A face anti-counterfeiting generalization method based on distribution robustness learning and related device

Abstract

The application discloses a face anti-counterfeiting generalization method based on distribution robustness learning and related devices, and acquires a face image to be detected; the face image to be detected is input into a face anti-counterfeiting network model, the face anti-counterfeiting network model outputs a face result, and face anti-counterfeiting generalization is realized; wherein the face anti-counterfeiting network model is obtained by training a distribution enhancement module in a domain and a distribution fusion module between domains. The application enhances the training distribution in various dimensions of main conditions (such as illumination, facial appearance and camera quality) and attack types (such as a new 3D mask made of new materials), so that the enhanced distribution covers part of the test distribution as much as possible, greatly alleviates the problem of distribution deviation, and enables the model to have good generalization capability in unknown attack types and unknown scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of face anti-spoofing, and particularly relates to a face anti-spoofing generalization method based on distribution robustness learning and a related device. BACKGROUND

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

[0003] Domain adaptation and domain generalization are two commonly used techniques for generalizing face anti-spoofing models to unknown domains. Domain adaptation techniques use knowledge from the target domain to bridge the gap between the source and target domains. Face anti-spoofing methods based on domain adaptation mainly include two categories: 1) domain mapping-based domain adaptation methods; 2) domain adversarial-based domain adaptation methods. Domain generalization techniques can directly learn a generalizable feature representation from multiple source domains without accessing target data. Face anti-spoofing methods based on domain generalization mainly include three categories: 1) domain adversarial-based methods; 2) domain decoupling-based methods; 3) domain fusion-based methods. In addition, for the generalization of unknown attack types, there are mainly two methods: 1) zero / few-shot learning-based methods; 2) anomaly detection-based methods.

[0004] In summary, the key to solving the face anti-spoofing generalization problem lies in alleviating the distribution shift caused by unseen scenarios or new attacks, while existing methods mainly focus on learning domain-invariant or generalizable features, which cannot alleviate the greater distribution shift caused by different lighting conditions, camera quality, background, facial appearance, attack type, and other conditions. SUMMARY

[0005] To address the distribution shift problem in existing face anti-spoofing generalization methods, the purpose of the present application is to provide a face anti-spoofing generalization method based on distribution robustness learning and a related device, which has good generalization ability in unknown attack types and unknown scenarios.

[0006] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0007] A face anti-spoofing generalization method based on distribution robustness learning, comprising the following steps:

[0008] obtaining a face image to be detected;

[0009] inputting the face image to be detected into a face anti-counterfeiting network model, and outputting a face result by the face anti-counterfeiting network model to realize face anti-counterfeiting generalization;

[0010] The face anti-counterfeiting network model is obtained by training a domain distribution enhancement module and a domain distribution fusion module.

[0011] The face anti-counterfeiting network model is further improved by the following process:

[0012] obtaining a face anti-counterfeiting dataset, and preprocessing the dataset by using an MTCNN algorithm, and then dividing the dataset into a training set and a test set;

[0013] extracting features of images in the training set, and calculating pseudo labels;

[0014] performing joint distribution estimation of face samples based on the pseudo labels to obtain joint distribution of the face samples;

[0015] enhancing the joint distribution of the face samples estimated by each domain to obtain a distribution enhancement loss;

[0016] mapping the joint distribution of each pair of face samples from different domains to a common space by a Kantorovich potential network, calculating a dual optimal transport distance between each pair of distributions by using a dual optimal transport distance, and calculating a distribution fusion loss according to the dual optimal transport distance between each pair of distributions;

[0017] obtaining a total loss function of the face anti-counterfeiting network model according to the distribution enhancement loss and the distribution fusion loss.

[0018] The pseudo labels are further improved by the following process:

[0019] The training set includes standard training data and unlabeled training data;

[0020] mapping each face sample in the labeled training data to a face sample feature space by using a feature extraction function to obtain a sample feature, and then mapping the sample feature to a sample feature label space by using a classifier to obtain a pseudo label.

[0021] The total loss function of the face anti-counterfeiting network model is further improved as follows: The calculation formula of the total loss function is as follows:

[0022]

[0023] In the formula, is a cross-entropy loss, For inter-domain distribution fusion loss, For the domain-specific fusion loss, As the first weight parameter, This is the second weighting parameter.

[0024] A further improvement of this invention is that the cross-entropy loss is calculated using the following formula:

[0025]

[0026] In the formula, For network parameters, For sample random variables, For the number of domains, For the first The number of samples in each domain For tags The corresponding one-hot vector, For the first The first in the domain One sample, For the field number, Number the sample;

[0027] The inter-domain distribution fusion loss is calculated using the following formula:

[0028]

[0029] In the formula, For balancing parameters, The KL divergence of the sample labels. For adversarial examples, This represents a triple marginal loss.

[0030] The domain-specific fusion loss is calculated using the following formula:

[0031]

[0032] In the formula, For Kantorovich potential network, For from the Random variables of samples in a domain, For from the Random variables of samples in a domain, It is a mean function. For the first The probability distribution of each domain, For the first The probability distribution of each domain, It is a constant.

[0033] The further improvement of the present application is that the distribution fusion loss is calculated by the following process: calculating the dual optimal transport distance between each pair of distributions using the dual optimal transport distance, and calculating the distribution fusion loss according to the dual optimal transport distance between each pair of distributions.

[0034] The further improvement of the present application is that the dual optimal transport distance between each pair of distributions is calculated by the following formula:

[0035]

[0036] wherein, is a random variable of the sample from the i-th domain, is a random variable of the sample from the i-th domain, is a random variable of the sample from the i-th domain, is a mean function, is a probability distribution of the i-th domain, is a probability distribution of the i-th domain, is a probability distribution of the i-th domain, is a constant. A face anti-counterfeiting generalization system based on distribution robustness learning comprises:

[0037] A face image acquisition module is configured to acquire a face image to be detected.

[0038] A face anti-counterfeiting network model calculation module is configured to input the face image to be detected into a face anti-counterfeiting network model for calculation, and the face anti-counterfeiting network model outputs a face result to realize face anti-counterfeiting generalization.

[0039] The face anti-counterfeiting network model is obtained by network training on a distribution enhancement module in a domain and a distribution fusion module between domains.

[0040] An electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to realize the face anti-counterfeiting generalization method based on distribution robustness learning.

[0041] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the face anti-counterfeiting generalization method based on distribution robustness learning.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] ​In one aspect, the intra-domain distribution enhancement module enhances the training distribution in various 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 as to cover part of the test distribution as much as possible, greatly alleviate the distribution shift problem, and enable the model to have good generalization ability in unknown attack types and unknown scenes. On the other hand, the distribution fusion module aggregates information from each training distribution to minimize the total cost of transferring each training distribution to a common product space, and the model trained using the fusion distribution has better generalization ability to the test distribution. Through the joint action of the two modules, a model with good generalization ability in the case of significant difference between the test distribution and the training distribution can be obtained.

[0045] Further, the intra-domain distribution enhancement module and the inter-domain distribution fusion module are both designed based on the joint distribution of sample features and labels (pseudo labels), which can enable the model to transfer information from the intra-class samples of the training distribution to the test distribution. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of the face anti-counterfeiting generalization method based on distribution robustness learning in the present application;

[0047] Figure 2 A detection effect display of unknown attack types in unknown scenes in the present application;

[0048] Figure 3 A schematic diagram of the face anti-counterfeiting generalization system based on distribution robustness learning in the present application. DETAILED DESCRIPTION

[0049] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0050] In order to solve the problem of distribution shift caused by samples of unknown attack types in unknown scenes in the test distribution of the face anti-counterfeiting model, the present application proposes a face anti-counterfeiting generalization method based on distribution robustness learning, including the following steps:

[0051] First, given real and fake face samples from different domains, a pre-trained feature extractor and a randomly initialized classifier are first used to estimate the joint distribution of real and fake face samples from different domains.

[0052] Then, the joint distribution of each domain estimate is augmented. The worst-case joint distribution is learned in the Wasserstein sphere around each training distribution by using the adversarial Stein variational gradient descent (A-SVGD) algorithm.

[0053] Meanwhile, by learning the Kantorovich potential network related to the dual optimal transport problem, each pair of joint distributions from different domains is mapped to a common space.

[0054] Finally, the parameters of the feature extractor and classifier are optimized using distribution augmentation 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-spoofing datasets—MSU-MFSD, IDIAP REPLAY-ATTACK, CASIA-FASD, OULU-NPU, and WMCA—show that the proposed face anti-spoofing generalization method based on distribution augmentation learning achieves the best performance under both unseen scenarios and unknown attacks, demonstrating its superior generalization ability, compared to state-of-the-art methods based on domain adversarial, domain decoupling, and domain fusion.

[0055] This invention is specifically implemented through the following steps:

[0056] Step 1: Collect a face anti-spoofing dataset, and preprocess the dataset using the MTCNN algorithm, then divide it into training and test sets.

[0057] Step 2: Extract features from the images in the training set and calculate pseudo-labels.

[0058] Specifically, the training set includes standard training data as well as unlabeled training data;

[0059] Annotated training data in the training set , It is the number of domains in the training data, here Indicates the first The number of samples in each domain The label representing the field. Indicates a facial sample. Let i represent the label, and i represent the sample number in the k-th field. Indicates label The corresponding one-hot vector.

[0060] Through feature extraction function labeled training data Each facial sample Mapping to facial samples The feature space is used to obtain the sample features, where, network parameters of the feature extractor; then the pseudo labels are obtained by the classifier mapping the sample features to the sample feature label space, to obtain the pseudo labels.

[0061] The feature of the i-th sample in the j-th domain can be represented as , i.e. , the pseudo label can be represented as , i.e. . .

[0062] Step 3: Constructing an intra-domain distribution enhancement module for learning robust distribution, and an inter-domain distribution fusion module for fusing multiple training distributions;

[0063] As shown in FIG. 3, the training process includes three steps, i.e., joint distribution estimation, intra-domain distribution enhancement, and inter-domain distribution fusion. Figure 1

[0064] Specifically, for the face samples from K domains, first, a weight-shared pre-trained feature extractor and a randomly initialized classifier are used to estimate the joint distribution of the face samples. The joint distribution of the face samples is composed of sample features and pseudo labels.

[0065] Then, the joint distribution of the face samples estimated by each domain is enhanced. By using the adversarial Stein variational gradient descent algorithm (A-SVGD), the most challenging (i.e., the worst case) distribution in the uncertainty set around the joint distribution of each face sample is learned, to obtain the distribution enhancement loss.

[0066] At the same time, first, the sample features of the joint distribution of the face samples of each domain are transmitted into the Kantorovich potential network , second, the dual optimal transport distance is used to measure the difference between each pair of training distributions, and finally, the sum of the dual optimal transport distance between each pair of training distributions is used to update the Kantorovich potential network, to obtain the distribution fusion loss. This is achieved by learning the Kantorovich potential network related to the dual optimal transport problem.

[0067] Finally, the parameters of the feature extractor and the classifier will be optimized under the distribution enhancement loss, the distribution fusion loss, and the cross-entropy classification loss, to obtain the total loss.

[0068] Once the feature extractor and the classifier are trained, they can be used in the test phase to estimate the labels of the samples from the test distribution.

[0069] ​​An intra-domain distribution enhancement module is constructed for learning robust distributions. The adversarial training-based Bayesian inference algorithm A-SVGD is used to generate adversarial examples and approximate the most challenging distributions.

[0070] The specifics of the intra-domain distribution enhancement module and inter-domain distribution fusion are as follows:

[0071] (1) Enhanced intra-domain distribution

[0072] For the The first in the domain Sample First, it is generated in a Euclidean sphere. of We first generate adversarial examples, then update these examples using a multi-particle search strategy. Finally, we utilize the first... Adversarial examples of all samples in each domain are used to approximate the most challenging distribution.

[0073] Specifically, firstly, by adding uniform noise... n adversarial examples Initialize the sample data, then project these adversarial examples onto the ball. , ,Right now

[0074]

[0075] In the formula, It is a point in the sphere. It is the radius of the sphere. These are the initial adversarial samples, where j is the adversarial sample label and n is the number of adversarial samples. It is the initial noise figure. It is random noise. It's a projection, a projection. Its function is to constrain all generated adversarial samples to be located inside or on the boundary of the sphere.

[0076] Based on the multi-particle search strategy of SVGD, the following iterations are possible:

[0077]

[0078] In the formula, Indicates the number of iterations. Indicates step size, Indicates the direction of the update. yes Adversarial examples in the next iteration.

[0079] According to the Stein variational gradient descent theory, the update direction of each particle is determined by the nucleated Stein difference between the approximate distribution and the target distribution, i.e.

[0080]

[0081] In the formula, This is for calculating the gradient, where t is the label of other adversarial samples when calculating the j-th adversarial sample. It is a positive definite radial fundamental function (RBF) kernel The definition is as follows:

[0082]

[0083] In the formula, It is a positive definite radial fundamental function (RBF) kernel. It is any sample. It is any adversarial example. It refers to the kernel width, which is set empirically and is related to the number of adversarial examples. proportional (i.e.) ).

[0084] The entire process of the above multi-particle search strategy is called the Adversarial Stein Variance Gradient Descent algorithm (A-SVGD).

[0085] (2) Inter-domain distribution fusion

[0086] Specifically, given the first The domain and the first Training distribution of each domain and Kantorovich power network The parameters are denoted as ,and The relevant optimization problem can be described as follows:

[0087]

[0088] In the formula, It is to maximize. It is the first Random variables of samples in a domain, It is the first The joint distribution of samples from each domain. It calculates the mean, where d is the distance function. It is a constant. The definition is as follows:

[0089]

[0090] In the formula, It is the first a random variable of a sample from the th domain, is the joint distribution of samples from the first and the th domain.

[0091] Kantorovich potential network is composed of three fully connected layers, and the optimized network is obtained. and between the dual optimal transport distance can be calculated by the optimized network , i.e.

[0092]

[0093] where is a random variable of a sample from the th domain, is a random variable of a sample from the th domain, is the mean function, is the probability distribution of the th domain, is the probability distribution of the th domain, is a constant.

[0094] When there are multiple training distributions, the dual optimal transport distance between each pair of distributions is calculated.

[0095] According to the dual optimal transport distance between each pair of distributions, the distribution fusion loss is calculated.

[0096] Step 4: Network training based on the intra-domain distribution enhancement module and the inter-domain distribution fusion module to obtain a face anti-counterfeiting network model. Design the intra-domain distribution enhancement loss and the 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 degenerate. In addition, as a typical classification task, the widely used cross-entropy loss is also essential. The final total loss includes the distribution enhancement loss, the distribution fusion loss, and the classification loss. is the network parameter, is the sample random variable, and is the adversarial sample random variable.

[0097] Specifically, the intra-domain distribution enhancement loss is composed of two parts. The first term is the triple marginal loss It can make the learning of better face embedding related to the generated adversarial samples. That is, given any anchor adversarial sample, the positive adversarial sample (i.e. of the same class as the anchor) will be pulled closer to the anchor, while the negative adversarial sample will be pushed away from the anchor. It can be expressed as follows:

[0098]

[0099] where, , and denote the anchor sample, the positive sample and the negative sample, respectively, is the one-hot vector corresponding to the label. denotes norm, is the margin, and the margin is a constant. It is noted that if the training samples are unlabeled (such as the samples of the target domain in the domain adaptation problem), this term will degenerate.

[0100] The second term is the KL divergence of the sample label , which measures the difference between the pseudo-label of the adversarial sample and the original training sample corresponding to it, that is:

[0101]

[0102] where, is the feature of the th sample in the th domain, is the network parameter of the classifier, is the number of the sample.

[0103] Therefore, the distribution enhancement loss of the th sample in the th domain can be expressed as follows:

[0104]

[0105] where, is the balance parameter.

[0106] The distribution enhancement loss of all training samples in all domains can be expressed as:

[0107]

[0108] where, is the number of domains, is the number of generated adversarial samples corresponding to each sample, is the number of domains, is the number of adversarial samples.

[0109] Inter-domain distribution fusion loss The inter-domain distribution fusion loss can be expressed as the dual optimal transport distance between a pair of training distributions corresponding to any pair of domains, i.e.,

[0110]

[0111] where, is the Kantorovich potential network.

[0112] All training distributions The distribution fusion loss can be expressed as:

[0113]

[0114] Kantorovich potential network is updated by optimizing the following objective function:

[0115]

[0116] In addition, as a typical classification task, the following cross-entropy loss is used :

[0117]

[0118] Finally, the total loss function of the face anti-counterfeiting network model is calculated as:

[0119]

[0120] where, , is the first weight parameter, is the second weight parameter.

[0121] The final feature extraction and classification network is trained by minimizing the function :

[0122]

[0123] Step 5: After training, the face anti-counterfeiting network model can be used to determine whether the image from the test set is a real face or a fake face.

[0124] The present application evaluates the generalization effect of the model on the test set, inputs face samples from four data sets as the training set, learns the network parameters of the feature extractor and the classifier under the joint supervision of the loss functions of the intra-domain distribution enhancement module and the inter-domain distribution fusion module, and discriminates between true and false faces from the test set. Figure 2The attack samples detected when performing cross-dataset cross-attack type testing under the O&C&I&M->W setting, under the action of the distribution fusion module, the distribution enhancement module, and the combination of the two modules are shown. The results show that the distribution fusion module has good detection effect on attack samples under uniform background daylight lighting. The distribution enhancement module has good detection effect on attack samples under uniform background bottom LED lighting. The combination of the two modules has good detection effect on attack samples under complex background.

[0125] Referring to Figure 3 Another embodiment of the present application provides a face anti-fraud generalization system based on distribution robustness learning, comprising:

[0126] A face image acquisition module is configured to acquire a face image to be detected.

[0127] A face anti-fraud network model calculation module is configured to input the face image to be detected into a face anti-fraud network model for calculation, and the face anti-fraud network model outputs a face result, thereby realizing face anti-fraud generalization.

[0128] The face anti-fraud network model is obtained by network training on the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

[0129] Another embodiment of the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to realize the face anti-fraud generalization method based on distribution robustness learning.

[0130] Another embodiment of the present application provides a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to realize the face anti-fraud generalization method based on distribution robustness learning.

[0131] In the present application, face anti-fraud datasets are collected, and the datasets are preprocessed. The images in the training set are subjected to feature extraction, and pseudo labels are calculated. An intra-domain distribution enhancement module for learning robustness distribution is constructed, and an inter-domain distribution fusion module for fusing multiple training distributions is constructed. Network training is performed based on the intra-domain distribution enhancement module and the inter-domain distribution fusion module. After training, the face anti-fraud network model can determine whether the images from the test set are real faces or fake faces. The present application enhances the training distribution in various dimensions of main conditions (such as illumination, facial appearance, and camera quality) and attack types (such as new 3D masks made of new materials), so that the enhanced distribution covers part of the test distribution as much as possible, greatly alleviates the problem of distribution shift, and enables the model to have good generalization ability in unknown attack types and unknown scenes.

[0132] The above description only illustrates the best mode of the present application and cannot be understood as a limitation of the claims. The present application is not limited to the above embodiments and the specific construction allows variations. Any variations made within the scope of the independent claims are within the scope of the present application.

[0133] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

Claims

1. A face anti-spoofing generalization method based on degenerate robustness learning, characterized in that, Includes the following steps: Acquire the image of the face to be detected; The face image to be detected is input into the face anti-spoofing network model, and the face anti-spoofing network model outputs the face result, thus realizing face anti-spoofing generalization; Among them, the face anti-spoofing network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module; The face anti-spoofing network model is established through the following process: We obtained a face anti-spoofing dataset, and then preprocessed the dataset using the MTCNN algorithm, and then divided it into training and testing sets. Feature extraction is performed on the images in the training set, and pseudo-labels are calculated; The joint distribution of facial samples is estimated based on pseudo-labels to obtain the joint distribution of facial samples; The joint distribution of the estimated face samples in each domain is enhanced to obtain the distribution enhancement loss; The joint distribution of each pair of face samples from different domains is mapped to a common space using a Kantorovich potential network. The dual optimal transmission distance between each pair of distributions is calculated using the dual optimal transmission distance. Based on the dual optimal transmission distance between each pair of distributions, the distribution fusion loss is calculated. The total loss function of the face anti-spoofing network model is obtained based on the distributed augmentation loss and the distributed fusion loss. The total loss function of the face anti-spoofing network model The calculation formula is: In the formula, For cross-entropy loss, To enhance the distribution loss, For distributed fusion loss, As the first weight parameter, This is the second weighting parameter; Cross-entropy loss is calculated using the following formula: In the formula, For network parameters, For sample random variables, For the number of domains, For the first The number of samples in each domain For tags The corresponding one-hot vector, For the first The first in the domain One sample, For the field number, Number the sample; No. The sample at the th The distribution enhancement loss in each domain is calculated using the following formula: In the formula, For balancing parameters, The KL divergence of the sample labels. For adversarial examples, This represents a triple marginal loss. The distribution augmentation loss for all training samples across all domains is expressed as: In the formula, The number of adversarial samples generated for each sample For the field number, The numbering of the adversarial sample; The inter-domain distribution fusion loss is calculated using the following formula: All training distributions The distributed fusion loss is expressed as: In the formula, For Kantorovich potential network, For from the Random variables of samples in a domain, For from the Random variables of samples in a domain, It is a mean function. For the first The probability distribution of each domain, For the first The probability distribution of each domain, It is a constant; The Kantorovich potential network is updated by optimizing the following objective function: The distribution fusion loss is calculated through the following process: The dual optimal transmission distance between each pair of distributions is calculated using the dual optimal transmission distance; based on the dual optimal transmission distance between each pair of distributions, the distribution fusion loss is calculated; the dual optimal transmission distance between each pair of distributions... Calculated using the following formula: In the formula, For from the Random variables of samples in a domain, For from the Random variables of samples in a domain, It is a mean function. For the first The probability distribution of each domain, For the first The probability distribution of each domain, It is a constant.

2. The face anti-spoofing generalization method based on degenerate robust learning according to claim 1, characterized in that, Pseudo-tags are obtained through the following process: The training set includes labeled training data; Each facial sample in the labeled training data is mapped to the facial sample feature space by a feature extraction function to obtain sample features. Then, the sample features are mapped to the sample feature label space by a classifier to obtain pseudo-labels.

3. A face anti-spoofing generalization system based on degenerate robust learning, characterized in that, include: The face image acquisition module is used to acquire the face image to be detected; The face anti-spoofing network model calculation module is used to input the face image to be detected into the face anti-spoofing network model for calculation. The face anti-spoofing network model outputs the face result, realizing face anti-spoofing generalization. The face anti-spoofing network model is obtained by training the intra-domain distribution enhancement module and the inter-domain distribution fusion module.

4. 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, it implements the face anti-spoofing generalization method based on degenerate robust learning as described in any one of claims 1 to 2.

5. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the face anti-spoofing generalization method based on degenerate robust learning as described in any one of claims 1 to 2.

Images