Federal learning-based privacy medical analysis method and system

Abstract

The invention relates to the technical field of data analysis and privacy protection, in particular to a privacy medical analysis method and system based on federated learning. The method comprises the steps that unified semantic modeling and ontology mapping are conducted on medical institution tasks through a server side, a multi-layer semantic ontology is constructed, and a mapping matrix is generated; the client performs semantic weighting by using the alignment features and confidence, constructs a local cross-domain feature association matrix, performs hierarchical privacy coding, generates abstract vectors and uploads the abstract vectors to the server; the server aggregates local model parameters based on abstract weights, optimizes a global model in combination with multilayer semantic consistency loss, and dynamically adjusts aggregation weights according to prediction errors; a dynamic self-adaptive fine tuning mechanism and a strategy generation mechanism are introduced, cross-domain migration and strategy optimization of the model are realized, and a closed-loop dynamic evolution system of the privacy strategy model is formed through a privacy constraint matrix and feedback credibility. According to the method, the accuracy, privacy and cross-mechanism adaptability of medical data analysis are improved.

Description

Technical Field

[0001] This invention relates to the fields of data analysis and privacy protection technology, specifically to a privacy-preserving medical analysis method and system based on federated learning. Background Technology

[0002] With the rapid development of medical informatization and artificial intelligence technologies, collaborative analysis of multi-source heterogeneous medical data has become a key way to improve the accuracy and efficiency of diagnosis and treatment. However, medical data is highly sensitive and requires privacy protection. Traditional centralized data processing models face challenges in data security and compliance. Federated learning, as a distributed machine learning paradigm, can achieve multi-party collaborative modeling without the data leaving the local machine.

[0003] CN117033797A discloses an information retrieval method, apparatus, device, and storage medium based on federated learning. The method includes: obtaining local model weight parameters by performing local updates on at least one client device; aggregating each local model weight parameter according to an aggregation algorithm to obtain server-side global model parameters; extracting semantics from the text context using the server-side global model to obtain context feature vectors; calculating the similarity between the vectors of the query text and the retrieved text using a pre-built encoder-decoder framework and the text context feature vectors; and displaying retrieval results associated with the query text from the set of retrieved texts based on the similarity between the vectors of each query text and each retrieved text, thereby retrieving the query text.

[0004] However, the application of federated learning technology in the medical field still faces technical challenges such as difficulties in cross-institutional semantic alignment, difficulty in balancing privacy protection and model performance, insufficient adaptability to dynamic environments, and lack of reliable policy coordination mechanisms. Therefore, it is urgent to build a federated learning scheme that can deeply integrate semantic alignment, hierarchical privacy protection, and dynamic policy optimization to support safe and efficient medical data analysis and application. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a privacy-preserving medical analysis method and system based on federated learning.

[0006] The technical solution of this invention: a privacy-preserving medical analysis method based on federated learning, comprising the following specific implementation steps: S1. The server performs unified semantic modeling and ontology mapping on the tasks submitted by several medical institutions, constructs a multi-layer semantic ontology and generates a mapping matrix, and embeds the features of each medical institution into a unified semantic space for feature alignment. S2. Each medical institution's client uses the aligned feature vectors and their confidence levels to perform semantic weighting, constructs a local cross-domain feature association matrix, performs hierarchical privacy coding according to the feature sensitivity level, and generates a local feature association summary vector which is then uploaded to the server. S3. The server receives the feature association summary vector uploaded by the client, aggregates the local model parameters based on the summary weight, optimizes the global model by combining multi-layer semantic consistency loss, and dynamically adjusts the aggregation weight according to the client's prediction error. S4. Based on the global aggregation model, a dynamic adaptive fine-tuning mechanism is introduced to perform multi-level parameter correction by combining the task deviations and feature differences of each client, and to dynamically generate differentiated optimization strategies for cross-domain model migration and strategy optimization. S5. Construct a dynamic privacy constraint matrix on the server side to constrain and optimize the global policy. The client executes the policy and generates trusted feedback. The server adjusts the fine-tuning weights in reverse based on the trust level of the feedback, forming a closed-loop dynamic evolution system for the privacy policy model.

[0007] Preferably, the process of unified semantic modeling and ontology mapping includes: The server receives task request descriptions from various medical institutions, extracts semantic tags through a natural language parsing model, and obtains a set of target concepts. The semantic similarity between concepts of each task is calculated using a concept similarity metric function, and a unified set of concepts is output. Construct a multi-layer semantic ontology containing a set of relationships between concepts and hierarchical labels, and establish an ontology mapping matrix based on concept similarity and relationship structure similarity; The mapping matrix is ​​used to map each feature in the original feature set submitted by each medical institution into a semantic embedding vector, and project it onto a unified semantic space to calculate the alignment confidence of each feature. Assess the consistency of tasks across institutions within a unified semantic space, calculate semantic biases, and provide dynamic feedback and adjustments to address semantic differences.

[0008] Preferably, the implementation process of hierarchical privacy coding includes: The client weights the aligned feature vectors based on the alignment confidence of the features to obtain the weighted enhanced feature vectors; A local cross-domain feature association matrix is ​​constructed using the enhanced feature vectors, and the elements of this matrix represent the association strength between features. Based on the sensitivity level of the features, a hierarchical privacy coding function is used to perform privacy-constrained coding on the enhanced feature vector and feature correlation matrix; Based on the privacy-encoded feature association matrix, a summary vector containing only the core information of cross-domain association is generated, and the summary vector is uploaded to the server. At the same time, the privacy-encoded complete feature vector and association matrix are stored locally.

[0009] Preferably, the process of optimizing the global model includes: The server constructs a global feature association graph based on the feature association summary vectors uploaded by all clients; Based on the correlation strength of features in the summary vector, the local model parameters on the client are weighted and aggregated to obtain the preliminary aggregated global model parameters; A multi-layer semantic consistency loss function is introduced to iteratively optimize the initial aggregation model; For highly sensitive features of the client, the server only uses the feature association summary information to guide the local model parameter update of the client; The prediction error of each client is evaluated on the global model, and the aggregation weight of each client in the next round of aggregation is dynamically adjusted based on the evaluation results.

[0010] The preferred dynamic adaptive fine-tuning mechanism includes the following steps: The server calculates a dynamic adaptive adjustment coefficient based on the feature association summary of each client and the global error response, and distributes this coefficient along with the global model parameters to each client. The client computes the local semantic consistency loss locally using its non-sensitive feature set and updates the model parameters with an adaptive learning rate that incorporates a dynamically adaptive adjustment coefficient. For the defined sensitive features, an indirect projection update mechanism based on feature summarization is used to adjust the model parameters; The server aggregates the local model outputs and feature decision mappings from each client to form a cross-domain strategy generation graph, and generates preliminary cross-domain strategies based on this graph through a constrained optimization function.

[0011] Preferably, the dynamic adaptive fine-tuning mechanism also includes a strategy optimization step, specifically: The server calculates the deviation between the initial cross-domain strategy and the global task objective; If the deviation exceeds the preset tolerance threshold, the dynamic adaptive adjustment coefficient of each client will be dynamically adjusted according to the deviation. After multiple iterations, the global strategy converges to the optimal solution that satisfies the common task objectives of multiple organizations; The converged global policy is distributed to each client, and the client updates its own decision-making logic and task scheduling according to the policy.

[0012] Preferably, the formation process of a closed-loop dynamic evolution system includes: The server constructs a privacy constraint matrix based on the client's sensitive data feature set and the institution's privacy weight coefficients, and generates dynamic privacy adjustment coefficients. By combining the global strategy with the dynamic privacy adjustment coefficient, a constraint optimization is performed to obtain an optimized strategy that meets privacy protection requirements. When the client executes the optimization strategy, it generates reliable execution feedback that includes the execution effect, privacy compliance, and task completion quality. The server constructs a feedback trust matrix based on trusted execution feedback and adjusts the client's fine-tuning weights according to this trust matrix; The server iteratively optimizes the strategy based on feedback trust and privacy constraints, and feeds the converged strategy and fine-tuned weights back to the global model for updating.

[0013] Preferably, the formation process of the closed-loop dynamic evolution system also includes: The server constructs a privacy constraint matrix based on the client's sensitive data feature set and the institution's privacy weight coefficients; Dynamic privacy adjustment coefficients are generated based on the privacy constraint matrix; The global policy is constrained and optimized using a dynamic privacy adjustment coefficient.

[0014] Preferably, the concept similarity measurement function measures semantic similarity by calculating the number of intersections of the attribute sets of two concepts; The elements in the ontology mapping matrix are calculated by weighting the concept similarity score and the structural similarity of the relationship between concepts.

[0015] The technical solution of this invention: A privacy-preserving medical analysis system based on federated learning, which is used to execute the aforementioned privacy-preserving medical analysis method based on federated learning, comprising: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement the aforementioned method for privacy-based medical analysis using federated learning.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a privacy-preserving medical analysis method and system based on federated learning. By introducing unified semantic modeling and a multi-layer ontology mapping mechanism, it effectively solves the semantic alignment problem of multi-source heterogeneous medical data in federated learning, significantly improving the accuracy and consistency of cross-institutional feature interactions. Combining a confidence-weighted feature enhancement method and hierarchical privacy coding technology, it achieves secure abstraction of local features and cross-domain association mining while fully protecting sensitive patient information, enhancing the global model's ability to balance medical data privacy and usability. Through dynamic weight adjustment, semantic consistency constraints, and feedback-driven strategy optimization, the system can adaptively optimize the aggregation process, improving the model's generalization performance and convergence efficiency in complex medical scenarios. The resulting privacy-policy-model closed-loop evolution system not only ensures data compliance and security but also improves the accuracy of medical diagnosis and system reliability. Attached Figure Description

[0017] Figure 1 This is a flowchart of a privacy-focused medical analysis method based on federated learning proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes a privacy-preserving medical analysis method based on federated learning, which includes the following specific implementation steps: S1. The server performs unified semantic modeling and ontology mapping on the tasks submitted by various medical institutions. It extracts core semantic concepts from the task objectives and calculates their similarity to form a unified concept set. A multi-layered semantic ontology is constructed, and a mapping matrix is ​​generated to describe the hierarchy and relationships between concepts. Based on this matrix, the features of each institution are embedded into the unified semantic space to achieve feature alignment. The task semantic consistency deviation is calculated and feedback corrections are made to ensure global semantic convergence. The specific implementation process is as follows: S11. The server first receives task request descriptions T from multiple medical institutions. i Each task request description T i It includes local diagnostic targets, feature fields, and constraints; The server uses a natural language parsing model to describe the task request T. i Semantic label extraction is performed to obtain the target concept set. ; Using concept similarity measurement functions: ; Calculate the semantic similarity between the concepts of each task; Among them, c ij Represents the j-th task concept of organization i (e.g., "abnormal blood glucose level" or "imaging features of lung nodules"); c klRepresents the l-th task concept of organization k; A(c ij ) represents concept c ij The set of attributes, including but not limited to concept-related fields, units of measurement, and clinical labels; A(c kl ) represents concept c kl The set of attributes; It represents the number of intersections of sets of concept attributes, used to measure the similarity between two concepts; This represents the concept similarity score, with a value range of [0,1]. A higher score indicates that the two concepts are more similar in attributes. Based on this: Output a unified set of concepts. This provides a constraint benchmark for subsequent feature semantic mapping; S12. Obtaining a unified set of concepts Then, the server constructs a multi-layered semantic ontology. ; Where R represents the set of relationships between concepts (such as "cause and effect" or "location-symptom"); L represents the hierarchical label (such as "macropathological layer" or "cellular imaging layer"). The server establishes a mapping matrix M based on the inter-layer constraints of the ontology, defined as follows: ; Among them, M pq Represents the elements of the ontology mapping matrix, representing the concept c. p With c q Comprehensive similarity in semantic and relational spaces; c p and c q To unify the set of concepts Two concepts in; This represents the concept similarity score; It represents the structural similarity between concepts, such as the degree of matching between two concepts in terms of ontology level, causal relationship, and dependency relationship, with a value range of [0,1]. Indicates the weighting coefficient. ∈[0,1], used to balance concept similarity and relation similarity; S13. The server performs embedding projection on the feature descriptions submitted by each institution based on the mapping matrix M in the semantic ontology O, specifically as follows: Obtain the original feature sets submitted by each institution. , for each original feature x ij Mapped to semantic embedding vector The following methods can be used: If the field is text / encoded: use TF-IDF; If the field is numeric / scalar: use one-hot encoding with embedding; The embedding vectors are projected onto the unified semantic space based on the ontology mapping matrix M: ; Where W is the projection matrix, obtained through the following optimization: ; Where, x ij M represents the j-th original feature field of the i-th institution; M represents the feature-concept semantic mapping matrix, M=[M jk ];f(x ij ) represents the aligned feature vector after projection; Representing concept c k The vector form; m is the number of features for each organization; The projected feature vector f(x) ij Normalize (e.g., unit vectorization) and calculate alignment confidence: ; Among them, Conf(x) ij () indicates alignment confidence; Representation and feature x ij The concept of the best match ; Based on this, output the aligned feature vector set. Alignment confidence Conf(x) for each feature ij ); S14. Assess the consistency of tasks across institutions within a unified semantic space, and provide dynamic feedback and adjustments to semantic differences to ensure that cross-institutional task definitions tend to be semantically unified, specifically as follows: The set of projected feature vectors for each institution With global average vector Compare and calculate semantic bias : ; right The server generates semantic feedback suggestions for the organization, recommending: adjusting the mapping relationship between features and concepts (such as re-annotation or fine-tuning the embedding); or adding weight adjustments in the next round of feature projection to enhance the alignment effect; After receiving feedback from various institutions, the updates were iteratively revised until... Alternatively, feature projection may be performed again once the maximum number of iterations has been reached.

[0019] S2. Each medical institution uses aligned feature vectors and their confidence levels to perform semantic weighting, enhancing the importance of high-confidence features while suppressing low-confidence features. This constructs a local cross-domain feature association matrix, uncovers potential dependencies between features, performs hierarchical privacy coding based on feature sensitivity levels, and generates local feature association summary vectors. These summary vectors and low-sensitivity features are then securely uploaded to the server to initialize local federated training model parameters, achieving secure cross-domain feature propagation and collaborative modeling. The specific implementation process is as follows: S21. Each client will align the feature vectors. Based on alignment confidence Conf(x) ij Weighting: ; in, This represents the weighted and enhanced feature vector; S22, The client utilizes the enhanced features Construct a local cross-domain feature association matrix A i : ; Among them, A i (p,q) represents the feature x ip With x iq The strength of the association; S23, The client will send the feature vector With the correlation matrix A i Perform hierarchical encryption mapping: ; ; in, This represents the feature vector after the privacy constraints are encoded. This represents the feature correlation matrix after privacy constraint encoding; L represents a hierarchical privacy coding function; priv Indicates the sensitivity level of the feature; S24. Based on the privacy-encoded feature association matrix Generate a summary vector S i Only the core information related to cross-domain relationships is retained, and no original sensitive data is included; The client stores the privacy-encoded complete feature vector locally. and complete association matrix Used for local model fine-tuning; can regenerate summaries in the next federated training iteration, improving global model adaptability; provides traceable references for anomalous data or semantic drift; The summary vector S i Uploaded to the server, it participates in global aggregation, providing cross-institutional feature association information; S25, The client utilizes local feature vectors and related summary S i Initialize the local federated training model parameters and output the client's local model initialization parameters. .

[0020] S3. The server receives the low-sensitivity feature vectors and feature association summaries uploaded by the client, constructs a global feature association graph, and performs preliminary aggregation of local model parameters based on the summary weights. The aggregated model is iteratively optimized through multi-layer semantic consistency loss to ensure cross-domain feature semantic alignment. High-sensitivity features are locally preserved and optimized. At the same time, the client aggregation weights are adaptively adjusted in combination with the prediction error to form a global closed-loop model output. The specific implementation process is as follows: S31. The server receives all low-sensitivity feature vectors uploaded by clients. Feature-related summary S i According to the summary vector S i Construct a global feature association graph G=(V,E): Node V={ } indicates low-sensitivity characteristics of each client; Edge E={S i (p,q)} represents the strength of potential cross-domain associations between features; For client-side local model parameters Perform weighted aggregation to obtain preliminary aggregated global model parameters. : ; Where N represents the total number of client organizations; m represents the number of features per client; This represents the association strength of the p-th and q-th features in the feature association summary of client i, derived from cross-domain feature association modeling: the client calculates the correlation matrix between features using local data, and after privacy constraints and summary processing, generates an uploadable summary vector. That is, the value at the corresponding position in the matrix; w i This represents the aggregate weight of client i; S32, Preliminary Aggregation Model Perform multi-level semantic constraint optimization: ; Where, f(x) ij ) represents the client's original feature vector; L represents the feature vector output by the global model. sem This represents the loss of global semantic consistency. Multi-level iterative updates are employed. ; in, Represents the global model parameters for the k-th iteration; Indicates the learning rate; Represents the global model parameters for the (k+1)th iteration; L represents the global semantic consistency loss. sem Gradient with respect to model parameters; S33. During the optimization process, for highly sensitive features, the server cannot directly access the original values. Instead, it guides local updates only through summary information, thereby protecting sensitive information while still allowing the global model to indirectly optimize the contribution of highly sensitive features. ; in, Indicates the local model parameters of client i; This represents the local parameter update guided by the summary, i.e., the model parameter increment vector; Indicates the privacy control factor; It should be noted that highly sensitive features refer to feature types that may directly expose personal privacy or core medical information, including but not limited to the following in medical scenarios: Personal identification information: name, ID number, contact information; Core information in clinical medical records: disease type, genetic information, and medical records; Biometric data: gene sequence fragments, raw electrocardiogram waveform data, and raw image data; Other data that laws or institutions require to be protected: medical insurance numbers, hospitalization numbers, sensitive experimental data, etc. S34. The server evaluates the prediction errors of each client on the global model: ; Dynamically adjust client aggregation weights based on errors: ; The next round of aggregation will use dynamically updated weights; Among them, E i This represents the prediction error of client i; This represents the j-th low-sensitivity feature vector of client i; Indicates the aggregate weight of client i in the next round; S35. After completing multiple iterations, the final global model is generated. The global model is distributed to each client.

[0021] S4. By introducing a dynamic adaptive fine-tuning mechanism on the basis of the global aggregation model, multi-level parameter correction is performed based on the task bias and feature differences of each client, and the self-adjustment of the model's cross-domain transfer capability is achieved based on gradient residuals and meta-learning optimization. At the same time, the strategy generation module dynamically generates differentiated optimization strategies based on domain performance feedback, enabling the model to achieve stable generalization and fast convergence in multi-source heterogeneous medical data environments, thereby improving the adaptability and accuracy of the global model to medical analysis tasks in different scenarios. The specific implementation process is as follows: S41. The server first processes the global model parameters. Structural analysis and fine-tuning initialization are performed to enable personalized adaptation for different medical institutions (clients). Specifically, the server associates a summary S based on the characteristics of each client. i With global error response E i Calculate the dynamic adaptive adjustment coefficient The global contribution of the client is coupled with the error feedback as a prior weight for personalized fine-tuning. That is, the model's fitness is enhanced on highly correlated and low-error mechanisms, while stability is maintained on low-correlation and high-error mechanisms. ; The server will then transfer the global model parameters. and dynamic adaptive adjustment coefficient Distribute to each client; S42, Any client receives global model parameters and dynamic adaptive adjustment coefficient Then, a personalized fine-tuning process is performed in the local environment, specifically as follows: The client utilizes its own set of non-sensitive features Calculate the local semantic consistency loss: ; and with adaptive learning rate Update parameters: ; in, N represents the local semantic consistency loss of client i; i f(x) represents the number of low-sensitivity features possessed by client i; ij ) represents the client-side local feature mapping vector; This represents the feature vector corresponding to the global model; Represents the local loss gradient; Indicates the adaptive learning rate; This represents the model parameters after local fine-tuning on client i; This represents the base learning rate, which is the preset global fine-tuning rate and can be adjusted according to the task. Meanwhile, for sensitive medical feature data (including but not limited to patient privacy indicators and disease gene fragments), an indirect projection update mechanism based on feature summaries is adopted: ; in, This indicates the amount of parameter adjustment guided by the highly sensitive feature summary; This represents the projection coefficient of the high-sensitivity feature summary, which controls the indirect fine-tuning of sensitive features and ensures privacy protection. S43. After the client completes adaptive fine-tuning, the server aggregates the local model outputs and feature decision mappings from each client to form a cross-domain strategy generation graph G. decision =(V d E d ); node This indicates the key feature nodes identified by each client in the local task, such as "abnormal indicator threshold" and "pathological pattern center point". Side E d The cross-domain collaborative relationship between nodes is determined by both feature relevance and task dependency. ; Among them, E d (u,v) represents the cross-domain feature collaborative edge weight; Let u and v represent the local key feature decision sets, respectively. Indicates global semantic confidence; Based on this, the server generates a cross-domain policy using a constraint optimization function: ; Among them, S cross This indicates an initial cross-domain strategy; R(S) represents the feature mapping function of the local model; R(S) represents the policy regularization term. V represents the policy regularization coefficient; G Indicate that the decision graph G decision Embedded vector space; It represents the policy's feasible domain, limiting the selection, weight, or execution threshold of policy nodes; S44, Server-side cross-domain computing policy S cross With global mission objective T global deviation : ; If the deviation exceeds the set tolerance threshold, the adjustment coefficient for each client will be dynamically adjusted based on the feedback deviation. ; in, This indicates that the client is fine-tuning the weights dynamically. Indicates the strategy sensitivity control parameters; This represents the fine-tuning coefficient for client i in the current round; The norm representing the strategy bias is taken as the L2 norm in this embodiment; Based on this, cross-round adaptive optimization is completed, that is: policy bias drives the model to adjust its intensity, and after multiple iterations, the global policy is optimized. It converges to the global optimum, satisfying the common mission objectives of multiple institutions; S45, When Global Strategy Once the convergence criteria are met, the strategy is distributed to each client execution end. The client updates its own decision-making logic or task scheduling module according to the strategy content (e.g., adjusting the diagnostic threshold or updating the disease risk classification rules). The client continuously generates policy execution feedback R during execution. i This includes, but is not limited to, performance evaluation values, strategy adaptability indicators, and characteristics of newly emerging abnormal samples, which are periodically uploaded to the server. The server then adjusts the model parameters and strategy templates based on the feedback data.

[0022] S5. By constructing a dynamic privacy constraint matrix on the server side, adaptive adjustment coefficients are generated based on client-side sensitive features and institutional weights to constrain and optimize the global policy and perform compatibility checks, forming a feasible privacy protection policy. The client executes the policy and generates trusted feedback. The server quantifies the trust level of the feedback based on the execution effect, task completion, and privacy compliance, and adjusts the fine-tuning weights in reverse to achieve iterative optimization of the policy and update of the global model, forming a closed-loop dynamic evolution system of privacy-policy-model. The specific implementation process is as follows: S51. The server constructs a privacy constraint matrix P based on the client's sensitive data feature set. i : ; in, Representation: The element in the p-th row and q-th column of the privacy constraint matrix of client i represents the feature. and A quantitative value for the potential information leakage risk between them; This represents the privacy weight coefficient of the organization to which client i belongs, reflecting the level of data sensitivity and compliance requirements of that organization; This represents a feature similarity metric function used to calculate features. and The degree of similarity; and These represent the p-th and q-th sensitive feature vectors in client i, respectively; Generate dynamic privacy adjustment coefficient : The smaller the value, the higher the privacy risk, and the more constrained the impact on fine-tuning or policy modifications. S52, Generate the global policy With dynamic privacy coefficient Combined, a constrained optimization objective is formed: ; in, Let i represent the set of sensitive features for client i, used for constraint optimization; This represents the penalty term for policy S deviating from privacy requirements on feature p; This represents the coefficient representing the trade-off between privacy and task loss. The task loss function represents the execution of the strategy; This represents the optimization strategy for client i under privacy constraints; like If the privacy threshold is violated, the policy node will be adjusted or sensitive features will be blocked from participating in fine-tuning; S53, The client executes the policy. Generate trusted execution feedback in real time : in, Indicators representing the actual effectiveness of the strategy; Indicates local privacy compliance metrics; Indicates the quality indicators for task completion; The server constructs a feedback trust matrix. i : ; in, , and This represents a weighting coefficient used to balance execution reliability, task completion rate, and privacy compliance. Based on feedback and trust levels, further adjust the client-side fine-tuning weights. To achieve reliable backflow; S54, The server relies on feedback trust level. i and privacy constraints strategy Perform iterative optimization: ; After multiple iterations, the strategy converges while balancing privacy protection and optimal task execution. ; in, This represents the policy iteration learning rate, used to control the magnitude of policy updates in each round; This represents the strategy of client i after the (t+1)th iteration; This represents the policy vector of client i in the t-th iteration, which includes the decision weights and execution paths of the policy nodes; The policy gradient operator represents the gradient of the policy optimization objective function. A coefficient representing the trade-off between privacy and task optimization; This indicates a penalty for a policy that deviates from privacy constraints, and is used to ensure that policies do not violate the requirements for the protection of sensitive information; S55, Convergence Strategy With the latest fine-tuned weights Feedback is jointly sent to the global model renew: ; in, This represents the parameter values ​​after the global model update, which incorporates the fine-tuned weights and strategy optimization results from each client. The server records the history of each round of strategy, feedback, and weight fine-tuning, forming a traceable and auditable global evolution log.

[0023] Example 2: A privacy-preserving medical analysis system based on federated learning proposed in this invention is used to execute a privacy-preserving medical analysis method based on federated learning proposed in Example 1, comprising: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement a privacy-based medical analysis method based on federated learning as described in Embodiment 1 above.

[0024] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for privacy medical analysis based on federated learning, characterized in that, The implementation steps include the following: S1, the server side performs unified semantic modeling and ontology mapping on the tasks submitted by a plurality of medical institutions, constructs a multi-layer semantic ontology and generates a mapping matrix, embeds the characteristics of each medical institution into a unified semantic space for feature alignment; S2, each medical institution client uses the aligned feature vector and its confidence to perform semantic weighting, constructs a local cross-domain feature correlation matrix, performs hierarchical privacy coding according to the feature sensitivity level, and generates a local feature correlation summary vector to upload to the server; S3, the server receives the feature correlation summary vector uploaded by the client, aggregates the local model parameters based on the summary weight, optimizes the global model in combination with the multi-layer semantic consistency loss, and dynamically adjusts the aggregation weight according to the client prediction error; S4, introduce a dynamic adaptive fine-tuning mechanism based on the global aggregation model, combine the task bias and feature difference of each client for multi-layer parameter correction, and dynamically generate a differentiated optimization strategy for model cross-domain migration and strategy optimization; S5, build a dynamic privacy constraint matrix on the server side to optimize the global strategy, and the client executes the strategy and generates a trusted feedback, the server reversely adjusts the fine-tuning weight based on the feedback trust degree, forming a closed-loop dynamic evolution system of the privacy strategy model. 2.The federated learning based privacy medical analysis method of claim 1, wherein, The process of unified semantic modeling and ontology mapping includes: The server receives the task request description of each medical institution, extracts semantic tags through a natural language parsing model, and obtains a target concept set; Calculate the semantic similarity between each task concept by a concept similarity measurement function, and output a unified concept set; Build a multi-layer semantic ontology containing the relationship set and hierarchical label between concepts, and establish an ontology mapping matrix based on concept similarity and relationship structure similarity; Map each feature in the original feature set submitted by each medical institution to a semantic embedding vector using the mapping matrix, and project it to a unified semantic space to calculate the alignment confidence of each feature; Evaluate the consistency of each institution's task in the unified semantic space, calculate the semantic deviation, and dynamically feedback and adjust the semantic difference. 3.The federated learning based privacy medical analysis method of claim 2, wherein, The implementation process of hierarchical privacy coding includes: The client weights the aligned feature vector according to the alignment confidence of the feature, and obtains an enhanced feature vector; Use the enhanced feature vector to construct a local cross-domain feature correlation matrix, and the elements of the matrix represent the correlation strength between features; According to the sensitivity level of the feature, use the hierarchical privacy coding function to encode the enhanced feature vector and the feature correlation matrix; According to the privacy coded feature correlation matrix, generate a summary vector containing only the core information of cross-domain association, and upload the summary vector to the server, while storing the complete feature vector and correlation matrix after privacy coding locally. 4.The federated learning based privacy medical analysis method of claim 3, wherein, The process of optimizing the global model includes: The server constructs a global feature correlation graph according to the feature correlation summary vectors uploaded by all clients; Based on the correlation strength of the features in the summary vector, the local model parameters of the client are weighted and aggregated to obtain the preliminary aggregated global model parameters; Introduce a multi-layer semantic consistency loss function to iteratively optimize the preliminary aggregated model; For the high-sensitive features of the clients, the server only guides the local model parameter update of the client through its feature association summary information; The prediction errors of each client on the global model are evaluated, and the aggregation weights of each client in the next round of aggregation are dynamically adjusted according to the evaluation results.

5. The federated learning based private medical analysis method according to claim 4, characterized in that, The process of the dynamic adaptive fine-tuning mechanism includes: The server calculates the dynamic adaptive adjustment coefficient according to the feature association summary of each client and the global error response, and distributes the coefficient and the global model parameters to each client; The client calculates the local semantic consistency loss using its non-sensitive feature set locally, and updates the model parameters with an adaptive learning rate combined with the dynamic adaptive adjustment coefficient; For the set sensitive features, an indirect projection update mechanism based on feature summary is used to adjust the model parameters; The server aggregates the local model outputs and feature decision mappings of each client to form a cross-domain strategy generation graph, and generates a preliminary cross-domain strategy based on the graph through a constraint optimization function.

6. The federated learning based private medical analysis method according to claim 5, characterized in that, The dynamic adaptive fine-tuning mechanism also includes a strategy optimization step, which is: The server calculates the deviation between the preliminary cross-domain strategy and the global task target; If the deviation exceeds the pre-set tolerance threshold, the dynamic adaptive adjustment coefficient of each client is dynamically adjusted according to the deviation; After multiple iterations, the global strategy converges to the optimal solution that meets the multi-agency common task target; The converged global strategy is distributed to each client, and the client updates its decision logic and task scheduling according to the strategy.

7. The federated learning based private medical analysis method according to claim 6, characterized in that, The formation process of the closed-loop dynamic evolution system includes: The server constructs a privacy constraint matrix based on the sensitive data feature set of the client and the agency privacy weight coefficient, and generates a dynamic privacy adjustment coefficient; Combine the global strategy with the dynamic privacy adjustment coefficient to perform constraint optimization to obtain an optimized strategy that meets privacy protection requirements; When the client executes the optimized strategy, it generates a trusted execution feedback that includes execution effect, privacy compliance, and task completion quality; The server builds a feedback trust matrix based on the trusted execution feedback, and adjusts the fine-tuning weight of the client according to the trust matrix; The server iteratively optimizes the strategy based on the feedback trust and privacy constraints, and feeds back the converged strategy and fine-tuning weight to the global model for updating. 8.The federated learning based privacy medical analysis method of claim 7, wherein, The formation process of the closed-loop dynamic evolution system also includes: The server constructs a privacy constraint matrix based on the sensitive data feature set of the client and the agency privacy weight coefficient; Generate a dynamic privacy adjustment coefficient according to the privacy constraint matrix; Constrained optimization of the global strategy using dynamic privacy adjustment coefficients. 9.The federated learning based privacy medical analysis method of claim 2, wherein, The concept similarity measure function measures the semantic similarity of two concepts by calculating the number of intersections of their attribute sets; The elements in the ontology mapping matrix are calculated by weighting the concept similarity score and the relationship structure similarity between concepts. 10.A privacy medical analysis system based on federated learning, used to perform the privacy medical analysis method based on federated learning in any one of claims 1-9. It includes: Memory; Processor; Computer program stored in the memory and executable on the processor; The processor executes the computer program to implement the privacy medical analysis method based on federated learning according to any one of claims 1-9.

Images