Dynamic personalized federal learning method for cross-spectrum palmprint recognition

Abstract

The invention discloses a dynamic personalized federal learning method for cross-spectrum palmprint recognition, and relates to the technical field of mode recognition, image processing and privacy protection. The method comprises the following steps of local training, aggregation and testing. The existing defects are as follows: (1) a federal average mode is not considered to be not suitable for a scene (for example, palm print pictures of each client are collected in different spectrums) when the client has isomerism, so that a trained model is not suitable for all clients; and (2) during each round of communication, a local model and a global model generally adopt a method with a fixed proportion, and the proportion is not considered to dynamically adapt to continuously trained and updated models so as to achieve the optimal performance. According to the method, privacy is protected, meanwhile, each client with heterogeneous data can have a personalized model with good performance, and meanwhile, the personalized model has high generalization ability. Therefore, the method has certain application value and significance.

Description

Technical Field

[0001] The present invention relates to pattern recognition, image processing technology and privacy protection technology. Specifically, each client has palmprint images from different spectrums and uses these images for local training, and then uploads them to a server for dynamic personalized federated aggregation to obtain a personalized model unique to each client. The model can not only adapt well to the local palmprint data set, but also has good generalization ability and can be applicable to cross-spectrum palmprint recognition scenarios. Under the premise of effectively protecting the privacy of client data, the present invention further improves the recognition performance and generalization ability of the model. Background Art

[0002] In the digital age, the accuracy and reliability of identity authentication are constantly increasing, and biometric recognition has gradually become the mainstream identity authentication method. Although face and fingerprint recognition technologies have been widely used, they each have obvious defects: face recognition has the risk of privacy leakage, fingerprints are easily damaged and forged, and are easily disturbed by the environment. In contrast, palmprint recognition technology has unique advantages: first, high recognition accuracy, because its texture features are rich and densely distributed; second, strong stability, the features basically remain unchanged throughout life and have good environmental adaptability; third, high security, with strong anti-counterfeiting performance. Therefore, palmprint recognition has been successfully applied to many fields such as payment verification and access control management.

[0003] Palmprint recognition research mainly includes three technical routes: global methods for statistical feature extraction, local methods for feature description, and automatic feature extraction methods based on deep learning. Traditional global and local methods rely too much on expert experience and manual design, and are difficult to adapt to diverse palmprint images. Although deep learning methods have excellent performance, they require centralized data processing, which easily leads to privacy and security issues. In order to protect data privacy, researchers have developed technical solutions such as differential privacy, homomorphic encryption, and federated learning. Among them, federated learning protects data privacy and builds high-performance global models through local training and server aggregation. However, existing federated learning methods face two major challenges in palmprint recognition: first, palmprint images collected under different spectra have data heterogeneity, which makes it difficult for the global model to adapt to each client; second, although personalized models can be constructed by weighted fusion of local models and global models, the fixed weight ratio cannot dynamically adapt to the model evolution process, which limits the overall performance of the system. Therefore, it is of great significance to develop a new personalized federated learning method that can effectively handle cross-spectral data heterogeneity and adaptively adjust the model fusion weights to promote the development of palmprint recognition technology. Summary of the invention

[0004] The present invention provides a dynamic personalized federated learning method for cross-spectral palmprint recognition, proposes a new loss function combination and a combined aggregation strategy, and further realizes high-accuracy palmprint recognition based on federated learning in cross-spectral scenarios.

[0005] To implement the above method, the specific steps are as follows:

[0006] Step 100: Each client divides the local data set into a training set and a test set, trains a local model using the training set data, and uploads it to the server;

[0007] Step 200, the server accepts the local models of all clients and selects a suitable strategy for federated aggregation. The present invention adopts a combined aggregation method: the first k rounds of communication process adopt the federated averaging method (FedAvg), and finally a global model is obtained and distributed to each client for the next round of training; the later communication process adopts the personalized federated aggregation method (Personalized Federated Learning, PFL), and finally a personalized model corresponding to each client is obtained and distributed to each client for the next round of training.

[0008] Step 300, testing phase, this step has two applications in the whole process: one is to use it in step 200 to select the best local model and the proportion of the global model in the final personalized model; the other is to finally evaluate the performance of the personalized model of each client generated by this method.

[0009] Wherein, each client has palmprint images with different spectra, and the step 100 of dividing these palmprint images into a training set and a test set and performing local training to obtain a local model includes:

[0010] First, each client divides the local palmprint dataset into a training set and a test set. A unified setting is made here, that is, the number of corresponding labels in the training set and test set of each client is the same, so as to avoid the influence of heterogeneity caused by different data distribution and only focus on the data heterogeneity influence caused by different spectra.

[0011] Then, each client uses the local training set for local training. First, a palmprint feature extraction network is randomly initialized (Comprehensive competition Network, CCNet, all clients use this network for random initialization, "Comprehensive competition mechanism in palmprint recognition"), and then one communication round (including 3 cycles) is trained under the constraints of the cross entropy loss, supervised contrast loss, triplet loss, and Arcface center loss proposed in the present invention, so as to obtain the local model corresponding to each client, and upload these local models to the server.

[0012] The server accepts all local models and selects corresponding aggregation strategies to aggregate them to obtain a new model, and finally sends the obtained new model to the corresponding client in step 200, which includes:

[0013] The server accepts all local models. The present invention adopts a combined federation aggregation method, and selects a corresponding aggregation method according to the current communication round: when the current communication round is less than or equal to k (artificially set, and in order to balance the training time cost and the final model performance, k accounts for a large proportion of the total communication rounds), a federal average method (FedAvg, "Communication-efficient learning of deep networks from decentralized data") is used to obtain a new global model, and the global model parameters are distributed to each client for local training in the next round of communication; when the current communication round is greater than k, the dynamic personalized aggregation method proposed in the present invention is adopted, that is, firstly, the test method of step 300 is used to select the best proportion m ([0,1], the present invention sets the proportion of local models in different clients to be the same to reduce the difficulty of the task) of the local model in the personalized global model, specifically, by traversing m on [0,1] with a step size of 0.1, and each time using the test method of step 300 to obtain the average accuracy as the evaluation criterion, select the m used when the average accuracy is the highest as the best m, and then each client obtains the best personalized model and distributes it to the corresponding client.

[0014] The step 300 of the test process for selecting the best m and the performance evaluation process of the final model includes:

[0015] First, the test process for selecting the best m: Specifically, there are S clients in total. Through the method of the present invention, each client will obtain its own unique personalized model. In order to comprehensively balance the personalization and generalization capabilities of the model, a cross-client test method is used. Each personalized model is used to test and match the training set corresponding to the client model and the test set of all clients, and the accuracy (Accuracy, ACC) of each test is recorded. Specifically, the trained personalized feature extractor is used to extract the features of the training set and the test set of all clients respectively and match them. Then the Euclidean distance is calculated, and the label corresponding to the feature vector with the smallest distance is selected as the predicted label, which is compared with the true label, and the accuracy (Accuracy, ACC) is calculated. Finally, S 2 The accuracy results are averaged as the final evaluation result, and the m used when the average accuracy is the highest is selected as the optimal m.

[0016] Secondly, the performance evaluation process of the final model: After all communication rounds are completed, the performance of the final model is evaluated. The same method as above is used to calculate the average accuracy ACC as the recognition performance and the average equal error rate (Equal Error Rate, EER) as the verification performance. The calculation of the average accuracy is the same as above, as the recognition performance. The calculation of the equal error rate is also to use the personalized model to extract the features of the training set and all client test sets and match them, then calculate the Euclidean distance, and then calculate the true positive rate (TPR), false positive rate (FPR) by setting different thresholds. The value when the equal error rate (EER) is 1-TPR=FPR is calculated by calculating this S 2 The average of the equal error rate results is used to evaluate the validation performance.

[0017] The present invention has the following advantages:

[0018] (1) The final personalized model is applicable to both local spectral datasets and other spectral datasets, which means that the present invention is applicable to heterogeneous scenarios under different spectra.

[0019] (2) The present invention is applicable to different tasks: identification or verification, by adjusting the hyperparameter k to control the average accuracy ACC to be optimal or the average error rate EER to be optimal;

[0020] (3) It has strong generalization ability. By simply adjusting the hyperparameter k, the extraction method of the present invention can be applied to other similar biometric recognitions, such as knuckle pattern recognition, finger vein recognition, and palm vein recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart of a dynamic personalized federated learning method for cross-spectral palmprint recognition

[0022] Figure 2 Schematic diagram of palmprint ROI images on different databases

[0023] Figure 3 ROC curves for cross-client testing of the final model under different federated learning methods DETAILED DESCRIPTION

[0024] The method will be described in detail below in conjunction with the accompanying drawings. It should be noted that, in general, the method can be applied to palmprint recognition under different spectra.

[0025] The present application provides a dynamic personalized federated learning method for cross-spectral palmprint recognition, which realizes high-performance cross-spectral palmprint recognition in a federated scenario.

[0026] Figure 1 A flowchart of a dynamic personalized federated learning method for cross-spectral palmprint recognition of the present application is shown.

[0027] This application proposes a dynamic personalized federated learning method for cross-spectral palmprint recognition. The specific steps are as follows:

[0028] Step 100: Each client divides the local data set into a training set and a test set, trains a local model using the training set data, and uploads it to the server;

[0029] Step 200, the server accepts the local models of all clients and selects a suitable strategy for federated aggregation. The present invention adopts a combined aggregation method: the first k rounds of communication process adopt the federated averaging method (FedAvg), and finally a global model is obtained and distributed to each client for the next round of training; the later communication process adopts the personalized federated learning method (PFL), and finally a personalized global model corresponding to each client is obtained and distributed to each client for the next round of training.

[0030] Step 300, testing phase, this step has two applications in the whole process: one is to use it in step 200 to select the best local model and the proportion of the global model in the final personalized model; the other is to finally evaluate the performance of the personalized model of each client generated by this method.

[0031] The step 100 includes the following sub-steps:

[0032] Sub-step 110, each client has a data set, and the ROI images of different data sets are as follows: Figure 2As shown, each client divides the local palmprint dataset into a training set and a test set, using a unified standard that selects the first three images of each category as the training set and the remaining images as the test set.

[0033] Sub-step 120, the present invention proposes a new loss function combination: cross entropy loss, supervised contrast loss, triple loss, Arcface center loss, which is used to constrain local training so that the local model has good classification ability. The cross entropy loss formula is:

[0034]

[0035] Where K represents the total number of images, M represents the total number of categories, and y i,c is the true label of the i-th sample in the c-th class, p i,c is the predicted probability of the i-th sample in the c-th class. The supervised contrast loss formula is:

[0036]

[0037] Where I = {1, ..., 2K} is the sample index set for comparison, P(i) is the positive sample set, A(i) is the index set that does not contain i, and z i is the feature of the anchor sample, z p is the feature of the positive sample, z a is the feature of negative samples, and τ = 0.07 is the temperature coefficient. The weighted sum of these two loss functions is used as the first loss:

[0038] L 1 =w ce L ce +w con L con (3)

[0039] where w ce =0.8, w con = 0.2. The triplet loss is used as the second loss function, and the formula is:

[0040] L 2 =max(0,d(a,p)-d(a,n)+α) (4)

[0041] Where a is the anchor sample, p is the positive sample, and n is the negative sample. is the Euclidean distance (c represents the first eigenvector, d represents the second eigenvector, c l Represents the value of the first eigenvector on the lth dimension, d l represents the value of the second eigenvector on the lth dimension), α is the minimum distance between positive samples and negative samples, and α=1 in the present invention. The Arc face loss formula is:

[0042]

[0043] Where s = 64 is the scaling factor used to adjust the scale of the loss, e is the natural number base, θ is the cosine value of the angle between the input sample feature vector and the center feature vector of the category, k = 0.5 is the angle boundary value used to increase the distance between classes, and y is the label. The center loss formula is:

[0044]

[0045] Where M represents the total number of categories in the data set, x i represents the feature vector of the i-th sample, c j Represents the central feature vector of the jth category. The arc face loss and the center loss are summed as the third loss:

[0046] L 3 =L Arc +L Cen (7)

[0047] The total loss formula is:

[0048] L = L 1 +L 2 +L 3 (8)

[0049] Eventually, each client trains a local model and uploads the local model to the server for subsequent aggregation.

[0050] The step 200:

[0051] The server accepts the local model from each client and selects an aggregation strategy based on the current number of communication rounds: the first k rounds of communication process adopt the federated average method (FedAvg), that is, the global model comes from the weighted sum of each client, and the global model is distributed to each client for the next round of training. The FedAvg formula is:

[0052]

[0053] Where S is the total number of clients, is the local model parameter of the sth client, θ 1is the global model parameter obtained by aggregation. In the later stage, the personalized federated learning method (PFL) is adopted, that is, the global model is first obtained by the federated averaging method (FedAvg), and then the personalized global model is obtained by weighted summation of the global model and the local model of this client, where the weighted ratio used by the local model is m, and the ratio used by the global model is 1-m. By traversing m from 0 to 1, the step size is set to 0.1, and the best m is selected through the test in step 300, and this m is used to obtain the best personalized global model for this round of communication, and finally the obtained personalized global model is sent to the corresponding client for the next round of training. The PFL formula is:

[0054]

[0055] Where m is the proportion of the best local model obtained through testing, θ 1 ′ is the global model parameter obtained by (9), is the local model parameter of the sth client, Personalize model parameters for the sth client.

[0056] The step 300:

[0057] The test process has two applications. One is to select the best m in step 200. Specifically, there are S clients in total. Through the method of the present invention, each client will obtain its own unique personalized model. In order to comprehensively balance the personalization and generalization capabilities of the model, a cross-client test method is used. Each personalized model is used to perform test matching on the training set corresponding to the client model and the data set of all clients, and the accuracy of each test is recorded. Specifically, the trained personalized feature extractor is used to extract the features of the training set and the test set of all clients respectively and match them, and then the Euclidean distance is calculated. in is the feature vector of each image in the training set, is the feature vector of an image in the test set. The label corresponding to the feature vector with the smallest distance in the training set is selected as the predicted label, and compared with the true label to calculate the accuracy. Finally, S 2 The accuracy results are averaged as the final evaluation result, and the m used when the average accuracy is the highest is taken as the best m. The second is suitable for evaluating the performance of the final model after all communications are completed. The same method is used as the first, except that the average accuracy is calculated as the recognition performance and the average equal error rate is calculated as the verification performance. The calculation of the average accuracy is the same as the first, as the recognition performance. The calculation of the equal error rate is also to extract the features of the training set and all client test sets and match them, and then calculate the Euclidean distance The Euclidean distance is normalized to the matching score, and then more than 1000 uniformly distributed thresholds are generated between the minimum matching score and the maximum matching score (a matching score greater than this threshold indicates a correct match, and a matching score less than this threshold indicates an incorrect match), and the correct acceptance rate under each threshold is calculated. TP is the number of the same category correctly identified as the same category, FN is the number of the same category identified as different categories, and the false acceptance rate Among them, FP is the number of heterogeneous objects identified as homogeneous objects, TN is the number of heterogeneous objects identified as heterogeneous objects, and the value when the error rate is 1-TPR=FPR is calculated by 2 The average of the error rates is taken as the validation performance.

[0058] The original palmprint images in the present invention come from three public palmprint databases: PolyU multispectral dataset, IITD dataset and CASIA dataset. Among them, the PolyU multispectral dataset contains four spectral sub-datasets, including red light (Red), green light (Green), blue light (Blue) and near infrared (NIR). A total of 250 people participated in the collection process, and 12 images were collected for each palm under each spectrum. Therefore, each sub-dataset contains 6,000 palmprint images under each spectrum. The IITD dataset contains 2,601 palmprint images collected from 460 different palms, and each palm contains 5 to 7 images. The CASIA dataset consists of 5,502 palmprint images, covering 312 different categories, and each category contains 8 to 17 images.

[0059] The experimental environment is a PC. The experiment is implemented based on the PyTorch framework. The Adam optimizer is used for training, and the learning rate is set to 0.001. The batch size, communication rounds, and local training cycles are set to 512, 100, and 3 respectively. The experimental environment configuration is: Intel(R) Xeon(R) Gold 5218CPU@2.30GHz processor, 62GB memory, NVIDIA GTX 3090GPU.

[0060] First, in the present invention, the number of clients S=4, and the data sets corresponding to each client are red light (Red), green light (Green), blue light (Blue) and near infrared (NIR). Using the method proposed in the present invention, 4 personalized models will eventually be obtained. These personalized models are used to evaluate the final performance using the method of step 300, and the method of the present invention is compared with other common federated learning methods (FedPer, FedProx, FedBN, FedFV, PSFed-Palm, and "w / o FL" (that is, no federation, only local data sets are used for training without communication). For the sake of fairness, these methods all use the loss function, data set and other experimental settings proposed by the present invention) for comparison. The experimental results show that the model obtained by the method of the present invention achieves the best performance in recognition (average ACC) and verification (average EER). The specific results are shown in Table 1. It can be observed that the method proposed in the present invention achieves the best accuracy ACC and equal error rate EER. In addition, the ROC curves under different methods are as follows Figure 3 As shown, it can be observed that the method of the present invention achieves nearly perfect performance in any cross-spectrum test, which is consistent with the results in Table 1.

[0061] Table 1 Accuracy ACC and equal error rate EER (%) of different recognition methods

[0062]

[0063]

[0064] Second, in order to verify the effectiveness of the loss function proposed in the present invention, an ablation experiment was carried out, that is, one loss function was removed each time, and the experiment was carried out under the same setting, with the hyperparameter k = 97. The experimental results are shown in Table 2. It can be observed that the loss function combination proposed in this paper can effectively enhance the performance of the model, and the cross entropy loss and ArcFace center loss play a major role, while the contrast loss and triplet loss play an auxiliary role. The use of these losses effectively ensures that the local model has good feature extraction capabilities.

[0065] Table 2 Average accuracy ACC (%) under different loss function combinations

[0066]

[0067] Third, two experiments were conducted to further demonstrate the generalization ability of the method proposed in the present invention: First, the underlying feature extraction network was replaced from the original CCNet to CO3Net (Co3net: Coordinate-aware contrastivecompetitive neural network for palmprint recognition), and other experimental settings remained unchanged. The experimental results are shown in Table 3. It can be observed that the present invention still has the best average accuracy ACC and average error rate EER; Second, the client data set was replaced. In order to explore the effectiveness of the method of the present invention in scenarios with greater data heterogeneity, the experimental settings are as follows: the number of clients S=2, the corresponding data sets are the IITD data set and the CASIA data set, the underlying network is CCNet, and other settings are the same as before. The experimental results of the method of the present invention are shown in Table 4. It can be observed that the method of the present invention still has good performance.

[0068] Table 3 Average accuracy ACC and average error rate EER (%) of different methods with CO3Net as the underlying network

[0069]

[0070] Table 4 Accuracy ACC and EER (%) of the method of the present invention in the cross-dataset scenario

[0071]

[0072] Fourth, to further explore the adaptability of the method proposed in the present invention to other privacy protection methods, two experiments were conducted: First, the adaptation exploration with the differential privacy method, specifically, adding Laplace noise with an expectation of 0 and different variances in the local model training process to blur the gradient to a certain extent to play a role in privacy protection. The performance of the final model under Laplace noise with different standard deviations is shown in Table 6. It can be observed that the performance decreases slightly as the variance increases, which shows that the method of the present invention has good adaptability with differential privacy technology. Second, the adaptation study with homomorphic encryption technology, by introducing small model parameter addition (i.e., addition between model parameters) and multiplication (i.e., model parameters multiplied by weight) deviations to simulate homomorphic encryption experiments, since most of the communication rounds of the present invention adopt the federated average method (FedAvg), the dynamic personalized aggregation process can be regarded as a fine-tuning process, so the main role is played by the early federated average process. For simplicity, encryption is only performed during the federated average, and decryption is performed after the federated average aggregation is completed, so as to simulate the entire homomorphic encryption process. The specific experimental results are shown in Table 7. It can be observed that the performance is slightly reduced after the introduction of homomorphic encryption, indicating that the method of the present invention and homomorphic encryption technology have good adaptability.

[0073] Table 6 Average accuracy ACC and average error rate EER (%) under Laplace noise with different variances

[0074]

[0075] Table 7 Average accuracy ACC and average error rate EER (%) with / without homomorphic encryption

[0076]

[0077] The dynamic personalized federated learning method for cross-spectral palmprint recognition in the present invention has the following advantages:

[0078] (1) The present invention still has good performance in palmprint recognition under different spectral scenarios based on federated learning.

[0079] (2) The present invention is adaptable to different underlying networks and still has good performance in cross-domain scenarios.

[0080] (3) The method of the present invention is compatible with technologies such as differential privacy and homomorphic encryption, and can further enhance privacy protection capabilities.

[0081] (4) The present invention has a strong generalization ability. By simply adjusting the hyperparameter k, the extraction method of the present invention can be applied to other similar biometric recognitions, such as knuckle pattern recognition, finger vein recognition, and palm vein recognition.

Claims

1. A dynamic personalized federated learning method for cross-spectral palmprint recognition, characterized in that: The following steps are involved: Step 100: Each client divides the local data set into a training set and a test set, trains a local model using the training set data, and uploads it to the server; Step 200: The server accepts the local models of all clients and selects a suitable strategy for federated aggregation. A combined aggregation method is used: the first k rounds of communication process adopt the method of Federated Averaging (FedAvg), and finally a global model is obtained and distributed to each client for the next round of training; the later communication process adopts the method of Personalized Federated Learning (PFL), and finally a personalized global model corresponding to each client is obtained and distributed to each client for the next round of training; Step 300, for PFL in step 200 to select the best local model, the ratio of the global model to the final personalized model.

2. The method according to claim 1, characterized in that: Step 100 includes: Each client has a data set. Each client divides the local palmprint data set into a training set and a test set. A unified standard is adopted, that is, the first three pictures of each category are selected as the training set, and the remaining pictures are used as the test set. A new loss function combination is proposed: cross entropy loss, supervised contrast loss, triple loss, Arcface center loss, which is used to constrain local training so that the local model has good classification ability; the cross entropy loss formula is: Where K represents the total number of images, M represents the total number of categories, and y i,c is the true label of the i-th sample in the c-th class, p i,c is the predicted probability of the i-th sample in the c-th class; the supervised contrast loss formula is: Where I = {1, ..., 2K} is the sample index set for comparison, P(i) is the positive sample set, A ( i ) is the index set that does not contain i, z i is the feature of the anchor sample, z p is the feature of the positive sample, z a is the feature of negative samples, τ = 0.07 is the temperature coefficient; the weighted sum of these two loss functions is taken as the first loss: L1=w ce L ce +w con L con (3) where w ce =0.8, w con =0.2; triplet loss is used as the second loss function, and the formula is: L2=max(0,d(a,p)-d(a,n)+α) (4) Where a is the anchor sample, p is the positive sample, and n is the negative sample. is the Euclidean distance, c represents the first eigenvector, d represents the second eigenvector, c l Represents the value of the first eigenvector on the lth dimension, d l represents the value of the second eigenvector on the lth dimension, α is the minimum distance between positive samples and negative samples, α = 1; the Arc face loss formula is: Where s = 64 is a scaling factor used to adjust the scale of the loss, e^ is a natural number base, θ is the cosine value of the angle between the input sample feature vector and the center feature vector of the category, k = 0.5 is the angle boundary value used to increase the distance between classes, and y is the label; the center loss formula is: Where M represents the total number of categories in the data set, x i represents the feature vector of the i-th sample, c j Represents the central feature vector of the jth category; the arc face loss and the center loss are summed as the third loss: L3=L Arc +L Cen (7) The total loss formula is: L=L1+L2+L3 (8) Eventually, each client trains a local model and uploads the local model to the server.

3. The method according to claim 2, characterized in that: Step 200 includes: The server accepts the global model from each client and selects an aggregation strategy based on the current number of communication rounds: the first k rounds of communication process adopt the federated average method (FedAvg), that is, the global model comes from the weighted sum of each client, and the global model is distributed to each client for the next round of training. The FedAvg formula is: Where S is the total number of clients, is the local model parameter of the sth client, and θ1 is the global model parameter obtained by aggregation; in the later stage, the personalized federated learning method (PFL) is adopted, that is, the global model is first obtained by the federated averaging method (FedAvg), and then the personalized global model is obtained by weighted summation of the global model and the local model, where the weighted ratio used by the local model is m, and the ratio used by the global model is 1-m. By traversing m from 0 to 1, the step size is set to 0.1, and the best m is selected through the verification in step 300, and this m is used to obtain the best personalized global model in this round of communication, and finally the obtained personalized global model is sent to each client for the next round of training; the PFL formula is: Where m is the best local model ratio obtained through testing, θ′1 is the global model parameter obtained through (9), is the local model parameter of the sth client, Personalize model parameters for the sth client.

4. The method according to claim 1, characterized in that: Step 300 includes: Specifically, there are S clients in total. The trained personalized feature extractor is used to extract the features of the training set and the test set of all clients, match them, and calculate the Euclidean distance. in For each image in the training set, the feature vector is mapped. After mapping the feature vector of an image in the test set, select the label corresponding to the feature vector with the smallest distance as the predicted label, compare it with the true label, calculate the accuracy, and finally get S 2 The accuracy results are averaged as the final evaluation result, and the m used when the average accuracy is the highest is taken as the optimal m.

Images