A federated contrastive structure distribution interaction active learning method and system

Abstract

The application discloses a kind of federal contrast structure distribution interaction active learning method and system, mainly related to federal learning, privacy protection and active learning cross technical field. Including: extracting local unlabeled sample features on client and constructing initial graph structure, generating privacy mask graph representation after irreversible masking and uploading;Server aggregates each client representation to form federal feature space, and introduces cross-client representation for contrast learning to expand negative samples;Calculate the global diversity score and coordination uncertainty score of candidate samples;Based on the score, select the sample to be labeled by the multi-stage strategy of preselection, fine screening and sorting, assign labels and update the model.The beneficial effects of the present application are that it realizes cross-client distribution collaboration and negative sample expansion, effectively reduces labeling redundancy by integrating diversity and uncertainty multi-stage sampling, significantly improves the sampling efficiency and model performance of federal active learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of federated learning, privacy protection and active learning, in particular to a federated contrast structure distribution interaction active learning method and system based on a privacy mask graph. BACKGROUND

[0002] With the wide application of deep learning in computer vision, medical imaging and other fields, the model performance is highly dependent on large-scale and high-quality labeled data. However, the cost of data labeling is high. Active learning, which selects the most informative unlabeled samples for labeling, is a key technology to improve labeling efficiency. Traditional active learning usually runs in a centralized data environment, and its core relies on accurate evaluation of global data distribution, model uncertainty and sample diversity.

[0003] However, in practical scenarios involving sensitive information such as medical and financial fields, data is often stored in different institutions (clients), and due to privacy regulations and compliance requirements, raw data cannot be directly shared. Federated learning provides a framework for collaborative model training without transmitting raw data. However, directly applying active learning to the federated environment faces serious challenges: first, privacy constraints make it impossible to exchange raw data or even fine-grained features, resulting in the failure of traditional global perspective-based diversity estimation and negative sample construction methods, reducing the global representativeness and efficiency of sample selection. Second, in the federated scenario, the data size of each client is limited, and there is a serious non-independent and identically distributed characteristic. Uncertainty estimation and representation learning based on local data are prone to bias and cannot reflect the global importance of samples.

[0004] Existing technologies attempt to combine federated learning and active learning, but most have obvious limitations: some methods only focus on applying differential privacy and other protection technologies in the transmission stage, still relying on frequent global information exchange; other methods completely perform sampling independently at each client, lacking collaboration, resulting in class imbalance and high redundancy in the final labeled set; some methods fail to effectively solve the fundamental problems of representation enhancement and negative sample expansion within the federated framework.

[0005] Therefore, the existing technology lacks a systematic solution that can simultaneously achieve cross-client collaborative data distribution estimation, solve the negative sample scarcity problem, and effectively integrate diversity and uncertainty indicators for efficient sample selection under strict privacy constraints. SUMMARY

[0006] The application aims to provide a privacy mask graph-based federated contrast structure distribution interaction active learning method and system, which realizes cross-client distribution collaboration and negative sample expansion, effectively reduces labeling redundancy through multi-stage sampling of diversity and uncertainty, and significantly improves the sampling efficiency and model performance of federated active learning.

[0007] To achieve the above-mentioned purpose, the application is implemented through the following technical solutions:

[0008] In one aspect, the application provides a federated contrast structure distribution interaction active learning method, comprising the following steps:

[0009] Step S1: extracting features of local unlabeled samples on each client, and constructing an initial graph structure reflecting the internal relationship of local data based on the extracted features;

[0010] Step S2: performing information aggregation and irreversible masking processing on each initial graph structure to generate a privacy mask graph representation, and uploading the privacy mask graph representation to a server, wherein the privacy mask graph representation cannot restore the original sample features;

[0011] Step S3: the server aggregates the privacy mask graph representations from each client to form a federated feature space; in the federated feature space, contrast learning is performed based on the privacy mask graph representations from different clients;

[0012] Step S4: for candidate unlabeled samples in the federated feature space, respectively calculating their global diversity scores and coordination uncertainty scores ;

[0013] Step S5, based on the coordination uncertainty score, the candidate unlabeled samples are pre-selected to obtain a first candidate set; based on the global diversity score, the samples in the first candidate set are fine-screened to obtain a second candidate set; for the samples in the second candidate set, the global diversity score and the coordination uncertainty score are fused for final sorting, and according to the sorting result, a preset number of samples are selected as labeled samples;

[0014] Step S6: distributing the labeled samples to the client to which they belong for labeling, and updating the federated learning model using the newly labeled data.

[0015] Preferably, in step S2, the information aggregation and irreversible masking processing are performed on each initial graph structure to generate a privacy mask graph representation, specifically:

[0016] For the initial graph structure , wherein is a node set, is an edge set, and node corresponds to a sample feature ;

[0017] The node features are weighted and aggregated by a graph attention mechanism to obtain a masked structured representation , that is represents the feature representation of the i-th sample after graph attention aggregation:

[0018] ;

[0019] ;

[0020] wherein, represents the edge weight between node i and node j, is a learnable linear transformation matrix, is an activation function, is a normalized attention weight; The irreversible transformation result of is taken as a privacy mask graph representation.

[0021] Preferably, when constructing the initial graph structure, the edge weight is calculated based on the cosine similarity between the sample feature vectors and

[0022] and a threshold value is set for sparse processing:

[0023] ;

[0024] wherein, is the i-th sample feature, representing the sample feature processed by the feature extraction unit of the local client; represents the i-th sample, which is also the i-th node in the graph structure; similarly, represents the feature of the i-th sample; , that is, the edge weight between the nodes corresponding to the two samples is calculated based on the features of the two samples and

[0025] . .

[0026] Preferably, in step S3, the contrastive learning includes:

[0027] ​​​​​​​a contrast loss function is constructed, which contains not only negative samples from the same client, but also privacy mask map representations from other clients as cross-client negative samples;

[0028] The contrast loss function is represented as:

[0029] ;

[0030] ;

[0031] wherein, 、 is the projection representation of the sample 、 , represents the cosine similarity, is a temperature parameter, is the current client index, represents a set of privacy mask map representations from clients , is a cross-client negative sample item, represents the features of the cross-client negative sample from the client.

[0032] Preferably, in step S4, the global diversity score is calculated, including:

[0033] For unlabeled samples , the global diversity score of the representation of the sample in the high-scattering feature space is calculated:

[0034] ;

[0035] wherein, represents the features of the unlabeled sample , represents the mean of the first 5 values, is a distance metric, the first term represents the distance to the nearest neighbor sample in the federated feature space ; the second term represents the average distance to its nearest 5 neighbors.

[0036] Preferably, in step S4, the coordination uncertainty score is calculated, including:

[0037] Based on the prediction probability vector of the sample by the target model, the basic uncertainty value :

[0038] ;

[0039] wherein, represents a sample, is a probability concentration measure function, is a variance of a probability vector is a probability vector of a target model based on current label pool deep learning training ; is a probability vector of a target model represents a number of categories, is a real probability;

[0040] training a trainable global uncertainty indicator , whose input is a hidden layer feature of a sample , and whose output is a predicted uncertainty value ; the uncertainty value is a floating point number greater than 0, so that through the output, the uncertainty value of the data is fitted without touching other private data;

[0041] the uncertainty indicator is attached to a local feature extraction unit of federated contrast learning, and its input is each hidden layer network feature of the feature extraction unit; the structure of the uncertainty indicator is a multilayer perceptron model of a fully connected neural network with the same number of layers as the feature extraction unit, and the last layer is a single-layer regression network with a dimension of 1024 compressed to 1 dimension and an activation function layer, and the output is a floating point number greater than 0;

[0042] the coordinated uncertainty score is obtained by fusing the outputs of the global uncertainty indicator and the local uncertainty indicators of each client:

[0043] ;

[0044] wherein, is the total number of clients, represents a trainable local uncertainty indicator, represents an unlabeled sample.

[0045] Preferably, the uncertainty indicator network is trained by a pair-wise ranking loss, and the ranking loss is:

[0046] ; ​

[0047] wherein, , is a pair of training samples, and satisfies the true value relationship , is a marginal hyperparameter greater than 0.

[0048] Preferably, step S5 comprises the following sub-steps:

[0049] Step S51: Each client calculates the coordination uncertainty score of all samples in its local unlabeled sample pool , and reports the score to the server;

[0050] Step S52: The server sorts all candidate samples according to , and selects the top samples to form the first candidate set, wherein is greater than the final number of samples to be selected ;

[0051] Step S53: Calculate the global diversity score of all samples in the first candidate set , and sort the samples in the set according to , and select the top samples to form the second candidate set, wherein ;

[0052] Step S54: For each sample in the second candidate set, calculate its comprehensive score , wherein and are weighting coefficients obtained by federated big data statistical analysis;

[0053] Step S55: According to the comprehensive score , sort the samples in the second candidate set in descending order, and select the top samples as the final unlabeled samples.

[0054] On the other hand, the present application also provides a kind of active learning system of federated contrast structure distribution interaction, for realizing like above-mentioned one kind of active learning method of federated contrast structure distribution interaction, comprising:

[0055] A plurality of client modules for performing local data processing; comprising:

[0056] A feature extraction unit for extracting features from local unlabeled samples;

[0057] A privacy mask graph construction unit for constructing an initial graph structure based on the features and performing information aggregation and irreversible masking to generate a privacy mask graph representation;

[0058] a local computing and communication unit configured to upload the privacy mask representation and receive the samples to be labeled and model updates;

[0059] a server module configured to perform global coordination processing, including:

[0060] a federated space management unit configured to aggregate the privacy mask representations uploaded by the clients to form a federated feature space;

[0061] a contrast learning enhancement unit configured to perform contrast learning on the privacy mask representations across the clients in the federated feature space to expand the negative samples;

[0062] a sample value evaluation unit configured to calculate a global diversity score and a coordination uncertainty score for the candidate samples;

[0063] a multi-stage sampling decision unit configured to perform pre-selection, fine screening, and final ranking based on the scores to select the samples to be labeled and assign them to the corresponding clients.

[0064] Compared with the prior art, the present application has the following advantages:

[0065] 1. By constructing and exchanging irreversible privacy mask representations, the present application can realize the interaction and coordination estimation of data distribution information without exposing the original samples and fine-grained features of each client, thereby fundamentally meeting the strict compliance requirements of the medical, financial, and other fields for data not leaving the local (such as "data not leaving the hospital").

[0066] 2. The contrast structure distribution interaction mechanism proposed by the present application uses the masked representations from other clients as negative sample sources, significantly expanding the diversity of the negative sample pool in contrast learning. This alleviates the problem of insufficient discriminability of representation learning caused by the scarcity or singleness of samples from a single client, thereby learning more robust and discriminative feature representations.

[0067] 3. By fusing global diversity evaluation based on the federated feature space and coordination uncertainty measurement combining local and global information, and adopting a multi-stage (pre-selection-fine screening-ranking) sampling process, the present application can systematically select samples with high information content (uncertainty) and strong global representativeness (diversity). This effectively avoids the redundancy sampling or class imbalance caused by relying solely on local uncertainty, achieving higher model performance improvement with fewer labeled samples.

[0068] 4、The application corrects the local deviation by global perspective diversity evaluation and coordination, so that the selected labeled sample set can better reflect the global data distribution. The updated federated model trained using these samples has significantly improved generalization performance and is more suitable for the data heterogeneity challenges that exist in the real world. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 is a method flowchart of the application;

[0070] Figure 2 is a system structure schematic diagram of the application. DETAILED DESCRIPTION

[0071] The application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not to limit the scope of the application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content taught by the application, and these equivalent forms also fall within the scope defined by the application.

[0072] In the application, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicate the orientation or positional relationship shown in the drawings, which is only a relationship word determined for the convenience of describing the structural relationship of the components or elements of the application, and cannot be understood as a limitation on the application.

[0073] Embodiment:

[0074] As shown in Figure 1 , the embodiment provides a federated contrast structure distribution interaction active learning method, which comprises the following steps:

[0075] Step S1: extracting features of local unlabeled samples at each client, and constructing an initial graph structure reflecting the internal relationship of local data based on the extracted features;

[0076] Step S2: performing information aggregation and irreversible masking processing on each initial graph structure to generate a privacy mask graph representation, and uploading the privacy mask graph representation to the server, wherein the privacy mask graph representation cannot restore the original sample features;

[0077] Step S3: the server aggregates the privacy mask graph representations from each client to form a federated feature space; in the federated feature space, contrast learning is performed based on the privacy mask graph representations from different clients;

[0078] Step S4: calculating the global diversity score of each candidate unlabeled sample in the federated feature space Coordinated uncertainty score ;

[0079] Step S5, based on the coordinated uncertainty score, preselect the candidate unlabeled samples to obtain a first candidate set; based on the global diversity score, fine-screen the samples in the first candidate set to obtain a second candidate set; for the samples in the second candidate set, fuse the global diversity score and the coordinated uncertainty score to perform final sorting, and select a preset number of samples as the to-be-labeled samples according to the sorting result;

[0080] Step S6: distribute the to-be-labeled samples to the clients to which the to-be-labeled samples belong for labeling, and update the federated learning model by using the newly labeled data.

[0081] Specifically:

[0082] The embodiment first constructs a privacy mask graph for the local samples in each client: for each sample, the encoder extracts a feature vector (wherein is the sample index, , is the feature dimension); based on the cosine similarity between the feature vectors, an initial candidate edge is constructed, and the calculation formula is:

[0083] ;

[0084] In the formula, denotes the similarity / edge weight of node and node , denotes the vector norm, and the threshold is used to control the connectivity sparsity (usually set by experience or cross-validation, for example ). When , it means no edge or is truncated to maintain sparsity.

[0085] On the initial graph , the neighborhood information is weighted and aggregated by using a graph attention or other graph aggregation operator to generate a masked structured representation, specifically:

[0086] ;

[0087] In the formula, is the masked structured representation of node , is a learnable linear transformation matrix, is an activation function, and the weight denotes the neighbor of node normalized attention coefficients. Through the aggregation, the original fine-grained features are hidden in the neighborhood aggregation results, thereby forming a privacy mask representation for cross-client exchange; the client only reports or its irreversible transform / summary, without uploading the original ; wherein, is the sample feature of the i-th sample processed by the feature extraction unit of the local client, is the feature representation of the i-th sample after graph attention aggregation, the upper wave line indicates that it has undergone one graph attention aggregation, i represents the i-th sample, and also the i-th node in the graph structure, represents the feature of the j-th sample, and represent the features of two samples, and the weight of the edge between the nodes corresponding to the two samples is calculated based on the features of the two samples ; In the formula, k is the subscript of the summation formula, and it is assumed that there are N samples in total, and the denominator of the formula represents , …, the sum of the weights with the first subscript of 1.

[0088] Based on the privacy mask graph, the embodiment proposes a contrastive structure distribution interaction (COSIN) mechanism to improve the representation learning quality and negative sample diversity. The local contrastive loss adopts a typical form:

[0089] ;

[0090] wherein is the projection representation (obtained by the projection head after the encoder) of the sample , represents the cosine similarity, and the temperature parameter controls the distribution smoothing (the empirical value is often ). To overcome the problem of scarcity of negative samples on a single client, the embodiment incorporates the masked representation from other clients into the negative sample pool, and the loss function is modified as:

[0091] ,

[0092] ;

[0093] In the formula, represents the set of masked representations from the client , is the feature of the cross-client negative sample from the client. The cross-client negative sample item is calculated by secure aggregation or summary exchange, which not only expands the diversity of negative samples, but also guarantees privacy.

[0094] In the federated feature space This embodiment proposes a global diversity score To measure the representativeness of a candidate sample, define:

[0095] ;

[0096] where is the unlabeled sample to be evaluated, is its representation in the federated feature space, a Euclidean norm or equivalent measure of cosine similarity, denotes the average over the 5 representations closest to The first term guarantees novelty (larger nearest distance is more novel), the second measures local coverage, and their combination is used to achieve a feasible global representativeness evaluation under privacy constraints.

[0097] To robustly evaluate uncertainty and reduce bias from distributed non-identical data, this embodiment proposes a coordinated uncertainty and uncertainty indicator .One implementation form is:

[0098] ;

[0099] The formula is the predicted probability vector of class by the model, is the variance of this vector (measuring dispersion), is a probability concentration function (e.g. information entropy or its normalized version), is the highest confidence class probability of the model output (if the true label is unknown, replace it with the maximum predicted probability). To implement end-to-end prediction, design the uncertainty indicator :

[0100] ;

[0101] where denotes the hidden layer representation extracted by the target model parameters , and the function is the indicator network. To maintain the consistency of the ranking of uncertainty, train :

[0102] ;

[0103] where satisfies the ground truth is a marginal hyperparameter (e.g. Under a federated setup, global training is possible. Also retain local Ultimately, the uncertainty is reconciled. The result obtained through fusion:

[0104] ;

[0105] in The total number of clients is used for merging, which can be achieved by weighted average or other weighting rules.

[0106] In terms of sampling strategy, this embodiment adopts a multi-stage candidate screening process to balance uncertainty and diversity: the client or center first bases the sampling on the coordinated uncertainty. Initial selection of unlabeled samples (e.g., by...) (Selecting the top 4B candidates); then utilizing within the candidate set. A diversity screening process is performed to remove redundancy and ensure coverage (e.g., selecting the top-2B). Finally, the screening results are sorted according to the fused diversity and uncertainty scores, and the final labeled samples (top-B) are selected. This process avoids redundant sampling caused by relying solely on local uncertainty, and ensures the representativeness and information content of the sample set under limited communication and privacy constraints, thereby effectively improving labeling efficiency and enhancing the generalization performance of subsequent models.

[0107] This embodiment also provides optional implementation variations for different application scenarios:

[0108] In cross-hospital annotation of medical images, noise can be added to the masked representation to meet hospital privacy policies;

[0109] In financial risk control scenarios, more stringent statistical summary exchange can be adopted to avoid the leakage of sensitive information;

[0110] On resource-constrained edge devices, reporting frequency can be reduced and sparser mask maps can be used to lower communication and computing costs.

[0111] like Figure 2 As shown, this embodiment also provides an active learning system for distributed interaction of federated contrast structures, including:

[0112] Several client modules are used to perform local data processing; including:

[0113] The feature extraction unit is used to extract features from local unlabeled samples.

[0114] A privacy mask graph construction unit is used to construct an initial graph structure based on the features and perform information aggregation and irreversible masking to generate a privacy mask graph representation;

[0115] a local computing and communication unit configured to upload the privacy mask graph representation and receive the samples to be labeled and model updates;

[0116] a server module configured to perform global coordination processing, including:

[0117] a federated space management unit configured to aggregate the privacy mask graph representations uploaded by the clients to form a federated feature space;

[0118] a contrast learning enhancement unit configured to perform contrast learning on the privacy mask graph representations across the clients in the federated feature space to expand the negative samples;

[0119] a sample value evaluation unit configured to calculate a global diversity score and a coordination uncertainty score of the candidate samples;

[0120] a multi-stage sampling decision unit configured to perform preliminary selection, fine screening and final sorting based on the scores to select the samples to be labeled and distribute them to the corresponding clients.

[0121] The embodiment is directed to the medical imaging field (e.g., CT, MRI or X-ray images), and is designed for a distributed privacy dataset construction system and model learner for multiple affiliated hospitals and cooperative hospitals of a university.

[0122] Each of the multiple hospitals (clients ) has a large number of unlabeled or labeled sensitive medical image datasets . Due to strict privacy regulations and compliance requirements for data not leaving the hospital, raw image data and fine-grained features cannot be shared between hospitals or with a central server. The embodiment aims to apply the method of the present application to cooperatively select the most valuable medical images for labeling under the premise of protecting patient privacy, thereby minimizing labeling costs and quickly training a high-performance medical diagnosis model. The specific process is as follows:

[0123] 1. Construction and exchange of privacy mask graphs:

[0124] Within each hospital (client ), the system extracts a high-dimensional feature vector x of each image x from the local medical images . Based on the similarity between local features , a graph structure reflecting the internal relationship of the local data is constructed . Then, irreversible masking is performed on the node features of the graph to generate a privacy mask graph representation. The client only uploads this masked structured distribution representation to the central server in the federated server cluster that aggregates the local models of each client.

[0125] 2. Contrast structure distribution interaction (COSIN) mechanism:

[0126] The center server in the federal server cluster aggregates the local models of each client, collects the privacy mask graph of all clients, and enhances the learning of feature representation proposed by the mechanism, thereby constructing a federal feature space, and expanding the cross-client negative samples: on the basis of the local contrast learning loss, the masked representation from other clients is incorporated into the negative sample pool as cross-client negative samples, and the cross-client negative samples are expanded.

[0127] 3. Distributed sample value evaluation and dataset construction:

[0128] In the federal feature space, the global value of the sample (including global diversity and coordination uncertainty) is calculated systemically to ensure that the selected sample has global representativeness and information richness. The embodiment adopts a multi-stage process, taking into account sampling efficiency and quality. First, the candidate samples are obtained through preliminary selection, and then the diversity is refined, and the refined samples are finally sorted and selected to select the most valuable samples. Finally, the hospital is sent for artificial expert annotation, and the newly annotated data is used for local model fine-tuning, and the global medical model is updated through the federal strategy of the privacy graph.

[0129] The performance of the above embodiment is evaluated: compared with the existing federal active learning method, the framework based on the privacy mask graph can realize data collaboration under the premise of privacy, by exchanging only irreversible privacy mask graph abstracts, strictly complying with the privacy requirement of "data not leaving the hospital", and selecting the most valuable data for annotation, enhancing the robustness and discriminability of medical image features. Compared with other methods, the model accuracy of the system is improved by 8.7%, and the number of samples required is reduced by 11.5% under the premise of achieving the same training effect, so that the quality and representativeness of the final selected sample set are higher, thereby training a diagnostic model with better generalization performance with a smaller annotation budget.

[0130] The above is a specific description of the preferred embodiment of the present application, but the present application is not limited to the described embodiments, and those skilled in the art can make various equivalent modifications or replacements without deviating from the spirit of the present application. These equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. A federated contrastive structured distribution interaction active learning method, characterized in that, The method comprises the following steps: Step S1: extracting features of local unlabeled samples at each client, and constructing an initial graph structure reflecting the internal relationship of local data based on the extracted features; Step S2: performing information aggregation and irreversible masking processing on each initial graph structure to generate a privacy mask graph representation, and uploading the privacy mask graph representation to a server, wherein the privacy mask graph representation cannot be restored to the original sample features; Step S3: the server aggregates the privacy mask graph representations from each client to form a federated feature space; in the federated feature space, contrastive learning is performed based on the privacy mask graph representations from different clients; Step S4: Calculate the global diversity score of each candidate unlabeled sample in the federated feature space with the coordination uncertainty score ; Step S5, based on the coordination uncertainty score, the candidate unlabeled samples are preliminarily selected to obtain a first candidate set; based on the global diversity score, the samples in the first candidate set are finely screened to obtain a second candidate set; for the samples in the second candidate set, the global diversity score and the coordination uncertainty score are fused for final sorting, and a preset number of samples are selected as the unlabeled samples according to the sorting result; Step S6: distributing the unlabeled samples to the clients to which they belong for labeling, and updating the federated learning model using the newly labeled data; In step S2, the information aggregation and irreversible masking processing are performed on each initial graph structure to generate a privacy mask graph representation, specifically: For an initial graph structure wherein is a set of nodes, is a set of edges, nodes corresponding sample features ; The node features are weighted and aggregated by a graph attention mechanism to obtain a masked structured representation That is, represents the feature representation of the i-th sample after graph attention aggregation: sample: ； ； wherein, represents a node and an edge weight between nodes , is a learnable linear transformation matrix, is an activation function, is a normalized attention weight; In the above equation, k is the index of summation, and the denominator of the equation represents , , the sum of these weights with the first index of 1. The irreversible transform results on are represented as privacy mask maps; In step S4, the global diversity score is calculated comprising: For unlabeled samples Calculate its characteristics in the highly dispersed feature space. The representation in Global diversity score : ； where, denotes the features of the unlabeled sample , denotes the mean of the first 5 values, is the distance metric, the first term denotes the distance to the nearest neighbor in the federated feature space ; the second term denotes the average distance to its 5 nearest neighbors; In step S4, the coordination uncertainty score is calculated comprises: predicting a probability vector for a sample based on a target model :​​ ； in, To represent a sample, It is a probability concentration measure function. Probability vector variance The target model is based on the current label pool deep learning training. of The probability vector, Indicates the number of categories. This represents the true probability. Train a trainable global uncertainty indicator Its input is the hidden layer features of the sample. The output is the uncertainty value of the prediction. The uncertainty value The output is a floating-point number greater than 0, allowing the data to be fitted without accessing other private data. Uncertain value; the uncertainty indicator a local feature extractor attached to the federated contrastive learning, whose input is the network features of each hidden layer of the feature extractor, the uncertainty indicator The structure of the uncertainty indicator is a multi-layer perceptron model with the same number of layers as the feature extractor, each layer is a fully connected neural network, the last layer is a single-layer regression network with 1024 dimensions compressed to 1 dimension and an activation function layer, and the output is a floating-point number greater than 0; The coordinated uncertainty score is obtained by fusing the output of the global uncertainty indicator with the output of the local uncertainty indicator of each client. ； wherein, is the total number of clients, denotes a trainable local uncertainty indicator, denotes an unlabeled sample.

2. The active learning method of federated contrastive structured distribution interaction according to claim 1, characterized in that, In constructing the initial graph structure, its edge weight Based on the sample feature vector The cosine similarity calculation of And set the threshold Sparse processing: ； in, For the first Each sample feature represents the sample features obtained by the feature extraction unit of the local client. Indicates the first The nth sample, which is also the nth in the graph structure. Each node; similarly, Then it means the first Features of each sample; i.e. based on two samples with The feature of the two samples is calculated to get the edge weight between the nodes corresponding to the two samples .

3. The active learning method of federated contrastive structured distribution interaction according to claim 1, characterized in that, In step S3, the contrastive learning comprises: A contrastive loss function is constructed, which not only contains negative samples from the same client, but also contains privacy mask graph representations from other clients as cross-client negative samples; The contrastive loss function is represented as: ； ； wherein, , is a projection representation of the sample , , denotes a cosine similarity, is a temperature parameter, is a current client index, denotes a set of privacy mask map representations from clients , is a cross-client negative sample item, denotes a feature from a cross-client negative sample from a client.

4. The active learning method of federated contrastive structured distribution interaction according to claim 1, characterized in that, The uncertainty indicator network Training proceeds through a pairwise ranking loss, which is: ； wherein, , is a pair of training samples, and satisfies the true value relationship , is a marginal hyperparameter greater than 0.

5. The active learning method of federated contrastive structured distribution interaction according to claim 1, wherein, Step S5 comprises the following sub-steps: Step S51: Each client calculates the coordinated uncertainty score for all samples in its local pool of unlabeled samples and reports the score to the server; Step S52: the server sorts all candidate samples according to the first candidate set is composed of the top samples in the sorting result, wherein the number of the top samples is greater than the final number of samples to be selected ; Step S53: Calculate the global diversity score of all samples in the first candidate set , and according to , the samples in the set are sorted, and the top samples are selected to form the second candidate set, wherein ; Step S54: for each sample in the second candidate set, calculate its comprehensive score wherein and are weighting coefficients, obtained by federated big data statistical analysis; Step S55: selecting the top samples in the second candidate set as the final samples to be annotated according to the comprehensive scores the second candidate set as the final samples to be annotated.​ 6. A federated contrastive structured distribution interaction active learning system for implementing the federated contrastive structured distribution interaction active learning method of any one of claims 1-5. Including: A plurality of client modules for performing local data processing; comprising: A feature extraction unit for extracting features of local unlabeled samples; A privacy mask graph construction unit for constructing an initial graph structure based on the features and performing information aggregation and irreversible masking to generate a privacy mask graph representation; A local computing and communication unit for uploading the privacy mask graph representation, receiving unlabeled samples for labeling, and updating the model; A server module for performing global collaborative processing, comprising: A federated space management unit for aggregating the privacy mask graph representations uploaded by each client to form a federated feature space; A contrastive learning enhancement unit for performing contrastive learning in the federated feature space using cross-client privacy mask graph representations to expand negative samples; A sample value evaluation unit for calculating the global diversity score and the coordination uncertainty score of the candidate samples; A multi-stage sampling decision unit for performing preliminary selection, fine screening, and final sorting based on the scores to select unlabeled samples and distribute them to the corresponding clients.

Images