stein variational gradient descent theory, the update direction of each particle is determined by the kernelized Stein difference between the approximate distribution and the target distribution, that is,

[0047] In the formula, is to find the gradient, t is the number of other adversarial samples when calculating the jth adversarial sample, is a positive definite radial basis function (RBF) kernel , defined as follows,

[0048] In the formula, is a positive definite radial basis function (RBF) kernel, is any sample, is any adversarial example, is the kernel width, which is set empirically and is related to the number of adversarial samples. Proportional (i.e. ).

[0049] The entire process of the above multi-particle search strategy is called the Adversarial Stein Variance Gradient Descent algorithm (A-SVGD).

[0050] (2) Inter-domain distribution fusion Specifically, given domain and Training distribution of domains and , Kantorovich potential network The parameters are recorded as ,and The relevant optimization problem can be expressed as follows:

[0051] In the formula, is maximized, It is A random variable with samples from a domain, It is The joint distribution of samples from domains, is the mean, d is the distance function, is a constant, is defined as follows:

[0052] In the formula, It is A random variable with samples from a domain, It is The joint distribution of samples from the domains.

[0053] Kantorovich Potential Network Consists of three fully connected layers, the optimized network is obtained back, and The dual optimal transmission distance between Can be optimized by the network Calculate, that is

[0054] In the formula, For the A random variable with samples from a domain, For the A random variable with samples from a domain, is the mean function, For the The probability distribution of a domain, For the The probability distribution of a domain, is a constant.

[0055] When there are multiple training distributions, use the dual optimal transmission distance Compute the pairwise optimal transmission distance between each pair of distributions.

[0056] The distribution fusion loss is calculated based on the dual optimal transmission distance between each pair of distributions.

[0057] Step 4: Perform network training based on the intra-domain distribution enhancement module and the inter-domain distribution fusion module to obtain the face anti-counterfeiting network model. Design intra-domain distribution enhancement loss and inter-domain distribution fusion loss It is worth noting that when only a single domain is used for training, the inter-domain distribution fusion loss will be degraded. In addition, as a typical classification task, the widely used cross entropy loss The final total loss is also essential. , including distribution enhancement loss, distribution fusion loss and classification loss. are network parameters, is a sample random variable and is an adversarial sample random variable.

[0058] Specifically, the intra-domain distribution enhancement loss It consists of two parts. The first term is the triple margin loss , which enables learning better facial embeddings associated with the generated adversarial examples. That is, given any anchor adversarial example, positive adversarial examples (i.e., the same type as the anchor) are pulled closer to the anchor, while negative adversarial examples are pushed away from the anchor. It can be stated as follows:

[0059] In the formula, , and Represent anchor samples, positive samples and negative samples respectively, is the one-hot vector corresponding to the label. express norm, For the remainder, the remainder is a constant. It should be noted that if the training samples are unlabeled (such as samples of the target domain in domain adaptation problems), this term will degenerate.

[0060] The second term is the KL divergence of the sample labels , which measures the difference between the pseudo-label of the adversarial sample and its corresponding original training sample, namely:

[0061] In the formula, For the In the domain The characteristics of the samples, are the network parameters of the classifier, The number of the sample.

[0062] Therefore, the The sample in The distribution enhancement loss in the domain can be expressed as follows:

[0063] In the formula, is a balance parameter.

[0064] The distribution enhancement loss for all training samples in all domains can be expressed as:

[0065] In the formula, is the number of domains, The number of adversarial samples generated for each sample, is the domain number, is the number of the adversarial sample.

[0066] Inter-domain distribution fusion loss It can be expressed as the dual optimal transmission distance between a pair of training distributions corresponding to any pair of domains, that is:

[0067] In the formula, It is a Kantorovich potential network.

[0068] All training distributions The distribution fusion loss can be expressed as:

[0069] Kantorovich Potential Network Then update by optimizing the following objective function:

[0070] In addition, as a typical classification task, the following cross entropy loss is used :

[0071] Finally, the total loss function of the face anti-counterfeiting network model is The calculation formula is:

[0072] In the formula, , is the first weight parameter, is the second weight parameter.

[0073] The final feature extraction and classification network is obtained by minimizing the function Training:

[0074] Step 5: After training is completed, the face anti-counterfeiting network model can be used to determine whether the image from the test set is a real face or a spoofed face.

[0075] The generalization effect of the model is evaluated on a test set. Face samples from four data sets are input as training sets. The network parameters of the feature extractor and classifier are learned under the joint supervision of the loss functions of the intra-domain distribution enhancement module and the inter-domain distribution fusion module, and the true and false faces from the test set are distinguished. Figure 2 The attack samples detected by the distribution fusion module, the distribution enhancement module, and the combination of the two modules under the O&C&I&M→W setting are shown when testing across data sets and attack types. The results show that the distribution fusion module has a good detection effect on attack samples under uniform background daylight illumination. The distribution enhancement module has a good detection effect on attack samples under uniform background bottom LED light illumination. The combination of the two modules has a good detection effect on attack samples under complex backgrounds.

[0076] See also Figure 3 Another embodiment of the present invention provides a face anti-counterfeiting generalization system based on distributional robustness learning, including: A face image acquisition module is used to acquire a face image to be detected; The face anti-counterfeiting network model calculation module is used to input the face image to be detected into the face anti-counterfeiting network model for calculation. The face anti-counterfeiting network model outputs the face result to achieve face anti-counterfeiting generalization; Among them, the face anti-counterfeiting network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

[0077] Another embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the face anti-counterfeiting generalization method based on distributional robustness learning when executing the computer program.

[0078] Another embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the face anti-counterfeiting generalization method based on distributional robustness learning is implemented.

[0079] In the present invention, a face anti-counterfeiting data set is collected and pre-processed; features are extracted from images in the training set and pseudo labels are calculated; an intra-domain distribution enhancement module for learning robustness distribution and an inter-domain distribution fusion module for fusing multiple training distributions are constructed; network training is performed based on the intra-domain distribution enhancement module and the inter-domain distribution fusion module; after the training is completed, the face anti-counterfeiting network model can determine whether the image from the test set is a real face or a deceptive face. The present invention enhances the training distribution in various dimensions of the main conditions (such as lighting, facial appearance and camera quality) and attack types (such as emerging 3D masks made of new materials), so that the enhanced distribution covers part of the test distribution as much as possible, greatly alleviating the problem of distribution offset, and enabling the model to have good generalization ability in unknown attack types and unknown scenarios.

[0080] The above description is only for the best embodiment of the present invention, but it should not be understood as limiting the claims. The present invention is not limited to the above embodiments, and its specific structure is allowed to be changed. However, all changes made within the protection scope of the independent claims of the present invention are within the protection scope of the present invention.

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

Claims

1. A generalized face anti-counterfeiting method based on distributional robustness learning, characterized in that: The following steps are involved: Get the face image to be detected; The face image to be detected is input into the face anti-counterfeiting network model, and the face anti-counterfeiting network model outputs the face result to achieve face anti-counterfeiting generalization; Among them, the face anti-counterfeiting network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

2. The face anti-counterfeiting generalization method based on distributional robustness learning according to claim 1 is characterized in that: The face anti-counterfeiting network model is established through the following process: Obtain a face anti-counterfeiting dataset, pre-process the dataset using the MTCNN algorithm, and then divide it into a training set and a test set; Extract features from images in the training set and calculate pseudo labels; The joint distribution of facial samples is estimated based on pseudo labels to obtain the joint distribution of facial samples; Enhance the joint distribution of face samples estimated in each domain, and obtain the distribution enhancement loss; The joint distribution of each pair of facial samples from different domains is mapped to a common space through a Kantorovich potential network, and the dual optimal transfer distance between each pair of distributions is calculated using the dual optimal transfer distance. According to the dual optimal transfer distance between each pair of distributions, the distribution fusion loss is calculated; The total loss function of the face anti-counterfeiting network model is obtained based on the distribution enhancement loss and distribution fusion loss.

3. The face anti-counterfeiting generalization method based on distributional robustness learning according to claim 2 is characterized in that: Pseudo labels are obtained through the following process: The training set includes standard training data and unlabeled training data; Each facial sample in the labeled training data is mapped to the facial sample feature space through the feature extraction function to obtain the sample features, and then the sample features are mapped to the sample feature label space through the classifier to obtain the pseudo label.

4. The face anti-counterfeiting generalization method based on distributional robustness learning according to claim 2 is characterized in that: Total loss function of face anti-counterfeiting network model The calculation formula is: In the formula, is the cross entropy loss, is the inter-domain distribution fusion loss, is the in-domain distribution fusion loss, is the first weight parameter, is the second weight parameter.

5. The face anti-counterfeiting generalization method based on distributional robustness learning according to claim 4 is characterized in that: The cross entropy loss is calculated as follows: In the formula, are network parameters, is a sample random variable, is the number of domains, For the The number of samples in the domain, For label The corresponding one-hot vector, For the The first samples, is the domain number, Number the sample; The inter-domain distribution fusion loss is calculated as follows: In the formula, is the balance parameter, is the KL divergence of sample labels, For adversarial samples, It is triple marginal loss; The intra-domain distribution fusion loss is calculated as follows: In the formula, is the Kantorovich potential network, For the A random variable with samples from a domain, For the A random variable with samples from a domain, is the mean function, For the The probability distribution of a domain, For the The probability distribution of a domain, is a constant.

6. The face anti-counterfeiting generalization method based on distributional robustness learning according to claim 2 is characterized in that: The distribution fusion loss is specifically calculated through the following process: the dual optimal transmission distance is used to calculate the dual optimal transmission distance between each pair of distributions, and the distribution fusion loss is calculated based on the dual optimal transmission distance between each pair of distributions.

7. The face anti-counterfeiting generalization method based on distributional robustness learning according to claim 6 is characterized in that: The dual optimal transmission distance between each pair of distributions Calculated by the following formula: In the formula, For the A random variable with samples from a domain, For the A random variable with samples from a domain, is the mean function, For the The probability distribution of a domain, For the The probability distribution of a domain, is a constant.

8. A face anti-counterfeiting generalization system based on distributional robustness learning, characterized in that: include: A face image acquisition module is used to acquire a face image to be detected; The face anti-counterfeiting network model calculation module is used to input the face image to be detected into the face anti-counterfeiting network model for calculation. The face anti-counterfeiting network model outputs the face result to achieve face anti-counterfeiting generalization; Among them, the face anti-counterfeiting network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the face anti-counterfeiting generalization method based on distributional robustness learning as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the face anti-counterfeiting generalization method based on distributional robustness learning as described in any one of claims 1 to 7 is implemented.

Images