Active learning method and system for federal contrast structure distribution interaction

Abstract

The invention discloses an active learning method and system for federal contrast structure distribution interaction, and mainly relates to the technical field of federal learning, privacy protection and active learning crossing. Comprising the following steps: extracting local unlabeled sample features at a client, constructing an initial graph structure, generating a privacy mask graph representation after irreversible masking, and uploading the privacy mask graph representation; the server aggregates each client representation to form a federated feature space, and introduces a cross-client representation to carry out comparative learning so as to expand a negative sample; calculating a global diversity score and a coordination uncertainty score of the candidate sample; and selecting a to-be-labeled sample through a multi-stage strategy of primary selection, fine screening and sorting based on the score, distributing labels and updating the model. The method has the advantages that cross-client distributed collaboration and negative sample expansion are achieved, labeling redundancy is effectively reduced through multi-stage sampling fusing diversity and uncertainty, and the sampling efficiency and model performance of federal active learning are remarkably improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of federated learning, privacy protection, and active learning, specifically an active learning method and system based on a federated contrastive structure distribution interaction using a privacy mask graph. Background Technology

[0002] With the widespread application of deep learning in fields such as computer vision and medical imaging, model performance is highly dependent on large-scale, high-quality labeled data; however, data labeling is costly. Active learning, by strategically selecting the most informative unlabeled samples for labeling, has become a key technology for improving labeling efficiency. Traditional active learning typically operates in centralized data environments, and its core relies on accurate assessment of global data distribution, model uncertainty, and sample diversity.

[0003] However, in real-world scenarios involving sensitive information, such as healthcare and finance, data is often scattered across 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 a federated environment faces significant challenges: First, privacy constraints prevent the exchange of raw data and even fine-grained features, rendering traditional global-perspective diversity estimation and negative sample construction methods ineffective, reducing the global representativeness and efficiency of sample selection. Second, in federated scenarios, the data scale of each client is limited, and they generally exhibit severe non-independent and identically distributed characteristics. Uncertainty assessment and representation learning based solely on local data are prone to bias and fail to reflect the global importance of samples. Third, self-supervised methods such as contrastive learning rely on a large number of diverse negative samples. Under distributed privacy constraints, the available sources of negative samples are severely limited, leading to a decline in representation quality and consequently affecting subsequent sample value assessment.

[0004] Existing technologies attempt to combine federated learning and active learning, but most have obvious limitations: some methods focus only on applying protection techniques such as differential privacy in the transmission process and still rely on frequent global information exchanges; other methods perform sampling independently on each client, lacking coordination, resulting in an imbalance of categories and high redundancy in the final labeled set; and some methods have failed to effectively solve the fundamental problems of representation enhancement and negative sample augmentation within the federated framework.

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

[0006] The purpose of this invention is to provide an active learning method and system based on privacy mask graph-based federated contrast structure distribution interaction. It realizes cross-client distribution collaboration and negative sample augmentation. By integrating multi-stage sampling with diversity and uncertainty, it effectively reduces label redundancy and significantly improves the sampling efficiency and model performance of federated active learning.

[0007] To achieve the above objectives, the present invention employs the following technical solution: On the one hand, this invention provides an active learning method for distributed interaction of federated contrastive structures, comprising the following steps: Step S1: Extract features from unlabeled local samples on each client and construct an initial graph structure that reflects the inherent relationships of local data based on the extracted features; Step S2: Perform information aggregation and irreversible masking on each initial graph structure to generate a privacy mask graph representation, and upload the privacy mask graph representation to the server, wherein the privacy mask graph representation cannot restore the original sample features; Step S3: The server aggregates privacy mask image representations from various clients to form a federated feature space; in the federated feature space, comparative learning is performed based on the privacy mask image representations from different clients; Step S4: For each candidate unlabeled sample in the federated feature space, calculate its global diversity score. With reconciliation uncertainty score ; Step S5: Initially select unlabeled candidate samples based on the coordination uncertainty score to obtain the first candidate set; further screen the samples in the first candidate set based on the global diversity score to obtain the second candidate set; finally sort the samples in the second candidate set by combining their global diversity score and coordination uncertainty score, and select a preset number of samples as unlabeled samples according to the sorting results. Step S6: Assign the samples to be labeled to their respective clients for labeling, and update the federated learning model using the newly labeled data.

[0008] Preferably, in step S2, information aggregation and irreversible masking are performed on each initial graph structure to generate a privacy mask graph representation, specifically as follows: For the initial graph structure ,in For a set of nodes, Let be a set of edges, nodes Corresponding sample features ; By using a graph attention mechanism to weighted aggregate node features, a masked structured representation is obtained. ,Right now This indicates the first step after graph attention aggregation. Feature representation of each sample: ; ; in, Represents a node With nodes The boundary rights between them It is a learnable linear transformation matrix. For activation function, Normalized attention weights; right The result of the irreversible transformation is represented as a privacy mask image.

[0009] Preferably, when constructing the initial graph structure, its edge weights Based on sample feature vectors and Calculate the cosine similarity and set a threshold. Perform sparsification: ; 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; That is, based on two samples and Feature calculation yields the edge weights between the nodes corresponding to the two samples. .

[0010] Preferably, in step S3, the comparative learning includes: Construct a contrastive loss function that includes not only negative samples from the same client, but also privacy mask map representations from other clients as cross-client negative samples; The contrastive loss function is expressed as follows: ; ; in, , For the sample , The projection representation, Represents cosine similarity. For temperature parameters, For the current client index, Indicates that it comes from the client. The privacy mask image represents the set. For cross-client negative sample items, This represents the features of cross-client negative samples from the client.

[0011] Preferably, in step S4, a global diversity score is calculated. include: For unlabeled samples Calculate its characteristics in the highly dispersed feature space. The representation in Global diversity score : ; in, Indicates unlabeled samples Features This represents the average of the first 5 values. For distance measurement, the first term express To the federal feature space The distance to the nearest neighbor sample in the middle; the second term express The average distance to its five nearest neighbors.

[0012] Preferably, in step S4, the consistency uncertainty score is calculated. include: Based on the target model, the samples Predicted probability vector Calculate the fundamental uncertainty value : ; 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 prediction uncertainty value. 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 The local feature extraction unit, which is attached to federated contrastive learning, takes as input the hidden layer network features of the feature extraction unit and an uncertainty indicator. The structure is a multilayer perceptron model with the same number of layers as the feature extraction unit, each layer being a fully connected neural network, and the last layer being a single-layer regression network compressed from 1024 dimensions to 1 dimension and an activation function layer, with the output being a floating-point number greater than 0. The coordination uncertainty score This is obtained by fusing the outputs of the global uncertainty indicator and the local uncertainty indicators of each client: ; in, Total number of clients This indicates that a trainable local uncertainty indicator is available. This indicates an unlabeled sample.

[0013] Preferably, the uncertainty indicator network Training is performed using a pairwise ranking loss, where the ranking loss is: ; in, , Given a pair of training samples that satisfy the truth relation , It is a marginal hyperparameter greater than 0.

[0014] Preferably, step S5 includes the following sub-steps: Step S51: Each client calculates the coordinated uncertainty score for all samples in its local unlabeled sample pool. The score will be reported to the server. Step S52: The server, according to Sort all candidate samples and select the top-ranked ones. The samples constitute the first candidate set, wherein 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 Sort the samples in the set and select the top-ranked ones. The samples constitute the second candidate set, wherein ; Step S54: For each sample in the second candidate set, calculate its comprehensive score. ,in and These are weighting coefficients, obtained through statistical analysis of federal big data. Step S55: Based on the overall score Sort the samples in the second candidate set in descending order and select the top-ranked samples. The samples were used as the final samples to be labeled.

[0015] On the other hand, the present invention also provides an active learning system for distributed interaction of federated contrastive structures, used to implement the active learning method for distributed interaction of federated contrastive structures as described above, comprising: Several client modules are used to perform local data processing; including: The feature extraction unit is used to extract features from local unlabeled samples. 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; The local computing and communication unit is used to upload the privacy mask image representation and receive samples to be labeled for labeling and model updates; The server module is used to perform global collaborative processing, including: The Federation Space Management Unit is used to aggregate the privacy mask image representations uploaded by various clients to form a federated feature space. A contrastive learning enhancement unit is used to augment negative samples by performing contrastive learning in the federated feature space using privacy mask graph representations across clients. The sample value assessment unit is used to calculate the global diversity score and the coordination uncertainty score of the candidate samples. A multi-stage sampling decision unit is used to perform preliminary selection, fine screening and final sorting based on the scores, so as to select samples to be labeled and allocate them to the corresponding clients.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing and exchanging irreversible privacy mask graph representations, this invention enables the interaction and collaborative estimation of data distribution information without exposing the original samples and fine-grained features of each client. This fundamentally meets the strict compliance requirements of fields such as healthcare and finance that data should not leave the local area (e.g., "data should not leave the hospital").

[0017] 2. The contrastive structure distribution interaction mechanism proposed in this invention uses masked representations from other clients as a source of negative samples, significantly expanding the diversity of the negative sample pool in contrastive learning. This alleviates the problem of insufficient discriminative power in representation learning caused by the scarcity or uniformity of samples from a single client, thereby learning more robust and discriminative feature representations.

[0018] 3. This invention integrates global diversity assessment based on federated feature space and a coordinated uncertainty measure combining local and global information, and employs a multi-stage (initial selection-refinement-ranking) sampling process. This allows for the systematic selection of samples that possess both high information content (uncertainty) and strong global representativeness (diversity). This effectively avoids redundant sampling or class imbalance caused by relying solely on local uncertainty, achieving higher model performance improvements with fewer labeled samples.

[0019] 4. This invention, through a global perspective of diversity assessment and the correction of local biases by coordinating uncertainties, enables the selected set of labeled samples to better reflect the global data distribution. Using these samples to train an updated federated model significantly improves its generalization performance and makes it better suited to the data heterogeneity challenges prevalent in the real world. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0021] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0022] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0023] Example: like Figure 1 As shown, this embodiment provides an active learning method for distributed interaction of federated contrast structures, including the following steps: Step S1: Extract features from unlabeled local samples on each client and construct an initial graph structure that reflects the inherent relationships of local data based on the extracted features; Step S2: Perform information aggregation and irreversible masking on each initial graph structure to generate a privacy mask graph representation, and upload the privacy mask graph representation to the server. The privacy mask graph representation cannot be used to restore the original sample features. Step S3: The server aggregates privacy mask image representations from various clients to form a federated feature space; in the federated feature space, comparative learning is performed based on the privacy mask image representations from different clients; Step S4: For each candidate unlabeled sample in the federated feature space, calculate its global diversity score. With reconciliation uncertainty score ; Step S5: Initially select unlabeled candidate samples based on the coordination uncertainty score to obtain the first candidate set; further screen the samples in the first candidate set based on the global diversity score to obtain the second candidate set; finally sort the samples in the second candidate set by combining their global diversity score and coordination uncertainty score, and select a preset number of samples as unlabeled samples according to the sorting results. Step S6: Assign the samples to be labeled to their respective clients for labeling, and update the federated learning model using the newly labeled data.

[0024] Specifically: This embodiment first constructs a privacy mask map for local samples within each client: for each sample, the encoder extracts the feature vector. (in For sample index, , (where the feature dimension is 1). Initial candidate edges are constructed based on the cosine similarity between feature vectors, and the calculation formula is as follows: ; In the formula Represents a node With nodes Similarity / edge weight, Represents vector norm, threshold Used to control connectivity sparsity (often set empirically or through cross-validation, for example) ).when The time indicates no boundary or truncation to maintain sparsity.

[0025] In the initial diagram The above uses graph attention or other graph aggregation operators to weighted aggregate neighborhood information to generate a masked structured representation, specifically: ; In the formula For nodes Masked structured representation It is a learnable linear transformation matrix. For activation function, weights Indicates neighbors For nodes The normalized attention coefficients. Through this aggregation, the original fine-grained features are hidden in the neighborhood aggregation result, thus forming a privacy-masked representation for cross-client exchange; the client only reports outwards. Or its irreversible transformation / summary, do not upload the original. ;in, It is the sample feature obtained by the feature extraction unit of the local client for the i-th sample. This is the feature representation of the i-th sample after graph attention aggregation. The wavy line above it indicates that it has undergone one graph attention aggregation. i represents the i-th sample, which is also the i-th node in the graph structure. Then it represents the feature of the j-th sample. and This represents the features of two samples, and the weights of the edges between the corresponding nodes of the two samples are calculated based on these features. ; In this formula, k is the subscript of the summation formula. Assuming there are N samples, the denominator of the formula represents... , ,…, The sum of the weights whose first index is 1.

[0026] Based on the privacy mask image, this embodiment proposes a Contrastive Structure Distribution Interaction (COSIN) mechanism to improve representation learning quality and negative sample diversity. The local contrastive loss takes a typical form: ; in For the sample The projection representation (obtained from the projection head behind the encoder). Represents cosine similarity, temperature parameter Controlling distribution smoothing (empirical values ​​are often taken as follows) To overcome the problem of scarce negative samples from a single client, this embodiment incorporates masked representations from other clients into the negative sample pool, and the loss function is modified as follows: , ; In the formula Indicates that it comes from the client. The masked representation of the set, These are features of cross-client negative samples from the client side. Cross-client negative sample items. Computation via secure aggregation or digest exchange expands negative sample diversity while protecting privacy.

[0027] In the federal characteristic space Above, this embodiment proposes a global diversity score. To measure the representativeness of candidate samples, it is defined as: ; in For unlabeled samples to be evaluated, Let it be represented in the federal feature space. Equivalent measures such as Euclidean norm or cosine similarity can be used. Indicates to and The five nearest neighbors are averaged. The first term guarantees novelty (the larger the nearest neighbor distance, the more novel), and the second term measures local coverage. The combination of the two is used to achieve a feasible assessment of global representativeness under privacy-constrained conditions.

[0028] To robustly assess uncertainty and reduce bias caused by distributed, non-similar data, this embodiment proposes a reconciled uncertainty. With uncertainty indicator . One implementation is as follows: ; Mode For model pair The predicted probability vector of the class. This is the variance of the vector (a measure of its dispersion). It is a probability concentration function (e.g., information entropy or its normalized version). The highest confidence class probability output by the model (if the true label is unknown, the maximum predicted probability is used instead). An uncertainty indicator is designed to achieve end-to-end prediction. : ; in Indicates the parameters of the target model Extracted hidden layer representation, function This is an indicator network. To maintain consistency in the ranking of uncertainties, a pairwise ranking loss is used during training. : ; in Satisfy truth value For marginal hyperparameters (e.g.) Under a federated setup, global training is possible. Also retain local Ultimately, the uncertainty of reconciliation The result obtained through fusion: ; in The total number of clients is used for merging, which can be achieved by weighted average or other weighting rules.

[0029] 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.

[0030] This embodiment also provides optional implementation variations for different application scenarios: In cross-hospital annotation of medical images, noise can be added to the masked representation to meet hospital privacy policies; In financial risk control scenarios, more stringent statistical summary exchange can be adopted to avoid the leakage of sensitive information; On resource-constrained edge devices, reporting frequency can be reduced and sparser mask maps can be used to lower communication and computing costs.

[0031] like Figure 2 As shown, this embodiment also provides an active learning system for distributed interaction of federated contrast structures, including: Several client modules are used to perform local data processing; including: The feature extraction unit is used to extract features from local unlabeled samples. 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; The local computing and communication unit is used to upload the privacy mask image representation and receive samples to be labeled for labeling and model updates; The server module is used to perform global collaborative processing, including: The Federation Space Management Unit is used to aggregate the privacy mask image representations uploaded by various clients to form a federated feature space. A contrastive learning enhancement unit is used to augment negative samples by performing contrastive learning in the federated feature space using privacy mask graph representations across clients. The sample value assessment unit is used to calculate the global diversity score and the coordination uncertainty score of the candidate samples. A multi-stage sampling decision unit is used to perform preliminary selection, fine screening and final sorting based on the scores, so as to select samples to be labeled and allocate them to the corresponding clients.

[0032] This embodiment focuses on the field of medical imaging (e.g., CT, MRI, or X-ray images), and designs a distributed privacy dataset construction system and model learner for multiple affiliated and partner hospitals of a university.

[0033] Several of these hospitals (client) Each possesses a large number of unlabeled or labeled sensitive medical image datasets. Due to strict privacy regulations and compliance requirements that prevent data from leaving the hospital, raw image data and fine-grained features cannot be shared between hospitals or with a central server. This embodiment aims to apply the method of the present invention to collaboratively select the most valuable medical images for annotation while protecting patient privacy, thereby minimizing annotation costs and rapidly training a high-performance medical diagnostic model. The specific process is as follows: 1. Construction and exchange of privacy mask images: In each hospital (client) Internally, the system extracts a high-dimensional feature vector for each image x from local medical images. Based on local features Based on the similarities between them, construct a graph structure that reflects the inherent relationships between local data. Then, regarding the diagram... Irreversible masking is applied to the node features to generate a privacy mask graph representation. (Client) Only this masked, structured distributed representation is uploaded to the central server in the federated server cluster that aggregates the local models of each client.

[0034] 2. Comparison of the structured distribution interaction (COSIN) mechanism: The central server in the federated server cluster aggregates the local models of each client. It collects the privacy masked image representations of all clients. The mechanism proposed in the patent enhances the learning of feature representations, thereby constructing a federated feature space and expanding the cross-client negative samples: Based on the local contrastive learning loss, this embodiment incorporates the masked representations from other clients as cross-client negative samples into the negative sample pool to expand the cross-client negative samples.

[0035] 3. Distributed sample value assessment and dataset construction: In the federated feature space, the system collaboratively calculates the global value of samples (including global diversity and coordinated uncertainty) to ensure that the selected samples have global representativeness and rich information. This embodiment adopts a multi-stage process that balances sampling efficiency and quality. First, an initial selection is performed to obtain candidate samples. Then, diversity screening is conducted, and the selected samples are finally ranked and chosen to select the most valuable samples. Finally, the samples are sent to hospitals for human expert annotation, and the newly annotated data is used for local model fine-tuning. Finally, the global medical model is updated through a federated strategy of privacy graph.

[0036] Performance evaluation of the above embodiments: Compared with existing federated active learning methods, this privacy-mask-based framework can achieve data collaboration while ensuring privacy. By exchanging only irreversible privacy-mask summary, it strictly adheres to the privacy requirement of "data not leaving the hospital" and selects the most valuable data for annotation, enhancing the robustness and discriminative power of medical image features. Compared with other methods, the model accuracy of this system is improved by 8.7%. Under the premise of achieving the same training effect, the number of samples required is reduced by 11.5%, resulting in a higher quality and more representative sample set. Thus, a diagnostic model with better generalization performance can be trained with a smaller annotation budget.

[0037] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. An active learning method for distributed interaction of federated contrastive structures, characterized in that, Includes the following steps: Step S1: Extract features from unlabeled local samples on each client and construct an initial graph structure that reflects the inherent relationships of local data based on the extracted features; Step S2: Perform information aggregation and irreversible masking on each initial graph structure to generate a privacy mask graph representation, and upload the privacy mask graph representation to the server, wherein the privacy mask graph representation cannot restore the original sample features; Step S3: The server aggregates privacy mask image representations from various clients to form a federated feature space; in the federated feature space, comparative learning is performed based on the privacy mask image representations from different clients; Step S4: For each candidate unlabeled sample in the federated feature space, calculate its global diversity score. With reconciliation uncertainty score ; Step S5: Initially select unlabeled candidate samples based on the coordination uncertainty score to obtain the first candidate set; further screen the samples in the first candidate set based on the global diversity score to obtain the second candidate set; finally sort the samples in the second candidate set by combining their global diversity score and coordination uncertainty score, and select a preset number of samples as unlabeled samples according to the sorting results. Step S6: Assign the samples to be labeled to their respective clients for labeling, and update the federated learning model using the newly labeled data.

2. The active learning method for distributed interaction of federated contrastive structures according to claim 1, characterized in that, In step S2, information aggregation and irreversible masking are performed on each initial graph structure to generate a privacy mask graph representation, specifically: For the initial graph structure ,in For a set of nodes, Let be a set of edges, nodes Corresponding sample features ; By using a graph attention mechanism to weighted aggregate node features, a masked structured representation is obtained. ,Right now This indicates the first step after graph attention aggregation. Feature representation of each sample: ； ； in, Represents a node With nodes The boundary rights between them It is a learnable linear transformation matrix. For activation function, Normalized attention weights; right The result of the irreversible transformation is represented as a privacy mask image.

3. The active learning method for distributed interaction of federated contrastive structures according to claim 2, characterized in that, When constructing the initial graph structure, its edge weights Based on sample feature vectors and Calculate the cosine similarity and set a threshold. Perform sparsification: ； 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; That is, based on two samples and Feature calculation yields the edge weights between the nodes corresponding to the two samples. .

4. The active learning method for distributed interaction of federated contrastive structures according to claim 2, characterized in that, In step S3, the comparative learning includes: Construct a contrastive loss function that includes not only negative samples from the same client, but also privacy mask map representations from other clients as cross-client negative samples; The contrastive loss function is expressed as follows: ； ； in, , For the sample , The projection representation, Represents cosine similarity. For temperature parameters, For the current client index, Indicates that it comes from the client. The privacy mask image represents the set. For cross-client negative sample items, This represents the features of cross-client negative samples from the client.

5. The active learning method for distributed interaction of federated contrastive structures according to claim 1, characterized in that, In step S4, the global diversity score is calculated. include: For unlabeled samples Calculate its characteristics in the highly dispersed feature space. The representation in Global diversity score : ； in, Indicates unlabeled samples Features This represents the average of the first 5 values. For distance measurement, the first term express To the federal feature space The distance to the nearest neighbor sample in the middle; the second term express The average distance to its five nearest neighbors.

6. The active learning method for distributed interaction of federated contrastive structures according to claim 1, characterized in that, In step S4, the consistency uncertainty score is calculated. include: Based on the target model, the samples Predicted probability vector Calculate the fundamental uncertainty value : ； 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 prediction uncertainty value. 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 The local feature extraction unit, which is attached to federated contrastive learning, takes as input the hidden layer network features of the feature extraction unit and an uncertainty indicator. The structure is a multilayer perceptron model with the same number of layers as the feature extraction unit, each layer being a fully connected neural network, and the last layer being a single-layer regression network compressed from 1024 dimensions to 1 dimension and an activation function layer, with the output being a floating-point number greater than 0. The coordination uncertainty score This is obtained by fusing the outputs of the global uncertainty indicator and the local uncertainty indicators of each client: ； in, Total number of clients This indicates that a trainable local uncertainty indicator is available. This indicates an unlabeled sample.

7. The active learning method for distributed interaction of federated contrastive structures according to claim 6, characterized in that, The uncertainty indicator network Training is performed using a pairwise ranking loss, where the ranking loss is: ； in, , Given a pair of training samples that satisfy the truth relation , It is a marginal hyperparameter greater than 0.

8. The active learning method for distributed interaction of federated contrastive structures according to claim 1, characterized in that, Step S5 includes the following sub-steps: Step S51: Each client calculates the coordinated uncertainty score for all samples in its local unlabeled sample pool. The score will be reported to the server. Step S52: The server, according to... Sort all candidate samples and select the top-ranked ones. The samples constitute the first candidate set, wherein 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 Sort the samples in the set and select the top-ranked ones. The samples constitute the second candidate set, wherein ; Step S54: For each sample in the second candidate set, calculate its comprehensive score. ,in and These are weighting coefficients, obtained through statistical analysis of federal big data. Step S55: Based on the overall score Sort the samples in the second candidate set in descending order and select the top-ranked samples. The samples were used as the final samples to be labeled.

9. An active learning system for distributed interaction of federated contrastive structures, used to implement the active learning method for distributed interaction of federated contrastive structures as described in any one of claims 1-8, characterized in that, include: Several client modules are used to perform local data processing; including: The feature extraction unit is used to extract features from local unlabeled samples. 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; The local computing and communication unit is used to upload the privacy mask image representation and receive samples to be labeled for labeling and model updates; The server module is used to perform global collaborative processing, including: The Federation Space Management Unit is used to aggregate the privacy mask image representations uploaded by various clients to form a federated feature space. A contrastive learning enhancement unit is used to augment negative samples by performing contrastive learning in the federated feature space using privacy mask graph representations across clients. The sample value assessment unit is used to calculate the global diversity score and the coordination uncertainty score of the candidate samples. A multi-stage sampling decision unit is used to perform preliminary selection, fine screening and final sorting based on the scores, so as to select samples to be labeled and allocate them to the corresponding clients.

Images