Non-contact palm vein anti-attack method based on lightweight network

Abstract

The invention provides a non-contact palm vein anti-attack method based on a lightweight network, and the method comprises the steps: obtaining a palm vein data set, and dividing the palm vein data set into a training data set and a test data set; constructing a lightweight palm vein anti-attack network; the lightweight palm vein anti-attack network is trained; and identifying the palm vein image by using the trained lightweight palm vein anti-attack network. According to the method, the recognition accuracy is 99.6% or above, the average detection time is 8 ms, the model size is 353 kb, the method is easy to deploy under low-computing-power embedded equipment, and real-time, rapid and efficient non-contact palm vein living body detection is achieved; according to the invention, the in-vivo detection of the input palm vein image is realized through the deep learning network, so as to prevent the palm vein image from being attacked and cheated by artificial prostheses.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a non-contact palm vein anti-attack method based on a lightweight network. Background Art

[0002] With the rapid development of science and technology, more and more devices are applied to personal identity authentication, and various applications have higher and higher requirements for the accuracy and security of identity authentication systems. Traditional methods such as password input and card swiping cannot meet the requirements of high security and are gradually replaced by biometric recognition technology. Existing biometric recognition technologies include face, iris, fingerprint, palm print, palm vein, etc. Non-contact palm vein recognition technology uses the different absorption characteristics of near-infrared light by the human palm epidermis and the internal venous blood of the palm. It uses a near-infrared camera to collect each person's unique palm vein pattern image, and combines deep learning and image processing algorithms to identify and verify the identity of the person. It has the advantages of high efficiency, speed, convenience and hygiene. At the same time, because the veins are distributed under the surface skin of the palm, they will not be easily affected by external conditions, nor will they be easily stolen and forged like fingerprints, palm prints, faces, etc. Therefore, compared with other biometric recognition technologies, it has higher security and convenience and can be more widely used in various places.

[0003] However, like non-contact biometric identification methods such as face, palm print, and iris, palm vein recognition also faces the risk of artificial prosthesis attacks and deception. Attackers can forge prostheses in a variety of different ways, such as printing palm grayscale images, drawing palm vein texture images, making palm molds, etc., to deceive and attack the palm vein recognition system and endanger system security. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a non-contact palm vein anti-attack method based on a lightweight network. The present invention realizes liveness detection of input palm vein images through a deep learning network to prevent attacks and deceptions by artificial prostheses.

[0005] The technical solution of the present invention is: a non-contact palm vein anti-attack method based on a lightweight network, comprising the following steps:

[0006] S1), obtaining a palm vein dataset and dividing it into a training dataset and a test dataset;

[0007] S2) Build a lightweight palm vein anti-attack network;

[0008] S3), training a lightweight palm vein anti-attack network;

[0009] S4) Use the trained lightweight palm vein anti-attack network to identify the palm vein image.

[0010] Preferably, in step S1), the palm vein dataset includes a prosthetic dataset and a real palm dataset, and the prosthetic dataset and the real palm dataset are mixed in a ratio of 1:1, and the prosthetic dataset and the real palm dataset are randomly divided into a training dataset and a test dataset in a ratio of 4:1.

[0011] Preferably, in step S1), the prosthesis dataset includes a paper prosthesis dataset and a mold prosthesis dataset.

[0012] Preferably, in step S2), the palm vein anti-attack network includes multiple feature extraction modules and a fully connected layer. Multiple feature maps are extracted by the multiple feature extraction modules, and then the multiple feature maps are input into the fully connected layer. The output of the fully connected layer is an unnormalized 1×2 feature vector, which represents the probability that the input image is a real palm and a non-real palm, respectively. Finally, the feature vector output by the fully connected layer is calculated by the softmax function to obtain the final predicted probability value.

[0013] Preferably, in step S2), a threshold value is set to determine that when the probability of a real palm is greater than the threshold value, it indicates that the target is a real palm, otherwise it is not a real palm.

[0014] Preferably, in step S2), each of the feature extraction modules includes two branches and a feature fusion module, and the multiple feature maps extracted by the two branches are fused by the feature fusion module.

[0015] Preferably, in step S2), the two branches respectively include corresponding residual blocks and depth-wise separable convolution DSC modules, one of which also includes a spatial attention enhancement module SAA, and the other branch also includes a channel attention enhancement module CAA.

[0016] Preferably, in step S2), each branch is first downsampled through the corresponding residual block to reduce the resolution of the feature map while extracting shallow semantic features; then it passes through the spatial attention enhancement module SAA and the channel attention enhancement module CAA respectively, and the sensitivity of the network to a certain part of the features is improved through a variety of different attention enhancements; finally, the feature map after attention enhancement is passed through the corresponding deep separable convolution module DSC for deep feature extraction, which extracts high-level semantic information while accelerating network calculation, and finally, the multiple feature maps extracted by the two branches are fused and output through the feature fusion module.

[0017] Preferably, in step S2), the residual block includes a first convolutional layer, a second convolutional layer, a third convolutional layer and a first BN layer and a second BN layer; wherein the first convolutional layer, the first BN layer and the second convolutional layer serve as a first branch; the third convolutional layer and the second BN layer serve as a second branch; the input feature map is divided into two branches, the first of which first passes through a first convolutional layer with a convolution kernel size of 3×3 and a step size of 2, then passes through a first BN layer, and then outputs with a Relu function as an activation function, and then passes through a second convolutional layer with a step size of 1 and a convolution kernel size of 3×3 for output; the second branch consists of a 1×1 third convolutional layer with a step size of 2 and a second BN layer, both branches downsample the feature maps, and finally the feature maps with a length and width of half of the original obtained by the two branches are fused, and finally activated and output by the Relu function.

[0018] Preferably, in step S2), the spatial attention enhancement module SAA performs two-line processing on the feature map, namely: one performs an average pooling operation on the channel dimension to generate a spatial average pooling feature map; the other performs maximum pooling on the channel dimension to generate a spatial maximum pooling feature map; then the two feature maps are fused, and a convolution layer is used to further extract the pooled spatial information, and finally the Sigmoid function is used to map the output to between 0 and 1 to obtain the weight information of each pixel in the space.

[0019] Preferably, in step S2), the channel attention enhancement module CAA processes the feature map through two branches respectively, namely: one performs average pooling in the spatial dimension, pools the size of the feature map into 1×1, and generates a multi-dimensional average pooling feature vector;

[0020] The other branch performs maximum pooling in the spatial dimension, pools the size of the feature map to 1×1, and generates a multi-dimensional maximum pooling feature vector; then the two feature vectors are point-convolved twice, activated by the Sigmoid function, and finally the two feature vectors are fused and passed through the second Sigmoid function to map the output to between 0 and 1, obtaining the weight information of each channel in the channel dimension.

[0021] Preferably, in step S2), the depthwise separable convolution DSC module first performs convolution on each input feature channel through depthwise convolution, and then performs point-by-point convolution on the output of the depthwise convolution through a 1×1 convolution kernel.

[0022] Preferably, in step S3), the lightweight palm vein anti-attack network is trained using the training data set, and the loss function in the training process adopts the cross entropy loss function, and the optimizer is the Adam optimizer; the training epoch is 120, and the initial learning rate is 0.001.

[0023] The beneficial effects of the present invention are:

[0024] 1. The recognition accuracy of the present invention is above 99.6%, and the average detection time is 8ms. The model size is 353kb, which is easy to deploy on low-computing embedded devices to achieve real-time, fast and efficient non-contact palm vein liveness detection;

[0025] 2. The present invention realizes liveness detection of input palm vein images through a deep learning network to prevent attacks and deceptions by artificial prostheses. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a framework diagram of the lightweight palm vein anti-attack network of the present invention;

[0027] Figure 2 A result schematic diagram of the residual block of the present invention;

[0028] Figure 3 It is a structural framework diagram of the spatial attention enhancement module SAA of the present invention;

[0029] Figure 4 It is a structural framework diagram of the channel attention enhancement module CAA of the present invention;

[0030] Figure 5 This is a structural framework diagram of the depth-separable convolutional DSC of the present invention. DETAILED DESCRIPTION

[0031] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0032] like Figure 1 As shown, this embodiment provides a non-contact palm vein attack prevention method based on a lightweight network, including the following steps:

[0033] S1), obtaining a palm vein dataset and dividing it into a training dataset and a test dataset;

[0034] The palm vein dataset described in this embodiment includes a prosthetic dataset and a real palm dataset, with a total of more than 4,800 images; and the prosthetic dataset and the real palm dataset are mixed in a ratio of 1:1, and the prosthetic dataset and the real palm dataset are randomly divided into a training dataset and a test dataset in a ratio of 4:1. Among them, the prosthetic dataset includes a paper prosthetic dataset and a mold prosthetic dataset, wherein it is clear that the paper prosthesis is a real palm vein grayscale image collected by a real palm under a collection device, which is printed and then cut out to simulate an attacker attacking through the palm vein texture of the original user. The mold prosthesis is a palm mold of various materials, or directly photographed under the collection camera, or a certain texture is drawn on it, or various gloves are worn to simulate an attacker attacking through a bionic mold of other materials. The palm vein dataset is shown in Table 1.

[0035] Table 1 Palm vein dataset

[0036]

[0037]

[0038] S2) Build a lightweight palm vein anti-attack network;

[0039] like Figure 1 As shown, in this embodiment, the palm vein attack prevention network includes multiple feature extraction modules and a fully connected layer. Multiple feature maps are extracted by multiple feature extraction modules, and then the multiple feature maps are input into the fully connected layer. The output of the fully connected layer is an unnormalized 1×2 feature vector, which represents the probability of the input image being a real palm and a non-real palm, respectively. Finally, the feature vector output by the fully connected layer is calculated by the softmax function to obtain the final predicted probability value. In this embodiment, a threshold value is set to determine that when the probability of a real palm is greater than the threshold, it indicates that the target is a real palm, otherwise it is not a real palm.

[0040] As a preferred embodiment of the present invention, the palm vein anti-attack network includes four feature extraction modules, each of which includes two branches and a feature fusion module, and the multiple feature maps extracted from the two branches are fused through the feature fusion module.

[0041] The two branches respectively include corresponding residual blocks and depth-separable convolution DSC modules, one of which also includes a spatial attention enhancement module SAA, and the other branch also includes a channel attention enhancement module CAA.

[0042] As a preferred embodiment of this embodiment, each branch is first downsampled through the corresponding residual block to reduce the resolution of the feature map while extracting shallow semantic features; then it passes through the spatial attention enhancement module SAA and the channel attention enhancement module CAA respectively, and the network's sensitivity to a certain part of the features is improved through a variety of different attention enhancements; finally, the feature map after attention enhancement is passed through the corresponding deep separable convolution module DSC for deep feature extraction, extracting high-level semantic information while accelerating network calculation, and finally, the multiple feature maps extracted by the two branches are fused and output through the feature fusion module.

[0043] As preferred in this embodiment, Figure 2 As shown, the residual block includes a first convolutional layer, a second convolutional layer, a third convolutional layer and a first BN layer and a second BN layer; wherein the first convolutional layer, the first BN layer and the second convolutional layer are used as the first branch; the third convolutional layer and the second BN layer are used as the second branch; the input feature map is divided into two branches, the first one first passes through the first convolutional layer with a convolution kernel size of 3×3 and a step size of 2, then passes through the first BN layer, and then outputs with the Relu function as the activation function, and then passes through a second convolutional layer with a step size of 1 and a convolution kernel size of 3×3 for output; the second branch is composed of a 1×1 third convolutional layer with a step size of 2 and a second BN layer. Both branches downsample the feature maps, and finally the feature maps with a length and width of half of the original obtained by the two branches are fused, and finally activated and output by the Relu function.

[0044] As preferred in this embodiment, Figure 3 As shown, the spatial attention enhancement module SAA performs two-line processing on the feature map, namely: one is to perform average pooling operation on the channel dimension to generate a spatial average pooling feature map; the other is to perform maximum pooling on the channel dimension to generate a spatial maximum pooling feature map; then the two feature maps are fused, and a convolution layer is used to further extract the pooled spatial information, and finally the Sigmoid function is used to map the output to between 0 and 1 to obtain the weight information of each pixel in the space.

[0045] As preferred in this embodiment, Figure 4As shown, the channel attention enhancement module CAA processes the feature map through two branches respectively, namely: one branch performs average pooling in the spatial dimension, pools the size of the feature map to 1×1, and generates a multi-dimensional average pooling feature vector; the other branch performs maximum pooling in the spatial dimension, pools the size of the feature map to 1×1, and generates a multi-dimensional maximum pooling feature vector; then the two feature vectors are point convolved twice, activated by the Sigmoid function, and finally the two feature vectors are fused, and after the second Sigmoid function, the output is mapped to between 0 and 1 to obtain the weight information of each channel in the channel dimension.

[0046] As preferred in this embodiment, Figure 5 As shown, when the depthwise separable convolution DSC module operates, each input feature channel is first convolved by depthwise convolution, and then the output of the depthwise convolution is convolved point by point by a 1×1 convolution kernel.

[0047] S3), training a lightweight palm vein anti-attack network;

[0048] In this embodiment, a lightweight palm vein anti-attack network is trained using a training data set, and the loss function in the training process adopts a cross entropy loss function, and the optimizer is an Adam optimizer; the training epoch is 120, and the initial learning rate is 0.001.

[0049] S4) Use the trained lightweight palm vein anti-attack network to identify the palm vein image. This embodiment is verified on the training data set and the test data set, respectively, and achieves the best accuracy of 99.7% and 99.6%, respectively. At the same time, the average detection time is 8ms, and the model size is 353kb. It is easy to deploy on low-computing embedded devices, realizing real-time, fast and efficient non-contact palm vein liveness detection.

[0050] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.

Claims

1. A non-contact palm vein anti-attack method based on a lightweight network, characterized in that: The steps include: S1), obtaining a palm vein dataset and dividing it into a training dataset and a test dataset; S2) Build a lightweight palm vein anti-attack network; S3), training a lightweight palm vein anti-attack network; S4) Use the trained lightweight palm vein anti-attack network to identify the palm vein image.

2. According to claim 1, a non-contact palm vein attack prevention method based on a lightweight network is characterized in that: In step S1), the palm vein dataset includes a prosthetic dataset and a real palm dataset, and the prosthetic dataset and the real palm dataset are mixed in a ratio of 1:1, and the prosthetic dataset and the real palm dataset are randomly divided into a training dataset and a test dataset in a ratio of 4:

1.

3. According to claim 1, a non-contact palm vein attack prevention method based on a lightweight network is characterized in that: In step S2), the palm vein attack protection network includes multiple feature extraction modules and a fully connected layer. Multiple feature maps are extracted by the multiple feature extraction modules, and then the multiple feature maps are input into the fully connected layer. The output of the fully connected layer is an unnormalized 1×2 feature vector, which represents the probability that the input image is a real palm and a non-real palm, respectively. Finally, the feature vector output by the fully connected layer is calculated by the softmax function to obtain the final predicted probability value; by setting a threshold threshold, it is judged that when the probability of the real palm is greater than the threshold, it indicates that the target is a real palm, otherwise it is not a real palm.

4. According to claim 3, a non-contact palm vein attack prevention method based on a lightweight network is characterized in that: In step S2), each of the feature extraction modules includes two feature extraction branches and a feature fusion module, and the feature fusion module is used to fuse the multiple feature maps extracted by the two feature extraction branches.

5. According to claim 4, a non-contact palm vein attack prevention method based on a lightweight network is characterized in that: In step S2), the two feature extraction branches respectively include corresponding residual blocks and deep separable convolution DSC modules, one of which also includes a spatial attention enhancement module SAA, and the other feature extraction branch also includes a channel attention enhancement module CAA; each feature extraction branch first downsamples through the corresponding residual block to reduce the resolution of the feature map while extracting shallow semantic features; then passes through the spatial attention enhancement module SAA and the channel attention enhancement module CAA respectively, and improves the network's sensitivity to a certain part of the features through a variety of different attention enhancements; finally, the feature map after attention enhancement is passed through the corresponding deep separable convolution module DSC for deep feature extraction, extracting high-level semantic information while accelerating network calculation, and finally, the multiple feature maps extracted by the two branches are fused and output through the feature fusion module.

6. The non-contact palm vein attack prevention method based on a lightweight network according to claim 5 is characterized in that: In step S2), the residual block includes a first convolutional layer, a second convolutional layer, a third convolutional layer and a first BN layer and a second BN layer; wherein the first convolutional layer, the first BN layer and the second convolutional layer serve as a first branch; the third convolutional layer and the second BN layer serve as a second branch; the input feature map is divided into two branches, the first of which first passes through a first convolutional layer with a convolution kernel size of 3×3 and a step size of 2, then passes through a first BN layer, and then outputs with a Relu function as an activation function, and then passes through a second convolutional layer with a step size of 1 and a convolution kernel size of 3×3 for output; the second branch is composed of a 1×1 third convolutional layer with a step size of 2 and a second BN layer, both branches downsample the feature maps, and finally the feature maps with a length and width of half of the original obtained by the two branches are fused, and finally activated and output by the Relu function.

7. The non-contact palm vein attack prevention method based on a lightweight network according to claim 5, characterized in that: In step S2), the spatial attention enhancement module SAA performs two-line processing on the feature map, namely: one performs average pooling operation on the channel dimension to generate a spatial average pooling feature map; the other performs maximum pooling on the channel dimension to generate a spatial maximum pooling feature map; then the two feature maps are fused, and a convolution layer is used to further extract the pooled spatial information, and finally the Sigmoid function is used to map the output to between 0 and 1 to obtain the weight information of each pixel in the space.

8. The non-contact palm vein attack prevention method based on a lightweight network according to claim 5, characterized in that: In step S2), the channel attention enhancement module CAA processes the feature map through two branches respectively, namely: one branch performs average pooling in the spatial dimension, pools the size of the feature map to 1×1, and generates a multi-dimensional average pooling feature vector; the other branch performs maximum pooling in the spatial dimension, pools the size of the feature map to 1×1, and generates a multi-dimensional maximum pooling feature vector; then the two feature vectors are point convolved twice, activated by the Sigmoid function, and finally the two feature vectors are fused, and after the second Sigmoid function, the output is mapped to between 0 and 1 to obtain the weight information of each channel in the channel dimension.

9. The non-contact palm vein attack prevention method based on lightweight network according to claim 5, characterized in that: In step S2), the depthwise separable convolution DSC module first performs convolution on each input feature channel through depthwise convolution, and then performs pointwise convolution on the output of the depthwise convolution through a 1×1 convolution kernel.

10. The non-contact palm vein attack prevention method based on lightweight network according to claim 1, characterized in that: In step S3), the lightweight palm vein anti-attack network is trained using the training data set, and the loss function in the training process adopts the cross entropy loss function, and the optimizer is the Adam optimizer; the training epoch is 120, and the initial learning rate is 0.001.

Images