A palm vein image defense method based on generative adversarial network

Abstract

The present invention belongs to the field of image defense technology, and relates to a palm vein image defense method based on a generative adversarial network. The palm vein image defense method includes data preprocessing, generative adversarial network training and test matching; SwinGAN, a DiscoGAN-like cross-domain relationship discovery generative adversarial network, is constructed based on the Unet structure of Swin Transformer to realize palm vein recognition and SwinGAN adversarial attack defense; adversarial attack defense is realized by a generator, a discriminator and a loss function; the generator includes shallow convolution, downsampling, upsampling and Swin Transformer; the method improves the feature extraction capability of the generative adversarial network and obtains a generator for defending against adversarial attacks through training with an improved DiscoGAN training framework, and reconstructs the original image that eliminates adversarial disturbances by performing image preprocessing at the input image stage.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing and image defense, and relates to a palm vein image defense system method based on a generative adversarial network. Background Art

[0002] At present, compared with the widely used traditional biometric technologies, such as fingerprint recognition and face recognition, palm vein authentication technology based on the physiological characteristics inside the palm has demonstrated its unique technical advantages. The texture of the palm vein is hidden in the deep tissue of the human skin, and the texture structure is complex, which cannot be directly observed by the naked eye.

[0003] However, despite the many advantages of palm vein recognition technology, no technology is perfect. Palm vein recognition technology also faces a series of challenges in practical applications. In particular, with the vigorous development of machine learning and artificial intelligence technologies, various attacks on machine learning models have also emerged, especially adversarial attacks, which can cause the misclassification of legitimate images through carefully designed small perturbations. This attack method seriously reduces the accuracy of the vein recognition system and also poses new security challenges to the vein recognition system. Some defense methods are urgently needed to protect the recognition system. These challenges have also prompted researchers to begin studying how to defend against these adversarial attacks and protect the recognition system.

[0004] In recent years, many scholars have studied adversarial defense and proposed many defense measures to mitigate adversarial attacks. The mainstream defense methods are divided into two methods: adversarial training and perturbation filtering. Adversarial training AT and a series of improved training methods after it are defense methods to enhance the robustness of deep learning models, which are mainly achieved by incorporating adversarial samples into the training process of the model. Another strategy focuses on filtering the perturbation noise in adversarial samples so as to reproduce legitimate samples before inputting them into the deep learning model for prediction. However, these existing defense methods all have some shortcomings. First, the existing methods are generally designed to be tested on benchmark datasets such as CIFAR-10, MNIST, and Fashion MNIST, and the performance of the methods on real datasets has not been deeply explored. The performance of palm vein recognition defense is relatively poor. Secondly, methods such as adversarial training AT require complex and computationally intensive retraining of the recognition network, which is not conducive to the use of actual palm vein recognition scenarios. In addition, these defense methods have weak generalization capabilities and can only defend against specific attack methods, and cannot adapt to the ever-evolving more powerful adversarial attacks.

[0005] Therefore, this application is committed to improving and solving the above-mentioned problems, and proposes a new defense method based on the Swin Transformer-based generative adversarial network, which aims to improve recognition accuracy, reduce computational complexity, and solve different types of high-intensity adversarial attacks, making it more suitable for various practical application scenarios. Summary of the invention

[0006] The purpose of the present invention is to address the existing defense methods for adversarial attacks in the scenario of palm vein recognition, which have the defects of low recognition accuracy after defense, high computational complexity, inability to solve different types of adversarial attacks and insufficient generalization ability. To this end, a palm vein image defense method based on a generative adversarial network is proposed. The method uses a generator based on SwinTransformer, a Unet structure based on Swin Transformer to improve the feature extraction ability of the generative adversarial network, and a generator is obtained by training through an improved DiscoGAN training framework for defense against adversarial attacks. At the input image stage, the generator is used to perform image preprocessing to reconstruct the original image that eliminates adversarial disturbances.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme:

[0008] As a first aspect of the present invention, a palm vein image defense system based on a generative adversarial network is proposed, which has two working modes: a model training mode and a matching evaluation mode; it includes a data preprocessing module, a generative adversarial network and a recognition module; the generative adversarial network includes a defense generator submodule, an attack generator submodule, a discriminator submodule and a loss function submodule; the data preprocessing module includes a data set classification module and an FGSM adversarial sample generation submodule;

[0009] The data preprocessing module is used to preprocess and classify the input data; the preprocessing is different according to the model training mode and the matching evaluation mode;

[0010] Preprocessing of the model training mode: the original palm vein image is received through the data set classification module and the data set is classified to obtain the classified original data set; the FGSM adversarial sample generation submodule is used to generate FGSM adversarial samples, and then the generated adversarial sample data set and the total image of the classified original data set are input into the generative adversarial network;

[0011] Preprocessing of the matching evaluation mode: the data preprocessing module receives the original palm vein image, passes it through the data set classification module to obtain the classified original data set, and then inputs the image of the test data set in the original data set to the generative adversarial network;

[0012] In the model training mode of the generative adversarial network, the received original palm vein image is reconstructed into a new image through the attack generator submodule and the defense generator submodule, and then the adversarial sample image is reconstructed into a new image through the attack generator submodule and the defense generator submodule. Then the generated image, the original adversarial sample and the pure image are input into the discriminator submodule to judge the quality of the two generated images; the loss function submodule continuously optimizes the defense generator and the attack generator based on the quality of the two generated images to obtain the parameters of the trained defense generator submodule and the attack generator submodule;

[0013] The matching evaluation mode of the generative adversarial network uses the trained parameters of the defense generator submodule and does not calculate the loss function. It only reconstructs the image through the defense generator network and then calculates the recognition accuracy through the classifier.

[0014] In the model training mode of the recognition module, after the training of the generative adversarial network is completed, the defense generator is fixed to the parameters of the trained defense generator submodule. In the matching evaluation mode of the recognition module, the test samples after the aforementioned image preprocessing are input into the defense generator submodule, and after reconstruction, they are input into the recognition classifier to evaluate the recognition accuracy.

[0015] The data preprocessing module, generative adversarial network and recognition module all operate differently due to the model training mode and the matching evaluation mode.

[0016] The generative adversarial network is a deep hashing network, namely the SwinGAN model, which relies on the Unet generator based on the Swin Transformer to build a DiscoGAN-like cross-domain relationship generation;

[0017] The SwinGAN model is built based on the architecture of the DiscoGAN model and has a model training mode and a matching evaluation mode.

[0018] The defense generator submodule and the attack generator submodule are collectively referred to as generator submodules. The generator submodule relies on the Unet architecture of Transformer, including Swin Transformer units, shallow convolution units, downsampling units and upsampling units.

[0019] As a second aspect of the present invention, a palm vein image defense method based on a generative adversarial network is proposed, comprising the following steps:

[0020] S1. For the original palm vein training data set, a preprocessing method is used to obtain the palm vein ROI and input it into the palm vein recognition module as training data;

[0021] S2. Build a generative adversarial network, use the original palm vein training data set and the adversarial data set after the FGSM (= 0.01) attack to train the network, and obtain the image reconstruction defense generator model, specifically:

[0022] S21, initializing the generator module and the discriminator module, and giving preset parameter values;

[0023] S22, first train the attack generator module, then use the attack generator to learn and imitate the adversarial perturbations of the FGSM-like attack and continuously adjust the parameters of the attack generator, so that the adversarial perturbation samples generated by the attack generator are more threatening;

[0024] S23, training the defense generator again, so that the defense generator has the ability to reconstruct the adversarial sample into a clean original sample;

[0025] S24, repeatedly and alternately performing the attack generator module training and the defense generator module training of S22 and S23, until the attack capability of the attack generator is strong and the defense generator can successfully reconstruct an image that deceives the discriminator, then fix the parameters of the defense generator and the generator module, and derive the parameter value of the defense generator;

[0026] The derived defense generator is an image reconstruction defense generator model;

[0027] S3. For the trained generative adversarial network, use the obtained defense generator module and corresponding parameters, connect the pre-trained classifier model, build a complete palm vein recognition module, test the recognition accuracy obtained by the defense generator and fine-tune the SwinGAN training parameters accordingly;

[0028] S4. After obtaining a high recognition accuracy, all parameters of the defense generator are finally fixed, and the defense generator is connected to the final classifier model to obtain the palm vein recognition and classification system protected by the SwinGAN defense system.

[0029] The preprocessing method described in S1 is to extract ROI and obtain a palm vein identifier with better recognition performance as an indicator for fine-tuning the parameters of the generative adversarial network.

[0030] The attack generator module training described in S22 specifically uses the original image that has not been adversarially disturbed as input to the attack generator.

[0031] S23 the defense generator training, that is, using the adversarial sample that has been attacked by FGSM to input into the defense generator module; the attack generator and the defense generator both belong to the generator submodule;

[0032] The generator submodule includes a SwinTransformer unit, a downsampling unit, an upsampling unit and a shallow convolution unit.

[0033] The working process of the generator submodule includes the following sub-steps:

[0034] Step A: First perform a shallow convolution unit;

[0035] Step B: Repeat the SwinTransformer unit and the downsampling unit three times after the shallow convolution unit;

[0036] Step C, perform the bottom bottleneck layer calculation again through a two-layer deep SwinTransformer unit to reduce the computational complexity and amount;

[0037] Step D, after the two-layer deep SwinTransformer unit, the SwinTransformer unit and the upsampling unit are repeated three times to restore the dimensions and pixels of the original image;

[0038] Step E: Finally, a shallow convolution layer is used to reconstruct the image generated by the generator.

[0039] In the specific implementation, a variety of white-box attacks and black-box attacks are used to attack the palm vein images to form adversarial attack samples, test the recognition and classification system, and demonstrate the defense effect;

[0040] The recognition methods of the palm vein recognition system protected by SwinGAN are mainly:

[0041] Step 1: Given a subject's palm vein image, use a data preprocessing method to obtain data output to the defense generator module. The data preprocessing method here is data segmentation.

[0042] Step 2: Input the preprocessed data into the trained defense generator module proposed in this method to obtain the reconstructed defense clean image;

[0043] Step 3: For the reconstructed image, use the recognition module to perform recognition matching to obtain the recognition accuracy.

[0044] Beneficial Effects

[0045] Compared with the prior art, the method for palm vein attack defense described in the present invention has the following beneficial effects:

[0046] 1. The method has the advantage of high recognition accuracy;

[0047] 2. The method relies on the generative adversarial network proposed by the present invention to have advanced defense capabilities against various black-and-white box attacks, and has the advantages of strong generalization and only one training required;

[0048] 3. The method relies on the Swin-Unet generator method proposed in the present invention, and has the advantages of high clarity and strong feature reconstruction capability when generating pictures;

[0049] 4. The method is plug-and-play and has the advantages of being convenient and fast. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a workflow diagram for training a generative adversarial network defense method;

[0051] Figure 2 It is a workflow diagram for identifying and matching target palm vein images after defense;

[0052] Figure 3 is a flow chart of a method for generating countermeasures against network attacks;

[0053] Figure 4 is a flow chart of the method for generating adversarial network defenses;

[0054] Figure 5 This is the structure diagram of the Swin-Unet generator. DETAILED DESCRIPTION

[0055] The method for identifying palm prints and palm veins of the present invention is further illustrated and described in detail below in conjunction with the accompanying drawings and embodiments.

[0056] Example 1

[0057] This example uses the Tongji palm print and vein dataset to train a deep hashing network model.

[0058] Mode 1: Train SwinGAN to generate adversarial networks. The process can be referred to Figure 1 :

[0059] S1: Dataset division: The Tongji-PV database is divided into training set, test set and validation set. The pure original image training set B1 has 8,400 images (14 images × 2 hands × 300 volunteers), the test set B2 for generating adversarial attack samples has 2,400 images (4 images × 2 hands × 300 volunteers), and the validation set B3 has 1,200 images (2 images × 2 hands × 300 volunteers).

[0060] S2: Data preprocessing: In the data preprocessing stage, FGSM attack is performed on each image in the pure original image training set to generate an adversarial sample dataset.

[0061] S3: Model training: Use the preprocessed data to train the SwinGAN generative adversarial network model. For the overall training process, refer to Figure 1 , the attack training process is as follows Figure 3 , the defense training process is as follows Figure 4 , the generator module is as follows Figure 5 . First, train the attack generator and use the discriminator to evaluate the performance of the generator. Next, train the attack generator and use the discriminator to evaluate the performance of the generator, and update the two generator parameters through the loss function. The two processes are trained repeatedly and interactively until the trained SwinGAN network is obtained. The recognition accuracy of the trained defense network is tested using the validation set. The recognition accuracy can be achieved by reconstructing the defense generator under white-box attack to achieve at least 93.6%, and the recognition accuracy can be achieved by reconstructing the defense generator under black-box attack to achieve at least 94.3%, which is far higher than the traditional method.

[0062] S4: After finding the best-performing defense generator parameters, the corresponding generator parameters are fixed, and the model structure and parameters are saved and transferred to mode 2 for use;

[0063] Mode 2: Defense and match the target image, process reference Figure 2 :

[0064] S1: Data preprocessing: For a given subject’s palm vein image, perform multiple white-box attacks and HSJA attacks to obtain the corresponding adversarial sample dataset.

[0065] S2: Recognition accuracy calculation: The preprocessed image is input into the trained defense generator model to obtain the reconstructed target image.

[0066] S3: Calculation of recognition accuracy: Four recognition matching methods are used to perform recognition matching, obtain the label of the target image and calculate the recognition accuracy. For the five white-box attacks and the HSJA black-box attack, the recognition accuracy after defense is higher than the existing more advanced defense methods, at least 15.7% higher than the existing methods, and the defense performance is very good.

[0067] Example 2

[0068] This embodiment describes in detail the connection relationship and test results of a method for defending against palm vein attacks in the specific implementation of the present invention.

[0069] The palm vein adversarial defense method relies on the Unet generator structure based on Swin Transformer to construct a DiscoGAN-like cross-domain relationship discovery generative adversarial network SwinGAN, and the method includes a palm vein recognition module and a SwinGAN adversarial attack defense module;

[0070] The SwinGAN adversarial attack defense module is divided into a generator sub-module, a discriminator sub-module, and a loss function sub-module; the generator module includes a shallow convolution unit, a downsampling unit, an upsampling unit, and a Swin Transformer unit.

[0071] The SwinGAN adversarial attack defense includes data preprocessing, generative adversarial network training and testing, and matching algorithms;

[0072] The data preprocessing, generative adversarial network training and testing, and matching algorithms all have corresponding operations according to the different model training modes and matching evaluation modes;

[0073] The palm vein detection method including SwinGAN adversarial defense is based on the SwinGAN palm vein defense system, including a data preprocessing module, a generative adversarial network and an identification and matching module;

[0074] The deep hashing network is constructed through an architecture based on the DiscoGAN model;

[0075] The preprocessing module is used to preprocess and classify the data input to the SwinGAN; the preprocessing varies according to the training mode and the recognition test mode of the SwinGAN model;

[0076] Preprocessing for generative adversarial network model training, including dataset classification module and FGSM adversarial sample generation submodule;

[0077] Data preprocessing module, including data set classification module and FGSM adversarial sample generation submodule;

[0078] In the model training mode, the data preprocessing receives the original palm vein image, and then goes through the data set classification and FGSM adversarial sample generation in sequence, and then the generated adversarial sample data set and the classified original data set images are input into the SwinGAN generative adversarial network;

[0079] In the matching test mode, the data preprocessing module receives the original palm vein image, and after passing through the dataset classification module, the image of the image test dataset is input to the generative adversarial network;

[0080] SwinGAN Generative Adversarial Network, including defense generator submodule, attack generator submodule, discriminator submodule and loss function submodule;

[0081] In the model training mode, the received original palm vein image is reconstructed into a new image through the attack generator submodule and the defense generator submodule, and then the adversarial sample image is reconstructed into a new image through the attack generator submodule and the defense generator submodule. Then the generated image, the original adversarial sample and the pure image are input into the discriminator submodule to judge the generation quality;

[0082] The loss function submodule will evaluate the quality of the two images generated by the generator and continuously optimize the parameters of the model generator based on this;

[0083] In the matching test mode, only the parameters of the defense generator trained in the generative adversarial network are used, and the loss function is not calculated. The image is only reconstructed through the defense generator network, and then the recognition accuracy is calculated through the palm vein recognition classifier;

[0084] Generate preprocessing for adversarial network model matching test, including dataset classification submodule;

[0085] The SwinGAN generative adversarial network module includes a defense generator submodule, an attack generator submodule, a discriminator submodule and a loss function submodule;

[0086] The generator submodule proposed in the present invention is a Transformer-based Unet architecture, including a SwinTransformer submodule, a shallow convolution submodule, a downsampling module, and a double upsampling module;

[0087] The recognition classifier submodule includes four types of recognizer submodules;

[0088] After the SwinGAN network training is completed, the defense generator is fixed as the defense submodule in the final test;

[0089] In the matching test mode, the test samples after the aforementioned image preprocessing are input into the defense submodule, and after reconstruction, they are input into the recognition classifier to evaluate the recognition accuracy;

[0090] The method for training palm vein recognition defense given palm vein images recorded by a number of subjects comprises the following steps:

[0091] S1. For a given palm vein training data set, a preprocessing method is used to obtain the palm vein ROI input to the palm vein recognition module as training data. The preprocessing method here is ROI extraction, and a palm vein recognizer with better recognition performance is obtained as an indicator for fine-tuning the parameters of the generative adversarial network.

[0092] S2. Build the SwinGAN generative adversarial network, use the original palm vein training dataset and the adversarial dataset after the FGSM (= 0.01) attack to train the network, and obtain the image reconstruction defense generator model.

[0093] S21. Initialize the generator module and the discriminator module and give preset parameter values.

[0094] S22, the next stage is to first train the attack generator module, using the original image that has not been perturbed by the adversarial as input to the attack generator G AB , then the attack generator learns to imitate and generate adversarial perturbations (p') similar to FGSM attacks, and enables the subsequent attack generator to continuously adjust its parameters, so that G AB The generated adversarial perturbation samples are more threatening.

[0095] S23. Then, the defense generator is trained by inputting the adversarial samples that have been attacked by FGSM into the defense generator module to train the defense generator. The purpose is to reconstruct the adversarial samples into clean original samples through the defense generator.

[0096] S24. Repeat the training of the attack generator module and the defense generator module alternately until the attack capability of the attack generator is strong and the defense generator can successfully reconstruct an image that deceives the discriminator.

[0097] S25. Fix the parameters of the generator module and export the parameter values ​​of the defense generator module.

[0098] The generator submodule of SwinGAN consists of four units: SwinTransformer unit, downsampling unit, upsampling unit, and shallow convolution unit.

[0099] First perform a shallow convolution unit;

[0100] Repeat the SwinTransformer unit and the downsampling unit three times after the shallow convolution unit;

[0101] The bottom bottleneck layer calculation is then performed through a two-layer deep SwinTransformer unit to reduce the computational complexity and amount.

[0102] After the two-layer deep SwinTransformer unit, the SwinTransformer unit and the upsampling unit are repeated three times to restore the dimensions and pixels of the original image.

[0103] Finally, a shallow convolution layer is used to reconstruct the image generated by the generator.

[0104] S3. For the trained generative adversarial network, use the defense generator module and corresponding parameters obtained in the previous step, connect the pre-trained classifier model, build a complete palm vein recognition module, test the recognition accuracy obtained by using the defense generator proposed in the present invention, and fine-tune the SwinGAN training parameters accordingly.

[0105] S4. After obtaining a high recognition accuracy rate, all parameters of the defense generator are finally fixed, and the defense generator is connected to the final classifier model to obtain the complete palm vein recognition and classification system protected by the SwinGAN defense system proposed in the present invention.

[0106] S5. Use a variety of white-box attacks and black-box attacks to attack palm vein images, form adversarial attack samples, test the recognition and classification system, and demonstrate the defense effect.

[0107] The recognition method of the palm vein recognition system protected by SwinGAN includes the following steps:

[0108] Step 1: Given a subject's palm vein image, use a data preprocessing method to obtain data output to the defense generator module. The data preprocessing method here is data segmentation.

[0109] Step 2: Input the preprocessed data into the trained defense generator module proposed in this method to obtain the reconstructed defense clean image;

[0110] Step 3: For the reconstructed image, use the recognition module to perform recognition matching to obtain the recognition accuracy;

[0111] The performance of white-box attack and black-box attack are shown in Table 1 and Table 2 respectively.

[0112] Table 1 Performance of the proposed system in white-box attack on palm vein images

[0113]

[0114] Table 2 Performance of the proposed system in black box attack on palm vein images

[0115]

[0116] The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A palm vein image defense system based on generative adversarial network, characterized in that: It has two working modes, namely model training mode and matching evaluation mode; the palm vein image defense system includes a data preprocessing module, a generative adversarial network and a recognition module; the generative adversarial network includes a defense generator submodule, an attack generator submodule, a discriminator submodule and a loss function submodule; the data preprocessing module includes a data set classification module and a FGSM adversarial sample generation submodule; The data preprocessing module is used to preprocess and classify the input data; The preprocessing is different according to the model training mode and the matching evaluation mode; Preprocessing of the model training mode, receiving the original palm vein image through the data set classification module and performing data set classification to obtain the classified original data set; then generating FGSM adversarial samples through the FGSM adversarial sample generation submodule, and then inputting the generated adversarial sample data set and the total image of the classified original data set into the generative adversarial network; Preprocessing of the matching evaluation mode, the data preprocessing module receives the original palm vein image, obtains the classified original data set after passing through the data set classification module, and then inputs the image of the test data set in the original data set into the generative adversarial network; In the model training mode of the generative adversarial network, the received original palm vein image is reconstructed into a new image through the attack generator submodule and the defense generator submodule, and then the adversarial sample image is reconstructed into a new image through the attack generator submodule and the defense generator submodule, and then the generated image, the original adversarial sample and the pure image are input into the discriminator submodule to judge the quality of the two generated images; the loss function submodule continuously optimizes the defense generator and the attack generator according to the quality of the two generated images, and obtains the parameters of the trained defense generator submodule and the attack generator submodule; The matching evaluation mode of the generative adversarial network uses the trained parameters of the defense generator submodule and does not calculate the loss function. It only reconstructs the image through the defense generator network and then calculates the recognition accuracy through the classifier. Identify the model training mode of the module, and after the training of the generative adversarial network is completed, fix the parameters of the defense generator to the trained defense generator submodule; The matching evaluation mode of the recognition module inputs the test samples after the aforementioned image preprocessing into the defense generator submodule, and after reconstruction, inputs them into the recognition classifier to evaluate the recognition accuracy.

2. According to claim 1, a palm vein image defense system based on generative adversarial network is characterized in that: The data preprocessing module, generative adversarial network and recognition module all operate differently due to the model training mode and the matching evaluation mode.

3. According to claim 2, a palm vein image defense system based on generative adversarial network is characterized in that: The generator submodule relies on the Unet architecture of Transformer, including Swin Transformer units, shallow convolution units, downsampling units and upsampling units.

4. The palm vein image defense system based on generative adversarial network according to claim 3 is characterized in that: The generative adversarial network is a deep hashing network, namely the SwinGAN model, which relies on the Unet generator based on the Swin Transformer to build a DiscoGAN-like cross-domain relationship generation.

5. The palm vein image defense system based on generative adversarial network according to claim 4 is characterized in that: The SwinGAN model is built based on the architecture of the DiscoGAN model, including a model training mode and a matching evaluation mode.

6. A palm vein image defense method based on generative adversarial network, characterized in that: The steps include: S1. For the original palm vein training data set, a preprocessing method is used to obtain the palm vein ROI and input it into the palm vein recognition module as training data; S2. Build a generative adversarial network, use the original palm vein training dataset and the adversarial dataset after FGSM attack to train the network, and obtain the image reconstruction defense generator model, specifically: S21, initializing the generator module and the discriminator module, and giving preset parameter values; S22, first train the attack generator module, then use the attack generator to learn and imitate the adversarial perturbations of the FGSM-like attack and continuously adjust the parameters of the attack generator, so that the adversarial perturbation samples generated by the attack generator are more threatening; S23, training the defense generator again, so that the defense generator has the ability to reconstruct the adversarial sample into a clean original sample; S24, repeatedly and alternately performing the attack generator module training and the defense generator module training of S22 and S23, until the attack capability of the attack generator is strong and the defense generator can successfully reconstruct an image that deceives the discriminator, then fix the parameters of the defense generator and the generator module, and derive the parameter value of the defense generator; The derived defense generator is an image reconstruction defense generator model; S3. For the trained generative adversarial network, use the obtained defense generator module and corresponding parameters, connect the pre-trained classifier model, build a complete palm vein recognition module, test the recognition accuracy obtained by the defense generator and fine-tune the SwinGAN training parameters accordingly; S4. After obtaining a high recognition accuracy, all parameters of the defense generator are finally fixed, and the defense generator is connected to the final classifier model to obtain the palm vein recognition and classification system protected by the SwinGAN defense system.

7. The palm vein image defense method based on generative adversarial network according to claim 6 is characterized in that: The preprocessing method described in S1 is to extract ROI and obtain a palm vein identifier with better recognition performance as an indicator for fine-tuning the parameters of the generative adversarial network.

8. The palm vein image defense method based on generative adversarial network according to claim 7 is characterized in that: The attack generator module training described in S22 specifically uses the original image that has not been adversarially disturbed as input to the attack generator.

9. The palm vein image defense method based on generative adversarial network according to claim 8, characterized in that: S23 the defense generator training, that is, using the adversarial sample that has been attacked by FGSM to input into the defense generator module; the attack generator and the defense generator both belong to the generator submodule; The generator submodule includes a SwinTransformer unit, a downsampling unit, an upsampling unit and a shallow convolution unit.

10. The palm vein image defense method based on generative adversarial network according to claim 9, characterized in that: The working process of the generator submodule includes the following sub-steps: Step A: First perform a shallow convolution unit; Step B: Repeat the SwinTransformer unit and the downsampling unit three times after the shallow convolution unit; Step C, perform the bottom bottleneck layer calculation again through a two-layer deep SwinTransformer unit to reduce the computational complexity and amount; Step D, after the two-layer deep SwinTransformer unit, the SwinTransformer unit and the upsampling unit are repeated three times to restore the dimensions and pixels of the original image; Step E: Finally, a shallow convolution layer is used to reconstruct the image generated by the generator.

Images