A Palm Vein Recognition Method Based on Semantic Communication Federated Learning

Abstract

The present invention discloses a palm vein recognition method based on semantic communication-based federated learning, which realizes high-security and high-reliability palm vein biometric recognition by means of federated learning (FL). By applying semantic communication to the joint algorithm of federated learning based on fedavg, the communication load is optimized, ensuring that even in a network environment with limited bandwidth, each end device can efficiently participate in the learning and optimization process of the global model. In addition, the present invention adopts advanced image processing technology, proposes an efficient deep learning algorithm based on the fusion of texture and convolutional neural network (CNN) methods, and combines federated learning technology to effectively solve the balance problem between the effectiveness and security of palm vein recognition. The simulation results show that compared with other methods, the proposed method can greatly improve the palm vein recognition performance of each client user on the premise of ensuring the security of user data, and at the same time has a good convergence effect.

Description

Technical Field

[0001] The present invention relates to the fields of wireless communication and biometric recognition, and specifically, to a palm vein biometric recognition technology that achieves high security and high reliability by means of Federated Learning (FL). A joint algorithm that applies semantic communication to Federated Learning based on FedAvg is proposed, which optimizes the communication load and ensures that even in a network environment with limited bandwidth, each end device can efficiently participate in the learning and optimization process of the global model. In addition, the present invention adopts advanced image processing technology, proposes an efficient deep learning algorithm based on the fusion of texture and Convolutional Neural Network (CNN) methods, and combines Federated Learning technology to effectively solve the balance problem between the effectiveness and security of palm vein recognition. The simulation results show that the proposed method greatly improves the palm vein recognition performance of each client user on the premise of ensuring the security of user data. Background Art

[0002] In recent years, as a new research direction in the field of biometric recognition, palm vein recognition has attracted people's attention due to its high security, high detection activity, good user acceptability, and convenience of use. Compared with face, iris, fingerprint, palmprint, signature, voice, etc., palm veins have unique advantages of universality, uniqueness, collectability, and stability. The palm vein blood vessels are located under the skin layer, and the red pigment in the veins has the characteristic of absorbing infrared light. The acquisition of vein images requires near-infrared irradiation, which is not convenient for forgery or surgical modification. In addition, the acquisition process is very user-friendly because using the palm is relatively natural and easy to be accepted by users. These advantages have made vein recognition technology one of the research focuses of researchers at home and abroad in recent years.

[0003] Current vein recognition algorithms primarily include structural feature-based methods, subspace-based methods, texture feature-based methods, and deep learning-based methods. Structural feature-based methods primarily utilize the geometric features of palm veins, such as point features or line features. Subspace-based methods primarily use projection transformations to project samples into a low-dimensional subspace and utilize classification feature methods based on low-dimensional spatial information, such as principal component analysis, local support projection, and linear discriminant analysis. Texture feature-based methods utilize the rich texture features of palm veins to extract features such as direction, amplitude, phase, and grayscale level for encoding. Deep learning-based methods primarily utilize existing deep learning networks for vein recognition. In "An Automated Biometric Identification System Using CNN-Based Palm Vein Recognition," Jhong et al. proposed a novel, high-performance, contactless palm vein recognition system that uses high-performance adaptive background filtering to acquire palm vein images of the region of interest. They employed a modified convolutional neural network to determine the optimal recognition model through training and testing. In addition, Felix et al. proposed a palm vein recognition algorithm that combines traditional texture features and neural networks in "Palm vein recognition through fusion of texture-based and CNN-based methods", which further improved the recognition performance.

[0004] However, it is equally important to consider the privacy protection of user data in advanced technologies such as palm vein recognition. In practical applications, it is necessary to ensure that the biometric data provided by users will not be accessed or abused by unauthorized third parties. Federated learning is an emerging technology that can make the training of machine learning models more efficient while improving data privacy protection. Federated learning allows multiple devices or servers to participate in the training process of a shared model through distributed computing without the need to centralize their data on a central server. Under this framework, the user's palm vein data can remain on the local device and does not need to be sent directly to the cloud server or other devices. Instead, only information about model updates (such as gradients or parameter updates) will be shared and aggregated to improve the model. This operation reduces the risk of sensitive data exposure while still allowing the model to learn from data on various devices and continuously optimize.

[0005] On this basis, considering that traditional federated learning occupies a large amount of communication resources during training and has low training efficiency, the present invention proposes semantic communication-based federated learning, which applies a semantic network to the communication process. Based on deep learning, it mainly focuses on the semantic screening of the transmitted gradient information, that is, the network applies decision screening to weight updates with significant semantics. The server decodes the transmitted information through the semantic network and updates the corresponding shared network. The results show that the proposed algorithm has good convergence and can greatly improve the performance of palm vein recognition for each client user while ensuring the security of user data. Summary of the Invention

[0006] The present invention aims at the problem of palm vein recognition for protecting user data privacy in actual application scenarios, and proposes an efficient deep learning algorithm based on the fusion of texture and CNN methods. By using a joint algorithm that applies semantic communication to federated learning based on fedavg, it maximizes the data security and efficiency in palm vein recognition training and application. The proposed method can greatly improve the performance of palm vein recognition for each client user while ensuring the security of user data.

[0007] The key technology proposed by the present invention to solve its technical problems is: a palm vein recognition method based on semantic communication-based federated learning, characterized in that the method comprises the following steps:

[0008] Step 1: Preprocess the palm vein data;

[0009] Step 1-1: Extract the ROI region;

[0010] Step 1-2: Uniformly set the image resolution;

[0011] Step 1-3: Perform image enhancement using the histogram equalization algorithm;

[0012] Step 2: Obtain the local palm vein data and perform primary feature extraction;

[0013] Step 2-1: Perform texture-based feature extraction on the local client training data;

[0014] Step 2-2: Perform score-level fusion on the features of each sub-region to obtain the feature fusion score;

[0015] Step 3: Obtain the local palm vein data, and the local client performs initial deep learning training;

[0016] Step 3-1: Perform deep learning training based on CNN on the local client data;

[0017] Step 3-2: Form a local initial model and send the local initial model to the server;

[0018] Step 4: Conduct federated learning on multiple client data;

[0019] Step 4-1: The server receives the local initial model, distributes secret keys to each client, and distributes the initial model;

[0020] Step 4-2: Iteratively train the clients, update the model parameters {M1, M2,..., M n}, introduce a pre-trained semantic network to perform semantic extraction on the model parameter update information, adopt a channel coding method to simulate the actual federated communication and encryption process, and introduce the semantic similarity ξ;

[0021] Step 4-3: Iteratively train and update the federated learning until the training of the shared palm vein recognition network is completed;

[0022] Step 5: Perform decision-level fusion, make a joint decision, and return the recognition result to the client user to complete user recognition.

[0023] The present invention has the following advantages:

[0024] 1. The hybrid deep learning algorithm adopted by this technology combines texture feature analysis and convolutional neural network (CNN). Among them, the texture analysis method is very sensitive to the details and patterns in the palm vein image, while CNN can learn complex feature representations through its deep structure. By combining the two, the present invention can extract richer and more robust features from the palm vein image, while improving the generalization ability of the overall model, thereby providing more accurate biometric recognition results in the application stage. The simulation results show that without using the federated learning method and only processing data locally, the performance of palm vein recognition is significantly improved under the application of this hybrid recognition method, and the recognition accuracy is increased by nearly 30% compared with single texture feature extraction.

[0025] 2. The present invention applies federated learning to the palm vein recognition task. The core advantage lies in the privacy protection of user data, because the data of multiple clients do not need to leave the local device for model training. With the help of federated learning, the client device calculates the model update locally and uploads the training information after multiple rounds of communication, effectively reducing the potential risk of privacy leakage. It not only protects the security of the client's personal palm vein data, but also provides an additional privacy protection layer even when the model is under malicious attack.

[0026] 3. Since traditional federated learning generates a large amount of communication requirements during training, this technology first proposed semantic communication-based federated learning, applying semantic networks to the communication link between the federated learning client and the server, and fully optimizing the communication load. By extracting key semantic information to replace traditional gradient information, the overall model training reduces the amount of data required to be transmitted in each learning iteration. This innovation ensures that even under network conditions with limited bandwidth, end devices (such as smartphones or other distributed devices) can efficiently participate in the training and updating process of the global model. This helps to relieve the network pressure of model training in actual application scenarios, effectively improving the training speed while saving communication resources, and enhancing the scalability and operability of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall block diagram of the method proposed by the present invention.

[0028] Figure 2 It is the convergence graph of the texture and CNN-based recognition algorithm proposed in the present invention under different federated learning communication rounds.

[0029] Figure 3 It shows the ablation experiment of the semantic network in the semantic federated learning palm vein recognition algorithm proposed in the present invention.

[0030] Figure 4 It shows the comparison of the recognition accuracy between the palm vein recognition algorithm without federated learning and the palm vein recognition algorithm with semantic federated learning under the same user test set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The following further elaborates on the invention in conjunction with the accompanying drawings of the specification.

[0032] The technical method of the present invention includes the following steps.

[0033] Step 1, preprocess the palm vein dataset, extract the region of interest (ROI), set the image resolution to 128*128, perform image enhancement using the histogram equalization algorithm, and allocate the processed data to 5 client training sets and test sets.

[0034] Step 1.1, extract the ROI region, regard the grayscale image of the palm vein as a three-dimensional surface z = h(u, v), let p be a point on the surface z, and C i represents all the curves passing through point p on the surface z, and the curvatures corresponding to these curves are represented by k i . Among k i , take the maximum curvature k1 and the minimum curvature k2 as the two principal curvatures of point p, K is the Gaussian curvature, and H is the mean curvature. Their relationship with the principal curvatures k1 and k2 can be expressed as:

[0035]

[0036] K = k1 × k2;

[0037] After transformation, we can get:

[0038]

[0039] By calculating the maximum principal curvature for all points on the original ROI image, the maximum principal curvature image is obtained.

[0040] Step 1.2, set the image resolution to 128 * 128;

[0041] Step 1.3, use the histogram equalization algorithm for image enhancement;

[0042] For a grayscale image G of size (m × n), for the pixel gray value size (referring to the total number of image gray levels) corresponding to any position, the occurrence probability of different gray levels can be expressed as:

[0043]

[0044] Therefore, for the grayscale image G, the method of using histogram equalization for image enhancement can be described as:

[0045]

[0046] Here, the function represents the mapping relationship between the gray levels of the image before enhancement and after enhancement. In the invention, the gray level mapping expression of histogram equalization is:

[0047]

[0048] Through the mapping, the histogram equalization result is obtained.

[0049] Step 2, obtain the local palm vein data, perform texture-based feature extraction on the local client training data, and perform score-level fusion on the features of each sub-region to obtain the feature fusion score;

[0050] Step 2.1, perform texture-based feature extraction on the local client training data. Use the histogram of the BSIF coding values of the pixels to characterize the texture attributes of each sub-region of the palm vein. The value of each element (i.e., bit) in the BSIF binary coding string is obtained by binarizing the response of a linear filter, taking zero as the threshold. Each bit is associated with a different filter, and the desired length of the bit string determines the number of filters used. The BSIF method considers an image block X of size l × l pixels and a linear filter W of the same size ito work. The filter response s i The calculation formula is

[0051]

[0052] where the vectors w and x contain the pixel values of W i and X. Therefore, the binarized feature b extracted using BSIF i is obtained in the following way: if s i > 0, then set b i = 1, otherwise set b i = 0. The filter W i is learned using independent component analysis (ICA) by maximizing the statistical independence of s i .

[0053] Step 2.2, perform score-level fusion on the matching scores of each sub-region to obtain a single matching score vector, which is used to classify the sample test image in the classification stage.

[0054] Step 3, obtain local palm vein data, perform deep learning training based on CNN on the local client training data, form a local initial model after training the initial model, and send the local initial model to the server;

[0055] Step 3.1, perform deep learning training based on CNN on the local client data. Use the CNN architecture based on deep learning for inline modeling with the AlexNet structure. Build the CNN architecture, which includes five convolutional layers, and each convolutional layer uses filters to map values from the input image. In the deep learning training process, feature mapping is performed, where the input image is represented by f and the filter is represented by h. The row index and column index of the resulting matrix are represented by m and n respectively. The feature mapping function G[m, n] is expressed as

[0056]

[0057] The activation function of each convolutional layer uses the rectified linear unit (ReLU), which is given by y = max(0, x), converting all positive values to linear and all negative values to zero. To make the data similar to the normal distribution, the batch normalization formula used is expressed as

[0058]

[0059] where B is the mean of a batch of training sets, is the variance of a batch of training sets. The maximum pooling layer follows each convolutional layer and creates a new output matrix by taking the maximum value of a 2×2 region, where each element is the maximum value of a region in the original input, thereby shrinking its input. The architecture is completed by a dropout layer and a fully connected layer, where Softmax is used as the activation function, which normalizes the input value to (0,1). The formula is specifically expressed as

[0060]

[0061] Where each input value x i The exponential of x is divided by the sum of the exponentials of all input values, i.e. j to x k The sum of the exponentials of all input values.

[0062] Step 3.2: Form a local initial model and send it to the server.

[0063] Step 4: Federated learning is performed on multiple client data. The server receives the local initial model, assigns a secret key to each client and distributes the initial model. The client {C1, C2, ..., Cn} reads and uses the initial convolutional neural network to train the local data and updates the model parameters {M1, M2, ..., M n A pre-trained semantic network is introduced to extract semantic information from model parameter updates. Channel coding is used to simulate actual federated communication and encryption processes. Channel decoding is performed on the receiving server. The semantic network then recovers key model parameter information. Local training gradients and losses are then uploaded to the server, and the shared palm vein recognition network is updated. Iterative training and updates are then performed to complete the training of the palm vein deep network.

[0064] Step 4.1: The server receives the local initial model, and the client {C1, C2, ..., Cn} reads and uses the initial convolutional neural network to train the local data and update the model parameters {M1, M2, ..., M n}; Data is distributed among k (k=5) different clients, P k Formed by the distributed training data, n k =|P k |,n is the total number of data The conditions are met

[0065]

[0066] in, The present invention uses a palm vein federated learning method based on FederatedAvg (FedAvg). In each round, randomly selected clients perform a gradient calculation once, and then obtain k gradient information and perform subsequent semantic extraction. Compared with traditional distributed machine learning methods that only calculate gradients at client nodes, the FedAvg method hopes that client nodes can perform more operations to obtain a better descent direction than the gradient. Since this descent direction is better than the gradient, the convergence speed is higher and the number of communications is correspondingly reduced. The gradient calculation process of this method can be expressed by the formula

[0067]

[0068] Through iteration, the calculation for each client is increased while the model parameters are updated. The formula is expressed as

[0069]

[0070] where η represents the learning rate, initially set to 0.01. Multiple hyperparameters are set in the training process. Among them, C represents the fraction of the number of clients. C = 0.0 represents the first client, which is equivalent to the proportion of the extraction number among K clients. The hyperparameter B represents the local training batch size of each client. The hyperparameter E represents the number of local iteration cycles of each client in each round.

[0071] Step 4.2: Introduce a pre-trained semantic network to perform semantic extraction and encryption on the model parameter update information. Adopt a channel coding method to simulate the actual federated communication and encryption process, perform channel decoding at the server receiving end, and then restore important model parameter information through the semantic network, so as to upload the local training gradient and loss, and the server updates the shared palm vein recognition network.

[0072] The semantic extraction of the model parameter update information by the semantic network and the channel coding process can be expressed as:

[0073] x' = B α (C β (x));

[0074] where B α (·) represents the semantic encoder based on the pre-trained semantic network model and the parameter set 'α', and C β (·) represents the channel encoder of the parameter set 'β'. The process of semantic decoding information is the inverse process of the encoding process, expressed as:

[0075] x' = B α (C β (x));

[0076] In addition, the present invention introduces a semantic similarity estimation method to characterize the importance degree of the extracted semantic information and the original gradient information in the model update task, and the formula is expressed as:

[0077]

[0078] Among them, B(·) represents a pre-trained semantic network model based on a transformer. The semantic similarity ξ is a continuous value between 0 and 1. When the comprehensive importance degree of the extracted semantic representation for model update is higher, the semantic similarity is higher, which means the higher the efficiency of the semantic network and the better the training performance of the entire federated learning.

[0079] In the simulation of channel encoding and decoding, for the semantic set x of the training information to be transmitted, the simulation transmission process can be expressed as:

[0080] y = hx'+ω,

[0081] Among them, h represents the fading coefficient, which is a random variable, and its probability density function (PDF) can be represented by a Rayleigh distribution. ω represents additive white Gaussian noise (AWGN). The probability density function of the fading coefficient can be represented as a Rayleigh distribution, and its expression is:

[0082]

[0083] Among them, σ is the scale parameter of the fading channel, which determines the severity of the fading. The fading coefficient h takes non-negative real values. In this study, the overall simulation of the basic communication environment is controlled by controlling the fading gain and fading loss in the fading channel.

[0084] Step 4.3, iterate the federated learning training and update until the shared palm vein recognition network training is completed.

[0085] Step 5, decision-level fusion, jointly make a decision on the recognition scores of texture-based features and the recognition scores of the deep learning network based on CNN, and combine the recognition losses of the first-level decision (texture-based) and the second-level decision (CNN-based) with weighted processing to obtain the final decision, and return to the local client to complete user recognition.

[0086] In the decision-level fusion process, what we mainly focus on is how to effectively integrate information from different sources to make a final decision. In the present invention, the recognition scores of texture-based features and the scores of the deep learning network based on CNN are particularly considered. Since the decision-making process is only affected by two networks, we choose linear weighted decision as the basic method. This method calculates a weighted sum by assigning different weights (W fex ) and (W cnn )) to the decision results from different sources, and uses this as the final recognition decision. The specific process can be expressed by the formula as:

[0087] S fin =W fex ·S f +W cnn ·S c

[0088] Among them, W fex and W cnn respectively represent the weights of the texture-based feature recognition score and the decision result of the CNN-based deep learning network in the final recognition score. They are affected by the model training parameters and are comprehensively judged and dynamically assigned according to the training losses of the two during the iteration process. S f and S c represent the primary decision and secondary decision scores respectively. After linear addition, the final decision score S fin is obtained. According to the predefined decision threshold, the user recognition result is obtained.

[0089] The performance effect of the present invention can be further illustrated by the following simulation:

[0090] 1. Simulation hardware conditions

[0091] The simulation experiment of the present invention is carried out on a simulation platform of Python 3.8 and Pytorch 11.7. The computer CPU model is E5-2680v4, and the number is 5. The GPU model is NVIDIA GeForce RTX ****************** GB).

[0092] 2. Simulation conditions

[0093] The convergence error tolerance is set to 0.05, the image resolution is 128*128, the initial federated learning learning rate η is set to 0.01, the number of clients is set to 10, and the training volume of each client's data is 50 different palm vein recognition images, which are incrementally processed through linear transformation. The number of local iteration cycles of each round of clients is set to 6 times.

[0094] 3. Simulation content

[0095] Appendix Figure 2 shows the training convergence and recognition accuracy graphs of the palm vein recognition algorithm based on semantic communication-based federated learning proposed in the present invention as the number of communication rounds gradually increases, where the number of communication rounds is set to 5 times. It is observed that as the number of communication rounds increases, the proposed recognition algorithm converges to a stable point under a limited number of communication rounds, which verifies the effectiveness of the algorithm we proposed. In addition, the overall model recognition accuracy gradually improves and reaches a recognition accuracy of 97.4% after training, which effectively illustrates the high feasibility of the present invention.

[0096] Appendix Figure 3Shows the comparison of the palm vein recognition algorithm based on semantic communication-based federated learning proposed in the present invention with other basic recognition algorithms, mainly including the recognition method based on convolutional neural network and the traditional method based on texture feature recognition. The form of the simulation diagram considered is the Receiver Operating Characteristic (ROC) diagram. It can be observed that the palm vein recognition algorithm based on semantic communication-based federated learning proposed by us performs best on the ROC curve, which represents that the proposed algorithm has obvious advantages in the recognition task compared with individual deep learning or texture feature recognition, effectively demonstrating the high effectiveness of the present invention.

[0097] Appendix Figure 4 Shows the comprehensive index comparison of the palm vein recognition algorithm based on semantic communication-based federated learning proposed in the present invention, the federated learning recognition method without semantic network addition, and the traditional recognition method in terms of training efficiency (the greater the communication resource occupancy, the lower the training speed), security, and recognition effectiveness. It can be observed that the proposed palm vein recognition algorithm based on semantic communication-based federated learning is superior to the traditional method in terms of training efficiency, security, and recognition effectiveness, and is only slightly lower than the federated learning recognition method without semantic network in terms of recognition effectiveness. This is mainly because the introduction of the semantic network will perform semantic encryption on the information while fully protecting user privacy, and will select more important semantic information to a certain extent to reduce the communication resource occupancy, which will inevitably cause a certain loss in recognition performance, and this loss is acceptable.

[0098] Based on the above simulation results and analysis, the palm vein recognition algorithm based on semantic communication-based federated learning proposed in the present invention can enable the entire system to obtain better security and recognition performance, and at the same time effectively reduce the communication resource occupancy. In addition, the proposed algorithm can obtain good convergence performance and does not require a large computational complexity, and can be effectively extended to distributed palm vein recognition systems in other practical application scenarios.

Claims

1. A palm vein recognition method based on semantic communication-style federated learning, characterized in that, The method includes the following steps: Step 1: Preprocess the palm vein data; Step 1-1: Extract the ROI region; Step 1-2: Uniformly set the image resolution; Step 1-3: Use the histogram equalization algorithm for image enhancement; Step 2: Obtain the local palm vein data and perform primary feature extraction; Step 2-1: Perform texture-based feature extraction on the local client training data; Step 2-2: Perform score-level fusion on the features of each sub-region to obtain the feature fusion score; Step 3: Obtain the local palm vein data, and the local client performs initial deep learning training; Step 3-1: Perform deep learning training based on CNN on the local client data; Step 3-2: Form a local initial model and send the local initial model to the server; Step 4: Perform federated learning on the data of multiple clients; Step 4-1: The server receives the local initial model, distributes keys to each client, and distributes the initial model; Step 4-2: Iteratively train the client, update the model parameters {M1, M2,..., Mn}, introduce a pre-trained semantic network to perform semantic extraction on the model parameter update information, and utilize the channel encoding and decoding model C of the parameter set 'β' β (·) Simulate the actual federated communication and encryption process. The semantic extraction of the model parameter update information by the semantic network and the channel encoding process can be expressed as: x = C β (B α (x)); Among which B α (·) represents a semantic encoder based on a pre-trained semantic network model and a parameter set 'α', C β (·) represents a channel encoder of a parameter set 'β'. The process of semantic decoding information is the inverse process of the encoding process, which is expressed as: x' = B α (C β (x)); In addition, a semantic similarity estimation method is introduced to characterize the importance degree of the extracted semantic information and the original gradient information in the model update task, and the formula expression is: Among them, B(·) represents a pre-trained semantic network model based on a transformer, s represents the original gradient information, represents the encoded semantic information, T represents the transpose of a matrix, the semantic similarity ξ is a continuous value between 0 and 1. When the comprehensively important degree of the extracted semantic representation for model update is higher, the semantic similarity is higher, which means the higher the efficiency of the semantic network and the greater the training performance of the entire federated learning; In the simulation of channel encoding and decoding, for the semantic set x of the training information to be transmitted, the simulation transmission process can be expressed as: y = hx' + ω, where h represents the fading coefficient, which is a random variable, and its probability density function (PDF) can be represented by the Rayleigh distribution, ω represents the additive white Gaussian noise (AWGN), and the probability density function of the fading coefficient can be represented by the Rayleigh distribution, and its expression is: where σ is the scale parameter of the fading channel, which determines the severity of the fading, and the fading coefficient h takes non-negative real values; Step 4-3: Iteratively perform federated learning training and update until the shared palm vein recognition network training is completed; Step 5: Perform decision-level fusion, jointly make a decision, and return the recognition result to the client user to complete user recognition.

2. The palm vein recognition method based on semantic communication-based federated learning according to claim 1, wherein, The said Step 1 includes: preprocessing the palm vein data, and the specific calculation steps are: (2a) Extract the ROI region. Consider the grayscale image of the palm vein as a three-dimensional surface z = h(u, v), where u represents the horizontal coordinate of the image, v represents the vertical coordinate of the image. Let p be a point on the surface z, and represent all the curves passing through point p on the surface z. The curvatures corresponding to these curves are represented by k. Among k, take the maximum curvature k1 and the minimum curvature k2 as the two principal curvatures of point p. K is the Gaussian curvature, and H is the mean curvature. Their relationship with the principal curvatures k1 and k2 can be expressed as: K = k1 × k2; After transformation, we can get: where h u represents the partial derivative of h with respect to u, h v represents the partial derivative of h with respect to v, represents h u squared, represents h v squared, h uu represents the second - order partial derivative of h with respect to u, h uv represents the second - order partial derivative of h with respect to u and v, h vv represents the second - order partial derivative of h with respect to v and v; By calculating the maximum principal curvature for all points on the original ROI image, the maximum principal curvature image is obtained; (2b) Set the image resolution to 128*128; (2c) Use the histogram equalization algorithm for image enhancement; For a grayscale image G of size (m×n), for the pixel gray value size corresponding to any position (L refers to the total number of image gray levels), the occurrence probability of different gray levels can be expressed as: where δ(·) represents the discrete impulse function, which takes 1 when g(x, y) = i. Therefore, for the grayscale image G, the method of using histogram equalization for image enhancement can be described as: Here, the function represents the mapping relationship between the gray level of the pre-enhanced image and that of the post-enhanced image; in the invention, the gray level mapping expression of histogram equalization is: Through the mapping, the histogram equalization result is obtained.

3. A palm vein recognition method based on semantic communication-style federated learning as described in claim 1, characterized in that, Step 2 includes: obtaining local palm vein data and performing primary feature extraction. The specific calculation steps are as follows: (3a)Perform texture-based feature extraction on the local client training data; use the histogram of the BSIF coding values of the pixels to characterize the texture attributes of each sub-region of the palm vein; the value of each element (i.e., bit) in the BSIF binary coding string is obtained by binarizing the response of a linear filter, taking zero as the threshold; each bit is associated with a different filter, and the expected length of the bit string determines the number of filters used; the BSIF method works by considering an image block X of size l×l pixels and a linear filter W of the same size i to work; the filter response s i The calculation formula is Among them, the vectors w and x contain the pixel values of W i and X; thus, the binarized feature b extracted using BSIF i is obtained in the following way: if s i > 0, then set b i = 1, otherwise set b i = 0; the filter W i is learned using independent component analysis (ICA) by maximizing the statistical independence of s i ; (3b) Perform score-level fusion on the matching scores of each sub-region to obtain a single matching score vector, which is used to classify the sample test image in the classification stage.

4. For a palm vein recognition method based on semantic communication-based federated learning as described in claim 1, it is characterized in that Step 3 includes: obtaining local palm vein data and performing initial deep learning training on the local client. The specific calculation steps are as follows: (4a) Perform deep learning training based on CNN on the local client data; perform inline modeling using the CNN architecture based on deep learning and the AlexNet structure; construct the CNN architecture, which includes five convolutional layers, and each convolutional layer uses filters to map values from the input image; in the deep learning training process, feature mapping is performed, where the input image is represented by f, and the filter is represented by h; the row index and column index of the obtained matrix are represented by m and n respectively; the feature mapping function G[m,n] is expressed as The activation function of each convolutional layer uses the rectified linear unit (ReLU), which is given by y = max(0, x), converting all positive values to linear and all negative values to zero; to make the data similar to the normal distribution, the batch normalization formula used is expressed as where B is the mean of a batch of training sets, is the variance of a batch of training sets; The max pooling layer follows each convolutional layer and creates a new output matrix by taking the maximum value of a 2×2 region, where each element is the maximum value of a region in the original input, thus shrinking its input; The architecture is completed by an exit layer and a fully connected layer, where Softmax is used as the activation function, which normalizes the input values to (0,1), and the formula is specifically expressed as wherein the exponent of each input value x i is divided by the sum of the exponents of all input values, i.e., the sum of the exponents of all input values from x j to x k ; the sum of the exponents of all input values (4b) Form the local initial model and send the local initial model to the server.

5. A palm vein recognition method based on semantic communication-style federated learning according to claim 1, characterized in that Step 4 includes: performing federated learning on the data of multiple clients. The server receives the local initial model, distributes keys to each client and distributes the initial model. The clients {C1, C2,..., Cn} read and use the initial convolutional neural network to train the local data, update the model parameters {M1, M2,..., Mn}, introduce a pre-trained semantic network to perform semantic extraction on the model parameter update information, adopt a channel coding method to simulate the actual federated communication and encryption process, perform channel decoding at the server receiving end, and then restore the important model parameter information through the semantic network, and upload the local training gradient and loss accordingly. The server updates the shared palm vein recognition network, and then performs iterative training and update until the training of the palm vein deep network is finally completed. The specific calculation steps are as follows: (5a) The server receives the local initial model, and the clients {C1, C2,..., Cn} read and use the initial convolutional neural network to train the local data and update the model parameters {M1, M2,..., Mn}; the data is distributed among k (k = 5) different clients, P k formed by the distributed training data, n k = |P k |, where n is the total number of data The global loss function f(w) satisfies the condition that Among them, F k (w) represents the local loss function of the k-th client, The present invention uses a palm vein federated learning method based on Federated SGD (FedSGD). In each round, a randomly selected client performs a gradient calculation, and then k gradient information is obtained for subsequent semantic extraction. The gradient calculation process can be expressed by the formula Through iteration, the calculation for each client is increased, which is expressed as Among them, η represents the learning rate, initially set to 0.01; multiple hyperparameters are set during the training process; (5b) Introduce a pre-trained semantic network to perform semantic extraction on the model parameter update information, adopt a channel coding method to simulate the actual federated communication and encryption process, perform channel decoding at the server receiving end, and then restore the important model parameter information through the semantic network, and upload the local training gradient and loss accordingly. The server updates the shared palm vein recognition network; in the present invention, the overall simulation of the basic communication environment is controlled by controlling the fading gain and fading loss in the fading channel; (5c) Iteratively perform federated learning training and update until the training of the shared palm vein recognition network is completed.

6. For a palm vein recognition method based on semantic communication-based federated learning described in claim 1, it is characterized in that, Step 5 includes: decision-level fusion, jointly making decisions on the recognition scores of texture-based features and the recognition scores of the deep learning network based on CNN, performing weighted processing on the recognition losses of the first-level decision (based on texture) and the second-level decision (based on CNN), obtaining the final decision, and returning to the local client to complete user recognition; In the decision-level fusion process, what we mainly focus on is how to effectively integrate information from different sources to make a final decision; in the present invention, the texture-based feature recognition score and the deep learning network score based on CNN are particularly considered. Since the decision-making process is only affected by two networks, we choose the linear weighted decision as the basic method. This method calculates a weighted sum by assigning different weights (W fex ) and (W cnn ) to the decision results from different sources, and uses this weighted sum as the final recognition decision. The specific process can be expressed by the formula as follows: S fin = W fex ·S f + W cnn ·S c , Among them, W fex and W cnn respectively represent the weights of the texture-based feature recognition score and the decision result of the CNN-based deep learning network in the final recognition score. They are affected by the model training parameters and are dynamically assigned according to the training losses of the two during the iteration process. S f and S c represent the first-level decision and second-level decision scores respectively. After linear addition, the final decision score S fin is obtained. According to the predefined decision threshold, the user recognition result is obtained.

Images