Palm vein representation attack detection method and system based on lightweight neural network

Abstract

The invention relates to the technical field of palm vein recognition, in particular to an efficient palm vein representation attack detection method and system based on a lightweight neural network. The method comprises the following steps: acquiring an image to be identified; inputting a to-be-recognized image into the Xception entry stream, and performing preliminary feature extraction on the to-be-recognized image to obtain an initial feature; inputting the initial features into Block modules of a star network, and further mining high-order nonlinear features in the image by each Block module through convolution and star operation; and performing classification identification on the high-order nonlinear features to obtain classification result data. Real and forged palm vein images are distinguished by capturing vein texture features in the images, a single-channel near-infrared palm vein image is input, and a binary result of true and false classification is output.

Description

Technical Field

[0001] The present invention relates to the technical field of palm vein recognition, and in particular to an efficient palm vein representation attack detection method and system based on a lightweight neural network. Background Art

[0002] Biometrics utilize unique physiological or behavioral characteristics of the human body to authenticate identity. Palm vein recognition, particularly due to its uniqueness and difficulty in forging, has become a research hotspot in the biometric field. Palm vein recognition uses the shape and distribution of veins within the palm to identify individuals. Because veins are located beneath the skin and are unique, they are typically difficult to obtain and replicate through external means. Therefore, compared to technologies such as fingerprint and facial recognition, palm vein recognition offers greater security against counterfeiting. Palm vein recognition typically relies on illuminating the palm with near-infrared light. Due to the strong absorption of near-infrared light by hemoglobin, the blood vessels in the palm veins appear dark under near-infrared light. Near-infrared imaging is then used to capture texture images of the palm veins. By extracting and analyzing the features of these texture images, unique identification of individuals can be achieved. This method offers the advantages of being contactless and highly private, demonstrating great potential in various identity authentication scenarios.

[0003] However, palm vein recognition technology still faces some challenges in preventing spoofing attacks. With the development of technologies such as 3D printing and image processing, attackers can try to deceive palm vein recognition systems in a variety of ways. For example, using printed near-infrared images of palm veins, drawing images that simulate palm vein textures, or even making three-dimensional molds of the palm, in order to forge real palm vein features and thus bypass the recognition system. The emergence of these attack methods has exposed the shortcomings of current palm vein recognition in its anti-counterfeiting attack capabilities. In response to the anti-counterfeiting needs of palm vein recognition technology, the present invention provides an efficient palm vein representation attack detection method and system based on a lightweight neural network. Summary of the Invention

[0004] In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides an efficient palm vein representation attack detection method and system based on a lightweight neural network.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] In a first aspect, in one embodiment provided by the present invention, an efficient palm vein representation attack detection method based on a lightweight neural network is provided, the method comprising the following steps:

[0007] Obtaining an image to be recognized;

[0008] The image to be recognized is input into the Xception input stream, and preliminary feature extraction is performed on it to obtain initial features. The initial features are then input into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations.

[0009] Classify and identify high-order nonlinear features to obtain classification result data.

[0010] As a further solution of the present invention, the Xception inlet stream performs preliminary feature extraction on the image to be identified to obtain initial features, and inputs the initial features into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations, including:

[0011] The BSCnvU blueprint convolution is used in the inlet flow to replace the original depth-separable convolution. The convolution layer extracts the basic features in the image by combining point convolution and depth-wise convolution, and a batch channel normalization (BCN) layer is added after the convolution layer for channel normalization.

[0012] As a further solution of the present invention, the Xception inlet stream performs preliminary feature extraction on the image to be identified to obtain initial features, and inputs the initial features into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations, including:

[0013] The basic features are further extracted through two residual blocks, each of which uses BSConvU convolution units and maximum pooling operations to obtain high-order nonlinear features.

[0014] As a further solution of the present invention, the classifying and identifying high-order nonlinear features to obtain classification result data includes:

[0015] The feature map is reduced to a fixed size through adaptive average pooling (AdaptiveAvgPool2d), and then converted into a one-dimensional vector through flattening operation (Flatten);

[0016] The fully connected layer outputs a feature vector of two categories of results based on the one-dimensional vector, and calculates the feature vector through the softmax function to obtain the final predicted probability value, that is, the classification result data.

[0017] In a second aspect, in another embodiment provided by the present invention, an efficient palm vein representation attack detection system based on a lightweight neural network is provided, the system comprising: an input preprocessing module, a feature extraction module, and a classification module;

[0018] The input preprocessing module is used to obtain the image to be recognized;

[0019] The feature extraction module is composed of the inlet flow structure of the Xception model and the Block module of the StarNet model, and is used to extract features of the image to be identified to obtain feature data;

[0020] The classification module is used to classify and identify high-order nonlinear features to obtain classification result data.

[0021] As a further solution of the present invention, the feature extraction module is used to input the image to be identified into the Xception input stream, perform preliminary feature extraction on it to obtain initial features; then input the initial features into the Block module of the star network, and each Block module further mines the high-order nonlinear features in the image through convolution and star operations.

[0022] As a further solution of the present invention, the classification module is used to reduce the feature map to a fixed size through adaptive average pooling (AdaptiveAvgPoold), and then convert it into a one-dimensional vector through a flattening operation (Flatten); the fully connected layer outputs a feature vector of two categories of results based on the one-dimensional vector, and calculates the feature vector through a softmax function to obtain the final predicted probability value, that is, the classification result data.

[0023] As a further solution of the present invention, the system adopts a cross entropy loss function (Cross Entropy Loss) to perform binary classification tasks.

[0024] The technical solution provided by the present invention has the following beneficial effects:

[0025] The present invention uses the Mish activation function in BSConv-U as an alternative to nonlinear transformations. By introducing the smoothing property of sliding, the Mish activation function overcomes the limitations of traditional activation functions such as ReLU when processing negative values, thereby providing better gradient transfer. This improves the model's expressiveness while also enhancing its ability to learn complex patterns. By using the star network's block structure, the model can more efficiently extract subtle features from palm vein images, demonstrating greater robustness and accuracy, especially when faced with complex backgrounds and forged images. This makes the star network a significant innovation in palm vein recognition technology and provides strong support for improving anti-spoofing capabilities.

[0026] These and other aspects of the present invention will become more readily apparent in the following description of the embodiments. It should be understood that the above general description and the following detailed description are merely exemplary and explanatory and are not intended to limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 This is a flowchart of an efficient palm vein representation attack detection method based on a lightweight neural network according to an embodiment of the present invention.

[0029] Figure 2 This is a structural block diagram of an efficient palm vein representation attack detection system based on a lightweight neural network according to an embodiment of the present invention.

[0030] Figure 3 This is a diagram of a palm vein prosthesis data set acquisition device in an efficient palm vein representation attack detection system based on a lightweight neural network according to an embodiment of the present invention.

[0031] Figure 4 This is a real palm vein image.

[0032] Figure 5 This is a cropped printed palm vein image.

[0033] Figure 6 This is an uncropped printed palm vein image.

[0034] Figure 7 Collect diagram for palm mold.

[0035] Figure 8 Collect images for textured hand stencils.

[0036] Figure 9 Collect images of gloved hands.

[0037] Figure 10 The generated digital palmar vein prosthesis.

[0038] In the figure: input preprocessing module-100, feature extraction module-200, classification module-300. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0040] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0041] It should be understood that the terms used in this specification are only for the purpose of describing particular embodiments and are not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0042] Specifically, the embodiments of the present invention are further described below with reference to the accompanying drawings.

[0043] See also Figure 1 , Figure 1 This is a flowchart of an efficient palm vein representation attack detection method based on a lightweight neural network provided by an embodiment of the present invention. Figure 1 As shown, the efficient palm vein representation attack detection method based on lightweight neural network includes steps S10 to S30.

[0044] S10, obtaining an image to be identified;

[0045] S20: Input the image to be recognized into the Xception input stream and perform preliminary feature extraction to obtain initial features. Then, input the initial features into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations.

[0046] In an embodiment of the present invention, the Xception inlet stream performs preliminary feature extraction on the image to be identified to obtain initial features, and inputs the initial features into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations, including:

[0047] In the inlet flow, the BSConvU blueprint convolution is used instead of the original depth-separable convolution. This convolution layer extracts the basic features of the image by combining point convolution and depth-wise convolution. A batch channel normalization (BCN) layer is added after the convolution layer to perform channel normalization to better standardize the distribution of data.

[0048] The basic features are further extracted through two residual blocks, each of which uses BSConvU convolution units and maximum pooling operations to obtain high-order nonlinear features; these operations enable the model to retain important palm vein texture information while reducing computational complexity.

[0049] Each Block module also retains the input features through skip connections (residual connections) to avoid the gradient disappearance problem, enabling the network to perform deeper learning.

[0050] It's important to note that Blueprint Convolution (BSConv) is based on the concept of Depthwise Separable Convolution (DSC), but it introduces a more efficient feature fusion mechanism during feature extraction, making it particularly suitable for processing high-dimensional data such as images. The core idea of ​​BSConv is to decompose the traditional convolution operation into two main components: pointwise convolution and depthwise convolution. This structural design enables the network to effectively extract features while significantly reducing computational complexity and the number of parameters, thereby improving model efficiency.

[0051] The BSConv-U structure consists of point convolution and depth convolution.

[0052] 1) Point convolution: This layer performs a 1x1 convolution on the input feature map to change the number of channels of the feature while maintaining the spatial dimension. This process can effectively fuse information from different channels and prepare for subsequent depth convolution.

[0053] 2) Deep convolution: In this stage, standard convolution operations are used to process each channel independently, which can capture local features. The advantage of deep convolution is that it significantly reduces the complexity of convolution calculations, allowing the network to extract high-dimensional features while maintaining a lightweight structure.

[0054] The introduction of the BSConv-U architecture not only improves network performance but also lays the foundation for subsequent deep learning tasks. In BSConv-U, we introduce Batch Channel Normalization (BCN) as an alternative to standard batch normalization. BCN normalizes across the channel dimension, maintaining a consistent mean and variance across each channel, thereby enhancing the model's adaptability to feature distributions. This improvement more effectively maintains information flow when processing diverse features, thereby improving model training efficiency and convergence speed.

[0055] This paper uses the Mish activation function in BSConv-U as an alternative to nonlinear transformations. By introducing a sliding smoothing property, the Mish activation function overcomes the limitations of traditional activation functions like ReLU in handling negative values, thereby providing better gradient propagation. This feature improves the model's expressiveness while also enhancing its ability to learn complex patterns.

[0056] The StarNet design is inspired by the Star Operation, a concept that differs from traditional convolutional methods. It constructs high-level features through element-by-element multiplication, thereby improving the network's expressive power. This paper adopts the Block structure of StarNet to optimize the feature extraction process for palm vein images.

[0057] StarNet captures complex feature relationships by mapping input features into high dimensions. Its core concept is to utilize star operations, generating high-order features through element-by-element multiplication across channels. This operation not only enhances the model's sensitivity to input features but also improves the interaction between features, achieving greater expressive power while maintaining computational efficiency.

[0058] In StarNet, the Block module is its basic building block and is responsible for feature extraction. This module contains several key operations, as follows:

[0059] 1) Deep Convolution: The Block module first extracts local information from the input features through deep convolution. Deep convolution is performed independently within each channel, effectively capturing the spatial structure in the feature map. Compared to traditional convolution, it significantly reduces computational effort while preserving rich feature information.

[0060] 2) Star Operation: After depthwise convolution, the Block module uses a star operation to perform element-wise multiplication of the two feature maps. This process effectively captures complex relationships between features and generates high-order features. This operation improves the network's nonlinear expression capabilities, enabling the model to better handle complex pattern recognition tasks.

[0061] 3) Residual Connections: The Block module uses residual connections to add input features to features after a series of convolutions and nonlinear transformations. Residual connections can effectively alleviate the vanishing gradient problem during training, making deep networks easier to optimize.

[0062] 4) Activation Function: To enhance nonlinear capabilities, the Block module in this paper uses the Mish activation function. By combining the advantages of ReLU and the smoothness of tanh, Mish can effectively improve model performance, especially in the feature extraction stage.

[0063] It's important to note that the basic principle of Batch Channel Normalization (BCN) is to normalize the mean and variance of each channel in the channel dimension of the feature map. This method allows the network to better capture the interdependencies between channels during learning, thereby improving the model's expressiveness and generalization capabilities. Specifically, by independently calculating the statistics for each channel, BCN effectively preserves the feature information within the channel while reducing instability caused by changes in data distribution during training.

[0064] S30. Classify and identify high-order nonlinear features to obtain classification result data.

[0065] In an embodiment of the present invention, S30, classifying and identifying high-order nonlinear features to obtain classification result data, includes:

[0066] The feature map is reduced to a fixed size through adaptive average pooling (AdaptiveAvgPool2d), and then converted into a one-dimensional vector through flattening operation (Flatten);

[0067] The fully connected layer outputs a feature vector of two categories of results based on the one-dimensional vector, and calculates the feature vector through the softmax function to obtain the final predicted probability value, that is, the classification result data.

[0068] The present invention uses the Mish activation function in BSConv-U as an alternative to nonlinear transformations. By introducing the smoothing property of sliding, the Mish activation function overcomes the limitations of traditional activation functions such as ReLU when processing negative values, thereby providing better gradient transfer. This improves the model's expressiveness while also enhancing its ability to learn complex patterns. By using the star network's block structure, the model can more efficiently extract subtle features from palm vein images, demonstrating greater robustness and accuracy, especially when faced with complex backgrounds and forged images. This makes the star network a significant innovation in palm vein recognition technology and provides strong support for improving anti-spoofing capabilities.

[0069] It should be understood that although the above is described in a certain order, these steps are not necessarily performed in sequence according to the above order. Unless clearly stated herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, a part of the steps of the present embodiment may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0070] In one embodiment, see Figure 2 As shown, in an embodiment of the present invention, an efficient palm vein representation attack detection system based on a lightweight neural network is also provided. The system includes an input preprocessing module 100, a feature extraction module 200 and a classification module 300.

[0071] The input pre-processing module 100 is used to obtain an image to be identified. The image to be identified may be a near-infrared image with a resolution of 1920×1080 obtained by a device.

[0072] The feature extraction module 200 includes an inlet flow structure of an Xception model and a Block module of a StarNet model, and is used to extract features from an image to be identified to obtain feature data.

[0073] Specifically, the feature extraction method of the feature extraction module 200 includes: the image to be identified is subjected to preliminary feature extraction through the Xception entry flow (EntryFlow) to obtain initial features, wherein the BSConvU blueprint convolution is used in the entry flow to replace the original depth-separable convolution. The convolution layer extracts the basic features in the image by combining point convolution and depth convolution. At the same time, a batch channel normalization (BCN) layer is added after the convolution layer for channel normalization to better standardize the distribution of the data. Afterwards, features are further extracted through two residual blocks, each of which uses a BSConvU convolution unit and a maximum pooling operation. These operations enable the model to retain important palm vein texture information while reducing computational complexity.

[0074] The initial features are input into the Block module of StarNet. Each Block module further mines the high-order nonlinear features in the image through convolution and star operations (element-by-element multiplication).

[0075] Star operation can increase the model's ability to express complex relationships between features without significantly increasing computational complexity. Each Block module also retains input features through skip connections (residual connections), avoiding the gradient vanishing problem and enabling the network to learn at a deeper level.

[0076] The classification module 300 is used to classify and identify high-order nonlinear features to obtain classification result data.

[0077] In an embodiment of the present invention, the classification module 300 is used to reduce the feature map to a fixed size through adaptive average pooling (AdaptiveAvgPool2d), and then convert it into a one-dimensional vector through a flattening operation (Flatten); the fully connected layer outputs a feature vector of two categories of results based on the one-dimensional vector, and calculates the feature vector through the softmax function to obtain the final predicted probability value, that is, the classification result data.

[0078] In an embodiment of the present invention, the system uses a cross entropy loss function for binary classification. The output is a probability distribution for each class, and the class with the higher probability is taken as the final classification result, that is, determining whether the image is "real" or "forged."

[0079] Performance test:

[0080] A dataset was constructed and divided into training and test sets, with the training set accounting for 75% of the total dataset and the test set for 25%. 4988 real images and 4991 attack images were used. The cross-entropy loss function was used as the loss criterion, and the AdamW optimizer was used for model training, with an initial learning rate of 0.001. The learning rate scheduler used the ReduceLROnPlateau strategy to dynamically adjust the learning rate based on the validation set loss, helping the model maintain optimal performance during training. The training epoch was set to 150, and to prevent overfitting, an early stopping mechanism was implemented; if the test set loss did not improve over multiple epochs, training was terminated early.

[0081] The images in the dataset are collected by the embedded device RV1106. Figure 3 As shown in the figure, the device is equipped with four 850nm near-infrared fill lights, capturing images with a resolution of 1920×1080, all of which are single-channel near-infrared images. Near-infrared imaging technology can penetrate the skin to a certain extent, revealing the internal vein structure of the palm, thereby capturing the unique texture characteristics of palm veins. This imaging method provides high discrimination of palm vein images, making them particularly suitable for identity recognition and anti-counterfeiting detection tasks.

[0082] It should be noted that the real images in the dataset are obtained by directly collecting near-infrared images of real people's palms, which contain the unique texture features of individual palm veins. Examples of real images are as follows: Figure 4 The real images were collected under various factors, including ambient lighting conditions, palm postures, and placement distances, ensuring data diversity and authenticity, and providing the model with rich training samples to effectively identify real-person palm vein features.

[0083] It should be noted that prosthesis pictures are mainly divided into three categories: paper prostheses, mold prostheses and digital prostheses.

[0084] Paper prosthesis: Paper prosthesis is made by printing a near-infrared image of the palm veins of a real palm, aiming to simulate the scenario where an attacker can attack by mastering the palm vein texture of the original user. This type of prosthesis is further divided into two types:

[0085] ① Cropped printed image: Print out the real palm vein grayscale image and then crop it into the shape of the palm. This method imitates the attacker's attack method of copying the palm vein texture through high-quality printing technology and trying to deceive the recognition system, such as Figure 5 shown.

[0086] ② Printed uncropped image: Directly print the uncropped real palm vein image with a white background. This type is intended to study the robustness of the recognition system to prosthetic images of different shapes and backgrounds, such as Figure 6 shown.

[0087] Mold prosthesis: Mold prosthesis simulates the palm vein attack method created by bionic means. This type of prosthesis uses a variety of materials to make a palm mold and generates images through different attack methods, including:

[0088] ① Palm molds of different shapes and hardness: Use soft or hard materials to make palm molds, and use near-infrared equipment to capture the mold image. This method tests the anti-counterfeiting performance of the recognition system for palm bionics of different materials and shapes, such as Figure 7 shown.

[0089] ② Textured hand mold: Manually draw vein texture on the palm mold to simulate the details of palm veins. This attack method is more challenging because it attempts to reproduce the unique texture features of palm veins on the mold, such as Figure 8 shown.

[0090] ③ Glove prosthesis: Wearing specially designed gloves on real hands, such as gloves with vein patterns, simulates the situation where an attacker deceives the recognition system by wearing gloves with forged palm vein features, such as Figure 9 shown.

[0091] The construction of the dataset takes into account a variety of real and forged palm vein image scenarios, providing a testing environment for the anti-attack detection method proposed in this paper and helping to evaluate the model's ability to cope with various deception attacks in practical applications.

[0092] Digital prosthesis: This type of image is generated through a generative adversarial network (GAN) or a diffusion model (such as Stable Diffusion) to simulate a real palm vein image. Specifically, these deep learning models are used to first learn the feature distribution of a real palm vein image, and then generate a forged image that is highly similar to the real palm vein. In this way, the attacker can generate a palm vein image that is visually difficult to distinguish, thereby attempting to deceive the recognition system. The generated example is as follows Figure 10 shown.

[0093] By introducing digital prostheses generated by GANs or diffusion models into the dataset, the model is further challenged to test its ability to resist advanced digital forgery attacks. This type of attack represents a potential future security threat, especially as generative models continue to advance. Using this type of digital prosthesis to evaluate the robustness of recognition systems is also very important.

[0094] The model ultimately achieved optimal training and validation accuracies of 100% and 99.8% on the training and validation sets, respectively, with an average detection time of 5ms. Vein liveness detection and liveness detection metrics were also used. These metrics include the Attack Presentation Classification Error Rate (APCER), the Bona Fide Presentation Classification Error Rate (BPCER), and the Average Classification Error Rate (ACER). APCER represents the proportion of forged samples incorrectly classified as real samples by the model. BPCER represents the proportion of real samples incorrectly classified as forged samples by the model. ACER is the average of APCER and BPCER, comprehensively considering the model's recognition performance on both forged and real samples. The final data results are summarized in Table 1.

[0095] Table 1 Statistics of training and validation data

[0096]

[0097] It should be understood that, as used herein, the singular form "a" or "an" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations of one or more of the items listed in association. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0098] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples. Within the spirit of the embodiments of the present invention, the technical features of the above embodiments or different embodiments may be combined, and there are many other variations of different aspects of the above embodiments of the present invention, which are not provided in detail for the sake of clarity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the scope of protection of the embodiments of the present invention.

Claims

1. An efficient palm vein representation attack detection method based on lightweight neural network, characterized by: The method includes: Obtaining an image to be recognized; The image to be recognized is input into the Xception input stream, and preliminary feature extraction is performed on it to obtain initial features. The initial features are then input into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations. Classify and identify high-order nonlinear features to obtain classification result data.

2. The efficient palm vein representation attack detection method based on lightweight neural network according to claim 1 is characterized in that: The Xception inlet stream performs preliminary feature extraction on the image to be identified to obtain initial features, and inputs the initial features into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations, including: The BSConvU blueprint convolution is used in the inlet flow to replace the original depth-separable convolution. The convolution layer extracts the basic features in the image by combining point convolution and depth-wise convolution, and a batch channel normalization layer is added after the convolution layer to perform channel normalization.

3. The efficient palm vein representation attack detection method based on lightweight neural network according to claim 2 is characterized in that: The Xception inlet stream performs preliminary feature extraction on the image to be identified to obtain initial features, and inputs the initial features into the Block module of the star network. Each Block module further mines high-order nonlinear features in the image through convolution and star operations, including: The basic features are further extracted through two residual blocks, each of which uses BSConvU convolution units and maximum pooling operations to obtain high-order nonlinear features.

4. The efficient palm vein representation attack detection method based on lightweight neural network according to claim 1 is characterized in that: The classifying and identifying high-order nonlinear features to obtain classification result data includes: The feature map is reduced to a fixed size through adaptive average pooling, and then converted into a one-dimensional vector through flattening operation; The fully connected layer outputs a feature vector of two categories of results based on the one-dimensional vector, and calculates the feature vector through the softmax function to obtain the final predicted probability value, that is, the classification result data.

5. An efficient palm vein attack detection system based on lightweight neural network, characterized by: The system includes: an input preprocessing module, a feature extraction module and a classification module; The input preprocessing module is used to obtain the image to be recognized; The feature extraction module includes an inlet flow structure of the Xception model and a Block module of the StarNet model, and is used to extract features of the image to be identified to obtain feature data; The classification module is used to classify and identify high-order nonlinear features to obtain classification result data.

6. The efficient palm vein representation attack detection system based on lightweight neural network according to claim 5, characterized in that: The feature extraction module is used to input the image to be identified into the Xception input stream and perform preliminary feature extraction on it to obtain initial features; The initial features are then input into the Block module of the star network. Each Block module further mines the high-order nonlinear features in the image through convolution and star operations.

7. The efficient palm vein representation attack detection system based on lightweight neural network according to claim 5, characterized in that: The classification module is used to reduce the feature map to a fixed size through adaptive average pooling, and then convert it into a one-dimensional vector through a flattening operation; the fully connected layer outputs a feature vector of two categories of results based on the one-dimensional vector, and calculates the feature vector through the softmax function to obtain the final prediction probability value, that is, the classification result data.

8. The efficient palm vein representation attack detection system based on lightweight neural network according to claim 7, characterized in that: The system uses the cross entropy loss function for binary classification tasks.

Images