Federal learning-based medical data privacy calculation method and system

Abstract

The invention relates to the technical field of federated learning, and discloses a federated learning-based medicine data privacy calculation method and system, and the method comprises the steps: carrying out the model parameter updating based on a plurality of adjacent medicine client clusters and a to-be-updated global privacy model, and obtaining a plurality of model updating parameter sets, aggregating the plurality of model update parameter sets by using a plurality of target center clients to obtain a plurality of aggregated update parameter sets, performing consensus screening on the plurality of aggregated update parameter sets based on a plurality of consensus medicine clients to obtain a plurality of qualified update parameter sets, aggregating the plurality of qualified update parameter sets to obtain a target update parameter set, and sending the target update parameter set to a server; and performing model parameter updating on the to-be-updated global privacy model by using the target updating parameter group to obtain a target global privacy model. According to the invention, the communication efficiency in the federated learning process can be improved, and the security risk caused by network delay and unreliable updating is reduced.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a method and system for calculating medical data privacy based on federated learning. Background Technology

[0002] In the medical field, data privacy and security are core challenges for cross-institutional collaborative research. Federated learning, as a privacy-preserving technology, allows medical institutions to collaboratively train models without sharing raw data. This is crucial for mining the value of valuable medical data scattered across different institutions. Therefore, developing efficient and reliable federated learning methods is of great significance for promoting scientific research innovation and clinical collaboration in the medical field.

[0003] Traditional federated learning methods typically employ a centralized parameter aggregation approach where the server directly communicates with all clients. This approach has significant drawbacks: first, direct communication between all clients and the central server results in high latency due to network heterogeneity, becoming a bottleneck for system performance; second, the lack of an effective mechanism for validating model updates uploaded by clients makes the system vulnerable to low-quality or malicious updates, threatening model security and effectiveness. Summary of the Invention

[0004] This invention provides a method and system for privacy computing of medical data based on federated learning. Its main purpose is to improve communication efficiency in the federated learning process and reduce security risks caused by network latency and unreliable updates.

[0005] To achieve the above objectives, this invention provides a method for privacy-preserving medical data based on federated learning, comprising: Receive federated learning instructions, and identify a distributed medical client set based on the federated learning instructions. The distributed medical client set includes multiple distributed medical clients. Based on the distributed medical client set, client clustering is performed to obtain multiple neighboring medical client clusters and multiple target center clients. Among them, the target center clients correspond one-to-one with the neighboring medical client clusters, and the neighboring medical client clusters contain the target center clients. Receive the global privacy model to be updated, update the model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated, and obtain multiple sets of model update parameters; Multiple target center clients are used to aggregate multiple model update parameter sets to obtain multiple aggregated update parameter sets; Multiple consensus-based medical clients are selected in a distributed medical client cluster. Based on these multiple consensus-based medical clients, consensus screening is performed on multiple aggregated update parameter groups to obtain multiple qualified update parameter groups. Multiple qualified update parameter sets are aggregated to obtain the target update parameter set. The model parameters of the global privacy model to be updated are updated using the target update parameter set to obtain the target global privacy model. Based on the target global privacy model, federated learning-based medical data privacy computation is completed.

[0006] Optionally, the client clustering based on the distributed medical client set to obtain multiple neighboring medical client clusters and multiple target center clients includes: Extract distributed medical clients sequentially from the distributed medical client set and query the distributed client address of the extracted distributed medical clients; Summarize the distributed client addresses to obtain the distributed client address set; Clustering of distributed medical client sets based on distributed client address sets yields multiple nearest-neighbor medical client clusters, where each nearest-neighbor medical client cluster includes multiple nearest-neighbor medical clients; Multiple target central clients are obtained by selecting central clients based on multiple neighboring medical client clusters.

[0007] Optionally, the selection of a central client based on multiple neighboring medical client clusters yields multiple target central clients, including: Perform the following operation on each of the multiple neighboring medical client clusters: Extract neighboring medical clients sequentially from multiple neighboring medical clients in the neighboring medical client cluster, and record the extracted neighboring medical clients as candidate center clients; The candidate center client is removed from the neighboring medical client cluster to obtain the comparison medical client set, which includes multiple comparison medical clients; The comparison medical clients are extracted sequentially from the comparison medical client set, and the communication delay of the extracted comparison medical clients and the candidate center clients is detected to obtain the client communication delay; The data size of the extracted comparative medical clients is evaluated to obtain the comparative data size index; The client latency index is obtained by adjusting the client communication latency using a comparative data size index. By summing up the client latency indices for each comparative pharmaceutical client, a set of client latency indices is obtained; The average communication latency index is obtained by averaging the client latency index set. The average communication latency index corresponding to each candidate center client is summarized to obtain the average communication latency index set; Identify the minimum communication delay index in the set of average communication delay indices, and denote the candidate center client corresponding to the minimum communication delay index as the target center client; By aggregating the target central clients corresponding to each neighboring medical client cluster, multiple target central clients are obtained.

[0008] Optionally, the step of evaluating the data size of the extracted comparative medical clients to obtain a comparative data size index includes: Obtain the extracted local dataset of the comparison client from the comparison medical client; Data statistics are performed based on the local dataset of the comparison client to obtain a set of data scale parameters, which include: number of samples, feature dimensions, and data storage size. The data size parameter group is weighted and summed to obtain the original data size index. The original data size index is then normalized to obtain the comparative data size index.

[0009] Optionally, the step of adjusting the client communication latency using a comparative data size index to obtain a client latency index includes: The client latency index is calculated using the following formula:

[0010] in, Indicates the client latency index. This indicates client communication latency. This indicates the preset average communication delay. This indicates a comparison of data size index. This represents the preset sensitivity coefficient. This indicates the preset maximum client latency.

[0011] Optionally, the step of updating model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated yields multiple sets of model update parameters, including: Multiple neighboring medical client clusters are denoted as multiple federated training client sets, and the following operations are performed on each of the multiple federated training client sets: Extract the federated training clients sequentially from the federated training client set; Based on the extracted federated training client, obtain the client's local dataset, use the client's local dataset to train the global privacy model to be updated, and obtain the model update parameter set; The model update parameter set is obtained by summing up the model update parameter sets corresponding to each federated training client; By aggregating the model update parameter sets corresponding to each neighboring medical client cluster, multiple model update parameter sets are obtained.

[0012] Optionally, the aggregation of multiple model update parameter sets using multiple target center clients to obtain multiple aggregated update parameter sets includes: For each of the multiple model update parameter sets, perform the following operation: Based on the federated training client set corresponding to the model update parameter set, the data processing center client is determined from multiple target center clients; Upload the model update parameter set to the data processing center client to obtain the aggregation center client; The training data set is obtained by acquiring the training data volume of each federated training client in the federated training client set corresponding to the model update parameter group based on the aggregation center client. Normalize the training data set to obtain a normalized data set; The statistical model updates the parameter set of the model update indicator set, and performs the following operations on each model update indicator in the model update indicator set: Based on the model update index, identify the set of parameters to be aggregated in the model update parameter group set, and perform a weighted summation of the set of parameters to be aggregated according to the normalized data volume set to obtain the aggregated update parameters. Summarize the aggregated update parameters corresponding to each model update metric to obtain the aggregated update parameter group; By aggregating the update parameter sets corresponding to each model update parameter set, multiple aggregated update parameter sets are obtained.

[0013] Optionally, the consensus screening of multiple aggregated update parameter groups based on multiple consensus medical clients to obtain multiple qualified update parameter groups includes: Extract the aggregated update parameter group sequentially from multiple aggregated update parameter groups, and record the extracted aggregated update parameter group as the parameter group to be updated for consensus. Multiple consensus-based medical clients were used to verify the data integrity of the parameter group to be updated in the consensus process, resulting in multiple consensus pass values. Based on multiple consensuses, the validity of the parameter group is determined by value pair aggregation and update, resulting in a data validity result, which is either data valid or data invalid. If the data validity result is that the data is valid, then the parameter group to be updated by consensus is recorded as the valid update parameter group; Summarize the valid update parameter sets to obtain multiple valid update parameter sets; Multiple consensus-based medical clients were used to evaluate the model update quality of multiple valid updated parameter sets, resulting in multiple model update qualities, where each model update quality corresponds one-to-one with a valid updated parameter set. Based on the update quality of multiple models and preset update quality thresholds, data qualification screening is performed on multiple valid update parameter groups to obtain multiple qualified update parameter groups.

[0014] Optionally, the step of using multiple consensus-based medical clients to evaluate the model update quality of multiple valid updated parameter sets yields multiple model update qualities, including: Perform the following operation for each of the multiple valid update parameter groups: Extract the consensus medical client sequentially from multiple consensus medical clients, and record the extracted consensus medical client as the verification medical client; Obtain the client-side local validation dataset and local global privacy model for the validated medical client; The local global privacy model is updated using an effective set of update parameters to obtain a validated global privacy model. The model validation quality is obtained by performing model validation on the global privacy model based on the client-side local validation dataset. The model validation quality for each validated medical client is summarized to obtain multiple model validation qualities. The average of these multiple model validation qualities is then calculated to obtain the model update quality. The model update quality corresponding to each valid update parameter group is summarized to obtain multiple model update quality values.

[0015] To achieve the above objectives, the present invention also provides a medical data privacy computing system based on federated learning, comprising: The learning instruction receiving module is used to receive federated learning instructions and identify a distributed medical client set based on the federated learning instructions. The distributed medical client set includes multiple distributed medical clients. The pharmaceutical client clustering module is used to cluster clients based on a distributed pharmaceutical client set, resulting in multiple nearest-neighbor pharmaceutical client clusters and multiple target center clients. Each target center client corresponds one-to-one with a nearest-neighbor pharmaceutical client cluster, and each nearest-neighbor pharmaceutical client cluster contains a target center client. The parameter aggregation module is used to receive the global privacy model to be updated, update the model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated, and obtain multiple sets of model update parameters. Multiple target center clients are used to aggregate the multiple sets of model update parameters to obtain multiple aggregated update parameter sets. The global model update module is used to select multiple consensus medical clients in the distributed medical client set, perform consensus screening on multiple aggregated update parameter groups based on multiple consensus medical clients to obtain multiple qualified update parameter groups, aggregate multiple qualified update parameter groups to obtain target update parameter groups, and use target update parameter groups to update the model parameters of the global privacy model to be updated to obtain the target global privacy model.

[0016] To address the above problems, the present invention also provides an electronic device, the electronic device comprising: Memory, storing at least one instruction; The processor executes the instructions stored in the memory to implement the above-described federated learning-based medical data privacy computing method.

[0017] To address the aforementioned issues, the present invention also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the aforementioned federated learning-based medical data privacy computation method.

[0018] To address the problems described in the background, this invention first performs client clustering based on a distributed set of medical clients, resulting in multiple clusters of neighboring medical clients and multiple target center clients. This step selects target center clients by clustering clients based on geographical location and network topology, combined with a dynamic evaluation algorithm considering communication latency and data scale. This effectively optimizes communication efficiency in federated learning. Compared to traditional methods that only consider geographical proximity, it more accurately reflects the actual data transmission burden, thereby reducing overall communication latency and improving system performance. Next, multiple target center clients are used to aggregate multiple model update parameter sets, resulting in multiple aggregated update parameter sets. This step utilizes target center clients for hierarchical aggregation and adjusts the data scale accordingly. The weighted parameter aggregation strategy reduces the burden on the server to directly handle all client updates. Compared to centralized aggregation methods, it reduces communication path complexity and improves aggregation fairness, thereby enhancing the scalability and efficiency of federated learning. Finally, multiple consensus-based medical clients are selected from a distributed medical client pool. Based on these clients, consensus filtering is performed on multiple aggregated update parameter groups to obtain multiple qualified update parameter groups. This step introduces a decentralized consensus mechanism, using multiple consensus-based medical clients to verify data integrity and evaluate model quality of the aggregation results. This effectively enhances the privacy and security of federated learning. Compared to traditional methods lacking verification, it can identify and filter low-quality or malicious updates, ensuring the reliability of model updates. Therefore, this invention can improve communication efficiency in the federated learning process and reduce security risks caused by network latency and unreliable updates. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a federated learning-based medical data privacy computation method according to an embodiment of the present invention; Figure 2 A functional block diagram of a medical data privacy computing system based on federated learning provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device that implements the federated learning-based medical data privacy computing method according to an embodiment of the present invention.

[0020] Explanation of reference numerals in the attached figures: 10. Electronic device; 11. Processor; 12. Memory; 13. Bus.

[0021] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] This application provides a method for privacy-preserving medical data computation based on federated learning. The executing entity of this method includes, but is not limited to, at least one electronic device that can be configured to execute the method provided in this application, such as a server or a terminal. In other words, the method can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0024] Reference Figure 1 The diagram shown is a flowchart illustrating a federated learning-based method for medical data privacy computation according to an embodiment of the present invention. In this embodiment, the federated learning-based method for medical data privacy computation includes: S1. Receive federated learning instructions and identify a distributed medical client set based on the federated learning instructions. The distributed medical client set includes multiple distributed medical clients.

[0025] It is clear that the federated learning instruction refers to a human-initiated instruction used to start a cross-institutional collaborative medical modeling task (i.e., to execute federated learning). The distributed medical client set refers to a collection of multiple distributed medical clients, wherein distributed medical clients refer to terminals or servers located in different physical locations, possessing local medical data and participating in the federated learning task, such as: a local data server of a tertiary hospital, an experimental data terminal of a pharmaceutical research and development center, etc.

[0026] S2. Based on the distributed medical client set, perform client clustering to obtain multiple neighboring medical client clusters and multiple target center clients. Among them, the target center clients correspond one-to-one with the neighboring medical client clusters, and the neighboring medical client clusters contain the target center clients.

[0027] Understandably, the "nearest neighbor medical client cluster" refers to a set of multiple distributed medical clients that have similar geographical locations after client clustering. The "target center client" refers to the distributed medical client with the best communication efficiency within each near neighbor medical client cluster, determined by comprehensively calculating its communication latency with other distributed medical clients in the cluster and the data size of the other party. The specific selection method for this target center client will be given later.

[0028] Specifically, the client clustering based on the distributed medical client set yields multiple neighboring medical client clusters and multiple target center clients, including: Extract distributed medical clients sequentially from the distributed medical client set and query the distributed client address of the extracted distributed medical clients; Summarize the distributed client addresses to obtain the distributed client address set; Clustering of distributed medical client sets based on distributed client address sets yields multiple nearest-neighbor medical client clusters, where each nearest-neighbor medical client cluster includes multiple nearest-neighbor medical clients; Multiple target central clients are obtained by selecting central clients based on multiple neighboring medical client clusters.

[0029] It should be explained that the distributed client address refers to the actual address location of the distributed medical client, which can be represented by geographical coordinates (such as latitude and longitude) or network IP address. The clustering of the distributed medical client set based on the distributed client address set refers to: according to the geographical or network topological proximity relationships reflected by the distributed client addresses, using a clustering algorithm (such as the K-means algorithm) to divide the entire distributed medical client set into several groups, where each group is a nearest-neighbor medical client cluster. The above clustering is existing technology and will not be elaborated further here. For example, when using the K-means algorithm to perform the above client clustering operation, when selecting the K-means algorithm, the Euclidean distance can be calculated based on the latitude and longitude coordinates of the distributed client addresses, and the distributed medical clients can be divided into K nearest-neighbor medical client clusters through iterative optimization.

[0030] Specifically, the selection of a central client based on multiple neighboring medical client clusters yields multiple target central clients, including: Perform the following operation on each of the multiple neighboring medical client clusters: Extract neighboring medical clients sequentially from multiple neighboring medical clients in the neighboring medical client cluster, and record the extracted neighboring medical clients as candidate center clients; The candidate center client is removed from the neighboring medical client cluster to obtain the comparison medical client set, which includes multiple comparison medical clients; The comparison medical clients are extracted sequentially from the comparison medical client set, and the communication delay of the extracted comparison medical clients and the candidate center clients is detected to obtain the client communication delay; The data size of the extracted comparative medical clients is evaluated to obtain the comparative data size index; The client latency index is obtained by adjusting the client communication latency using a comparative data size index. By summing up the client latency indices for each comparative pharmaceutical client, a set of client latency indices is obtained; The average communication latency index is obtained by averaging the client latency index set. The average communication latency index corresponding to each candidate center client is summarized to obtain the average communication latency index set; Identify the minimum communication delay index in the set of average communication delay indices, and denote the candidate center client corresponding to the minimum communication delay index as the target center client; By aggregating the target central clients corresponding to each neighboring medical client cluster, multiple target central clients are obtained.

[0031] It should be explained that the "comparison pharmaceutical client set" refers to the cluster of neighboring pharmaceutical clients after removing candidate center clients, where the comparison pharmaceutical clients refer to the distributed pharmaceutical clients within the cluster of neighboring pharmaceutical clients after removing candidate center clients. The "client communication latency" refers to the communication latency between the comparison pharmaceutical client and the candidate center client, which can be expressed in milliseconds (ms) or round-trip time (RTT). The "communication latency detection" of the extracted comparison pharmaceutical clients and the candidate center clients refers to quantifying the network transmission latency between them by sending preset test data packets and measuring the response time. The "comparison data size index" refers to a value quantifying the data size of the comparison pharmaceutical client; the larger the comparison data size index, the larger the data size of the comparison pharmaceutical client. The "client latency index" refers to a value quantifying the communication efficiency after considering the combined effects of communication latency and data size; the larger the client latency index, the heavier the communication burden and the lower the communication efficiency between the candidate center client and the comparison pharmaceutical client. The "average communication latency index" refers to the average of all client latency indices in the client latency index set. The "minimum communication latency index" refers to the average communication latency index with the smallest value in the average communication latency index set.

[0032] Furthermore, since frequent data transmission is required between the distributed medical clients in the distributed medical client set during the subsequent federated learning model training and parameter aggregation process, this scheme first divides the distributed medical client set into multiple neighboring medical client clusters based on geographical location. This allows distributed medical clients within the same neighboring medical client cluster to perform prioritized, low-latency transmission, thereby reducing the decrease in data transmission efficiency caused by cross-regional communication. Next, a target central client is selected from the neighboring medical client clusters, making this target central client the endpoint for data transmission from other distributed medical clients in the cluster, thus reducing the complexity of the communication path and the overall communication latency.

[0033] It should be explained that, since the differences in data size among different distributed medical clients affect transmission time, this scheme introduces the aforementioned client communication latency. This client communication latency represents the inherent transmission performance of the network link. Since simply considering communication latency cannot accurately reflect the actual efficiency when transmitting large-scale data, this scheme introduces the aforementioned client latency index. This client latency index represents the actual impact of communication latency under a specific data size, that is, dynamically capturing the amplification or suppression effect of the comparison data size index on the communication burden. This step is as follows: if the client communication latency itself is low, it indicates that the basic communication efficiency between the comparison medical client and the candidate center client is high. At this time, even if the comparison client has a large comparison data size, its communication advantage can still be maintained, and thus the client latency index of the comparison client is small. Conversely, if the client communication latency itself is high, a larger comparison data size will non-linearly exacerbate the negative impact of communication latency, that is, the client latency index increases. This dynamic evaluation mechanism ensures that the selected target center client is not only centrally located in the network, but also optimal in terms of comprehensive communication efficiency, thereby effectively improving the overall performance of the federated learning system.

[0034] In detail, the data scale evaluation of the extracted comparative medical clients to obtain the comparative data scale index includes: Obtain the extracted local dataset of the comparison client from the comparison medical client; Data statistics are performed based on the local dataset of the comparison client to obtain a set of data scale parameters, which include: number of samples, feature dimensions, and data storage size. The data size parameter group is weighted and summed to obtain the original data size index. The original data size index is then normalized to obtain the comparative data size index.

[0035] Understandably, the "comparison client-side local dataset" refers to the client-side local dataset used for comparing medical clients. This dataset is used for machine learning on the subsequent global privacy model to be updated. This client-side local dataset will be explained in detail in later steps. The "data scale parameter set" refers to the set consisting of the number of samples, feature dimensions, and data storage size of the comparison client-side local dataset. The number of samples refers to the number of data samples contained in the comparison client-side local dataset. The feature dimension refers to the number of features contained in each sample data. The data storage size refers to the space volume occupied by the comparison client-side local dataset on the storage medium. For example, a comparison client-side local dataset might be an electronic medical record dataset for cardiology diagnosis from a hospital, containing indicators such as patient age, blood pressure, and cholesterol levels. This dataset consists of multiple specific patient medical records and contains 10,000 records, each with 50 feature fields, and a data storage size of 150MB. Therefore, the number of samples, feature dimensions, and data storage size of this comparison client-side local dataset would be 10,000, 50, and 150MB, respectively.

[0036] It is clear that the original data size index refers to the value obtained after weighted summation. The weighted summation of the data size parameters involves: first, normalizing the sample size, feature dimensions, and data storage size separately; then, multiplying and summing these parameters according to multiple manually set size weights (corresponding to the sample size, feature dimensions, and data storage size, respectively). The resulting value is the original data size index. Normalizing the original data size index refers to normalization using methods such as Min-Max normalization or Z-score standardization.

[0037] In detail, the method of adjusting client communication latency using a comparative data size index to obtain a client latency index includes: The client latency index is calculated using the following formula:

[0038] in, Indicates the client latency index. This indicates client communication latency. This indicates the preset average communication delay. This indicates a comparison of data size index. This represents the preset sensitivity coefficient. This indicates the preset maximum client latency.

[0039] As is clear, the average communication latency refers to the average communication latency of all the client-side communication latency corresponding to the comparison medical client set. The sensitivity coefficient is a manually set constant used to control the degree of non-linear influence of the comparison data size index on the client latency index. The maximum client latency is a manually set constant, which can be the maximum value of all client communication latency or selected from historical client communication latency.

[0040] It should be explained that the above formula for calculating the client latency index introduces the average communication latency. and data size index A nonlinear correction factor was constructed for the original communication delay. The formula operates dynamically, and its mechanism is as follows: when communication delay... At lower levels (e.g.) < ), the adjustment term in the formula The value is close to 1, making the client latency exponent... Approximately equal to client communication latency Even when comparing data size index Larger, due to The value itself is small, and the calculated result is... I It remains at a low level, reflecting the phenomenon that even when transmitting large amounts of data, good communication efficiency can be maintained on high-efficiency communication links. When communication latency... At higher levels (e.g.) > ), in the adjustment items It will amplify high latency The adverse effects, due to Follow Increase and increase rapidly (e.g., when the sensitivity coefficient increases) When the value is 2, it increases quadratically, which leads to an increase in the client latency index. The significant increase reflects the phenomenon that transmitting large amounts of data on high-latency communication links will drastically increase the communication burden.

[0041] S3. Receive the global privacy model to be updated, update the model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated, and obtain multiple sets of model update parameters.

[0042] Understandably, the global privacy model to be updated refers to the initial machine learning model issued by the federated learning server that needs to undergo a new round of training to improve performance. This global privacy model to be updated has integrated privacy protection mechanisms (such as differential privacy). The model update parameter set refers to the set of model update parameters generated by all neighboring medical clients in a certain cluster of neighboring medical clients after they have trained the global privacy model to be updated locally using their respective client-local datasets.

[0043] Specifically, the model parameter update is performed based on multiple neighboring medical client clusters and the global privacy model to be updated, resulting in multiple sets of model update parameters, including: Multiple neighboring medical client clusters are denoted as multiple federated training client sets, and the following operations are performed on each of the multiple federated training client sets: Extract the federated training clients sequentially from the federated training client set; Based on the extracted federated training client, obtain the client's local dataset, use the client's local dataset to train the global privacy model to be updated, and obtain the model update parameter set; The model update parameter set is obtained by summing up the model update parameter sets corresponding to each federated training client; By aggregating the model update parameter sets corresponding to each neighboring medical client cluster, multiple model update parameter sets are obtained.

[0044] It should be explained that the "client-local dataset" refers to a dataset stored locally on each federated training client, containing its private medical data and not shared externally. For example, de-identified electronic medical record data of diabetic patients stored locally on a hospital client, including features such as blood glucose levels and medication records. The "model update parameter set" refers to the update amount or gradient information of model parameters (such as neural network weights, biases, etc.) generated after the federated training client completes model training locally. The "training the global privacy model to be updated using the client-local dataset to obtain the model update parameter set" means that the federated training client uses its private data (i.e., the client-local dataset) locally, and performs several rounds of iterative training on the received global privacy model to be updated using a specified training algorithm (such as stochastic gradient descent). Finally, it calculates the parameter difference between the local model (i.e., the global privacy model to be updated after iterative training) and the initial global model (i.e., the global privacy model to be updated before iterative training), and uses this difference as the model update parameter set for this training.

[0045] S4. Use multiple target center clients to aggregate multiple model update parameter sets to obtain multiple aggregated update parameter sets.

[0046] Understandably, the aggregated update parameter set refers to the new model update parameter set obtained by aggregating the model update parameter set.

[0047] In detail, the aggregation of multiple model update parameter sets using multiple target center clients yields multiple aggregated update parameter sets, including: For each of the multiple model update parameter sets, perform the following operation: Based on the federated training client set corresponding to the model update parameter set, the data processing center client is determined from multiple target center clients; Upload the model update parameter set to the data processing center client to obtain the aggregation center client; The training data set is obtained by acquiring the training data volume of each federated training client in the federated training client set corresponding to the model update parameter group based on the aggregation center client. Normalize the training data set to obtain a normalized data set; The statistical model updates the parameter set of the model update indicator set, and performs the following operations on each model update indicator in the model update indicator set: Based on the model update index, identify the set of parameters to be aggregated in the model update parameter group set, and perform a weighted summation of the set of parameters to be aggregated according to the normalized data volume set to obtain the aggregated update parameters. Summarize the aggregated update parameters corresponding to each model update metric to obtain the aggregated update parameter group; By aggregating the update parameter sets corresponding to each model update parameter set, multiple aggregated update parameter sets are obtained.

[0048] It is clear that the data processing center client refers to the target center client corresponding to the model update parameter set. The aggregation center client refers to the data processing center client after receiving the model update parameter set. The purpose of uploading the model update parameter set to the data processing center client is to enable the aggregation center client to perform aggregation operations on the model update parameter set. The training data set refers to a collection of multiple training data sets, where the training data size refers to the comparative data size index of the local dataset corresponding to a certain federated training client in the federated training client set. The normalized data set refers to the training data set after normalization. The model update index set refers to the set of identifiers of all model parameters (such as weights and biases) that need to be updated contained in the model update parameter set. For example, if a model update parameter set is: the model update parameter set of federated training client A {W1: 0.5, W2: -0.2, b: 0.1} and the model update parameter set of federated training client B {W1: 0.3, W2: 0.4, b: -0.05}, then the model update index set is: [W1, W2, b]. The set of parameters to be aggregated refers to the collection of multiple model update parameters corresponding to the model update metric in the model update parameter set. For example, for the model update metric W1, the set of parameters to be aggregated is W1: 0.5 for federated training client A and W1: 0.3 for federated training client B. The aggregated update parameter refers to the value after weighted summation. Weighted summation of the set of parameters to be aggregated based on the normalized data set means multiplying each normalized data quantity in the normalized data set by its corresponding parameter to be aggregated, and then summing the multiplication results corresponding to different normalized data quantities. The calculated result is the aggregated update parameter.

[0049] S5. Select multiple consensus-based medical clients in the distributed medical client pool, and perform consensus screening on multiple aggregated update parameter groups based on the multiple consensus-based medical clients to obtain multiple qualified update parameter groups.

[0050] It is clear that the consensus medical client refers to a distributed medical client selected from a set of distributed medical clients, responsible for verifying and evaluating the aggregation results. The steps for selecting this consensus medical client are as follows: it is determined from the set of distributed medical clients based on factors such as client reliability, computing resources, or random sampling. For example, parameters such as the model update quality of each distributed medical client in the previous round of federated learning, the number of consensus pass values ​​of 1, and the number of times it correctly verified data integrity when acting as a consensus medical client are obtained. Correct verification means that the consensus pass value obtained from the data integrity verification of the consensus medical client is consistent with the data validity result (i.e., a consensus pass value of 1 is consistent with the data validity result). These parameters are then weighted and summed to obtain a consensus stable value, resulting in multiple consensus stable values. The largest consensus stable value is then selected from these multiple consensus stable values. Each consensus stable value corresponds to a distributed medical client, which serves as multiple consensus medical clients. This refers to the number of consensus-based medical clients. The qualified update parameter group refers to the aggregated update parameter group that has been verified as having complete data and whose model update quality meets the update quality threshold after consensus screening.

[0051] In detail, the process involves consensus screening of multiple aggregated update parameter groups based on multiple consensus-based medical clients to obtain multiple qualified update parameter groups, including: Extract the aggregated update parameter group sequentially from multiple aggregated update parameter groups, and record the extracted aggregated update parameter group as the parameter group to be updated for consensus. Multiple consensus-based medical clients were used to verify the data integrity of the parameter group to be updated in the consensus process, resulting in multiple consensus pass values. Based on multiple consensuses, the validity of the parameter group is determined by value pair aggregation and update, resulting in a data validity result, which is either data valid or data invalid. If the data validity result is that the data is valid, then the parameter group to be updated by consensus is recorded as the valid update parameter group; Summarize the valid update parameter sets to obtain multiple valid update parameter sets; Multiple consensus-based medical clients were used to evaluate the model update quality of multiple valid updated parameter sets, resulting in multiple model update qualities, where each model update quality corresponds one-to-one with a valid updated parameter set. Based on the update quality of multiple models and preset update quality thresholds, data qualification screening is performed on multiple valid update parameter groups to obtain multiple qualified update parameter groups.

[0052] As is clear, the consensus pass value refers to a Boolean value indicating whether the verification passed after data integrity verification. Each consensus pass value corresponds to a consensus medical client, and the value is either 1 or 0. If the consensus pass value is 1, it means the consensus medical client corresponding to that value has passed verification, and the data in the parameter group to be updated is considered complete and unaltered. If the consensus pass value is 0, it means the consensus medical client verification failed, and the data may be incomplete or tampered with. The aforementioned data integrity verification involves: sequentially extracting consensus medical clients from multiple consensus medical clients, using those clients to calculate the hash value of the parameter group to be updated, and recording the calculated hash value as the consensus hash value. Then, retrieving the hash value representing the parameter group to be updated uploaded to that consensus medical client (recorded as the uploaded hash value). If the consensus hash value is the same as the uploaded hash value, the consensus pass value is recorded as 1; otherwise, it is recorded as 0. The consensus pass values ​​of each consensus medical client are then aggregated to obtain multiple consensus pass values.

[0053] It should be explained that the data validity result refers to the result obtained after validity judgment. The data validity result includes data validity or data invalidity. Data validity means that more than half of the consensus medical clients recognize the integrity of the aggregated update parameter group. In this case, the parameter group to be updated by consensus can be recorded as a valid update parameter group. Data invalidity means that sufficient consensus has not been obtained, which indicates that the reliability of the aggregated update parameter group is questionable. In this case, the aggregated update parameter group can be discarded. The above-mentioned validity judgment of the aggregated update parameter group based on multiple consensus pass values ​​means that if the number of consensus pass values ​​with a value of 1 in the multiple consensus pass values ​​exceeds half, then the data is considered valid as the data validity result; otherwise, the data is considered invalid as the data validity result.

[0054] Furthermore, the model update quality refers to the numerical value of the effectiveness of the effective update parameter set in improving the performance of the global privacy model to be updated. The higher the model update quality, the greater the contribution of the effective update parameter set to improving the performance of the global privacy model to be updated. The detailed method for obtaining the model update quality will be given later. The update quality threshold is a human-set constant. If the model update quality is not greater than the update quality threshold, it means that the quality of the effective update parameter set does not meet the requirements. The data qualification screening of multiple effective update parameter sets based on multiple model update qualities and the preset update quality threshold refers to: removing model update qualities whose values ​​are not greater than the update quality threshold from the multiple model update qualities to obtain multiple qualified update qualities, and confirming multiple qualified update parameter sets corresponding to the multiple qualified update qualities among the multiple effective update parameter sets.

[0055] In detail, the method of using multiple consensus-based medical clients to evaluate the model update quality of multiple valid updated parameter sets yields multiple model update quality metrics, including: Perform the following operation for each of the multiple valid update parameter groups: Extract the consensus medical client sequentially from multiple consensus medical clients, and record the extracted consensus medical client as the verification medical client; Obtain the client-side local validation dataset and local global privacy model for the validated medical client; The local global privacy model is updated using an effective set of update parameters to obtain a validated global privacy model. The model validation quality is obtained by performing model validation on the global privacy model based on the client-side local validation dataset. The model validation quality for each validated medical client is summarized to obtain multiple model validation qualities. The average of these multiple model validation qualities is then calculated to obtain the model update quality. The model update quality corresponding to each valid update parameter group is summarized to obtain multiple model update quality values.

[0056] It is clear that the client-side local validation dataset refers to the client-side local dataset used to validate the medical client. The local global privacy model refers to the global privacy model to be updated, deployed on the medical client for validation. The validation global privacy model refers to the local global privacy model updated using an effective update parameter set. Updating the local global privacy model using an effective update parameter set means fusing the effective update parameter set (such as gradients or weight increments) with the current parameters of the local global privacy model, such as updating model weights through weighted averaging or direct replacement. The model validation quality refers to the numerical value that quantifies the predictive performance of the validation global privacy model on the local validation dataset. The higher the model validation quality, the better the performance of the validation global privacy model. Parameters such as model accuracy, model loss value, and F1 score can be used as the model validation quality, or a weighted result of these parameters can be used as the model validation quality.

[0057] S6. Aggregate multiple qualified update parameter groups to obtain the target update parameter group. Use the target update parameter group to update the model parameters of the global privacy model to be updated to obtain the target global privacy model. Based on the target global privacy model, complete the federated learning-based medical data privacy computation.

[0058] It is clear that the target update parameter set refers to the new qualified update parameter set obtained by aggregating multiple qualified update parameter sets. The step of aggregating multiple qualified update parameter sets to obtain the target update parameter set is the same as the step of aggregating multiple model update parameter set sets using multiple target center clients to obtain multiple aggregated update parameter sets, and will not be repeated here. The target global privacy model refers to the global privacy model to be updated after being updated by the target update parameter set.

[0059] To address the problems described in the background, this invention first performs client clustering based on a distributed set of medical clients, resulting in multiple clusters of neighboring medical clients and multiple target center clients. This step selects target center clients by clustering clients based on geographical location and network topology, combined with a dynamic evaluation algorithm considering communication latency and data scale. This effectively optimizes communication efficiency in federated learning. Compared to traditional methods that only consider geographical proximity, it more accurately reflects the actual data transmission burden, thereby reducing overall communication latency and improving system performance. Next, multiple target center clients are used to aggregate multiple model update parameter sets, resulting in multiple aggregated update parameter sets. This step utilizes target center clients for hierarchical aggregation and adjusts the data scale accordingly. The weighted parameter aggregation strategy reduces the burden on the server to directly handle all client updates. Compared to centralized aggregation methods, it reduces communication path complexity and improves aggregation fairness, thereby enhancing the scalability and efficiency of federated learning. Finally, multiple consensus-based medical clients are selected from a distributed medical client pool. Based on these clients, consensus filtering is performed on multiple aggregated update parameter groups to obtain multiple qualified update parameter groups. This step introduces a decentralized consensus mechanism, using multiple consensus-based medical clients to verify data integrity and evaluate model quality of the aggregation results. This effectively enhances the privacy and security of federated learning. Compared to traditional methods lacking verification, it can identify and filter low-quality or malicious updates, ensuring the reliability of model updates. Therefore, this invention can improve communication efficiency in the federated learning process and reduce security risks caused by network latency and unreliable updates.

[0060] like Figure 2 The diagram shown is a functional block diagram of a medical data privacy computing system based on federated learning provided in an embodiment of the present invention.

[0061] The federated learning-based medical data privacy computing system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the federated learning-based medical data privacy computing system 100 may include a learning instruction receiving module 101, a medical client clustering module 102, an update parameter aggregation module 103, and a global model update module 104. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0062] The learning instruction receiving module 101 is used to receive federated learning instructions and identify a distributed medical client set based on the federated learning instructions, wherein the distributed medical client set includes multiple distributed medical clients. The medical client clustering module 102 is used to perform client clustering based on a distributed medical client set to obtain multiple neighboring medical client clusters and multiple target center clients, wherein the target center clients correspond one-to-one with the neighboring medical client clusters, and the neighboring medical client clusters contain the target center clients; The update parameter aggregation module 103 is used to receive the global privacy model to be updated, update the model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated, obtain multiple model update parameter sets, and aggregate the multiple model update parameter sets using multiple target center clients to obtain multiple aggregated update parameter sets. The global model update module 104 is used to select multiple consensus medical clients in the distributed medical client set, perform consensus screening on multiple aggregated update parameter groups based on the multiple consensus medical clients to obtain multiple qualified update parameter groups, aggregate the multiple qualified update parameter groups to obtain the target update parameter group, and use the target update parameter group to update the model parameters of the global privacy model to be updated to obtain the target global privacy model.

[0063] In detail, the modules in the federated learning-based medical data privacy computing system 100 described in this embodiment of the invention employ the same methods as described above. Figure 1 The method uses the same technical means as the federated learning-based medical data privacy computing method described in the article and can produce the same technical effect, so it will not be repeated here.

[0064] like Figure 3 The diagram shown is a schematic representation of an electronic device that implements a federated learning-based method for medical data privacy computation, according to an embodiment of the present invention.

[0065] The electronic device 1 may include a processor 10, a memory 11 and a bus 12, and may also include a computer program stored in the memory 11 and capable of running on the processor 10, such as a federated learning-based method for medical data privacy computation.

[0066] The memory 11 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the electronic device 1, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device 1, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 1. Furthermore, the memory 11 includes both internal storage units and external storage devices of the electronic device 1. The memory 11 can be used not only to store application software and various types of data installed on the electronic device 1, such as the code of a medical data privacy computing method program based on federated learning, but also to temporarily store data that has been output or will be output.

[0067] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., a federated learning-based medical data privacy computation method program) and calls data stored in the memory 11 to perform various functions of the electronic device 1 and process data.

[0068] The bus 12 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 12 can be divided into an address bus, a data bus, a control bus, etc. The bus 12 is configured to realize the connection and communication between the memory 11 and at least one processor 10, etc.

[0069] Figure 3 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 3 The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0070] For example, although not shown, the electronic device 1 may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management system, thereby enabling functions such as charging management, discharging management, and power consumption management through the power management system. The power supply may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0071] Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish communication connections between the electronic device 1 and other electronic devices.

[0072] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device 1 and to display a visual user interface.

[0073] The federated learning-based medical data privacy computation method program stored in the memory 11 of the electronic device 1 is a combination of multiple instructions. When run in the processor 10, it can achieve the following: Receive federated learning instructions, and identify a distributed medical client set based on the federated learning instructions. The distributed medical client set includes multiple distributed medical clients. Based on the distributed medical client set, client clustering is performed to obtain multiple neighboring medical client clusters and multiple target center clients. Among them, the target center clients correspond one-to-one with the neighboring medical client clusters, and the neighboring medical client clusters contain the target center clients. Receive the global privacy model to be updated, update the model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated, and obtain multiple sets of model update parameters; Multiple target center clients are used to aggregate multiple model update parameter sets to obtain multiple aggregated update parameter sets; Multiple consensus-based medical clients are selected in a distributed medical client cluster. Based on these multiple consensus-based medical clients, consensus screening is performed on multiple aggregated update parameter groups to obtain multiple qualified update parameter groups. Multiple qualified update parameter sets are aggregated to obtain the target update parameter set. The model parameters of the global privacy model to be updated are updated using the target update parameter set to obtain the target global privacy model. Based on the target global privacy model, federated learning-based medical data privacy computation is completed.

[0074] Specifically, the processor 10's implementation method for the above instructions can be found in [reference needed]. Figures 1 to 3 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0075] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or system capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0076] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following: Receive federated learning instructions, and identify a distributed medical client set based on the federated learning instructions. The distributed medical client set includes multiple distributed medical clients. Based on the distributed medical client set, client clustering is performed to obtain multiple neighboring medical client clusters and multiple target center clients. Among them, the target center clients correspond one-to-one with the neighboring medical client clusters, and the neighboring medical client clusters contain the target center clients. Receive the global privacy model to be updated, update the model parameters based on multiple neighboring medical client clusters and the global privacy model to be updated, and obtain multiple sets of model update parameters; Multiple target center clients are used to aggregate multiple model update parameter sets to obtain multiple aggregated update parameter sets; Multiple consensus-based medical clients are selected in a distributed medical client cluster. Based on these multiple consensus-based medical clients, consensus screening is performed on multiple aggregated update parameter groups to obtain multiple qualified update parameter groups. Multiple qualified update parameter sets are aggregated to obtain the target update parameter set. The model parameters of the global privacy model to be updated are updated using the target update parameter set to obtain the target global privacy model. Based on the target global privacy model, federated learning-based medical data privacy computation is completed.

[0077] In the embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and actual implementations may have other classification methods.

[0078] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0080] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A medical data privacy computing method based on federated learning, characterized in that, The method comprises: receiving a federated learning instruction, confirming a distributed medical client set based on the federated learning instruction, wherein the distributed medical client set comprises a plurality of distributed medical clients; performing client clustering based on the distributed medical client set to obtain a plurality of neighbor medical client clusters and a plurality of target center clients, wherein the target center clients correspond to the neighbor medical client clusters one by one, and the neighbor medical client clusters contain the target center clients; receiving a global private model to be updated, performing model parameter updating based on the plurality of neighbor medical client clusters and the global private model to be updated to obtain a plurality of model update parameter group sets; aggregating the plurality of model update parameter group sets by using the plurality of target center clients to obtain a plurality of aggregated update parameter groups; selecting a plurality of consensus medical clients from the distributed medical client set, performing consensus screening on the plurality of aggregated update parameter groups based on the plurality of consensus medical clients to obtain a plurality of qualified update parameter groups; aggregating the plurality of qualified update parameter groups to obtain a target update parameter group, performing model parameter updating on the global private model to be updated by using the target update parameter group to obtain a target global private model, and completing federated learning-based medical data privacy calculation based on the target global private model.

2. The federated learning based medical data privacy computing method of claim 1, wherein, The client clustering based on the distributed medical client set to obtain the plurality of neighbor medical client clusters and the plurality of target center clients comprises: extracting distributed medical clients from the distributed medical client set in sequence, and querying distributed client addresses of the extracted distributed medical clients; aggregating the distributed client addresses to obtain a distributed client address set; performing clustering on the distributed medical client set based on the distributed client address set to obtain a plurality of neighbor medical client clusters, wherein each neighbor medical client cluster comprises a plurality of neighbor medical clients; performing center client selection based on the plurality of neighbor medical client clusters to obtain a plurality of target center clients. 3.The federated learning based medical data privacy computing method of claim 2, wherein, The center client selection based on the plurality of neighbor medical client clusters to obtain the plurality of target center clients comprises: performing the following operations on each neighbor medical client cluster in the plurality of neighbor medical client clusters: extracting neighbor medical clients from the plurality of neighbor medical clients of the neighbor medical client cluster in sequence, and taking the extracted neighbor medical clients as candidate center clients; excluding the candidate center clients from the neighbor medical client cluster to obtain a comparison medical client set, wherein the comparison medical client set comprises a plurality of comparison medical clients; extracting comparison medical clients from the comparison medical client set in sequence, performing communication delay detection on the extracted comparison medical clients and the candidate center clients to obtain client communication delays; performing data size evaluation on the extracted comparison medical clients to obtain comparison data size indexes; adjusting the client communication delays by using the comparison data size indexes to obtain client delay indexes; aggregating the client delay indexes corresponding to each comparison medical client to obtain a client delay index set; performing mean value calculation on the client delay index set to obtain an average communication delay index; aggregate the average communication delay indexes corresponding to each candidate center client to obtain an average communication delay index set; identify the minimum communication delay index in the average communication delay index set, and record the candidate center client corresponding to the minimum communication delay index as a target center client; aggregate the target center clients corresponding to each neighbor medical client cluster to obtain a plurality of target center clients.

4. The federated learning based medical data privacy computing method of claim 3, wherein, The data scale evaluation on the extracted comparative medical clients is performed to obtain a comparative data scale index, which includes: obtaining a comparative client local data set of the extracted comparative medical clients; performing data statistics based on the comparative client local data set to obtain a data scale parameter group, wherein the data scale parameter group includes: sample quantity, feature dimension and data storage size; performing weighted summation on the data scale parameter group to obtain an original data scale index, and performing normalization on the original data scale index to obtain the comparative data scale index.

5. The federated learning based medical data privacy computing method of claim 4, wherein, The client communication delay is adjusted by using the comparative data scale index to obtain a client delay index, which includes: The client delay index is calculated by using the following formula: ； wherein, represents a client delay index, represents a client communication delay, represents a preset average communication delay, represents a comparison data size index, represents a preset sensitivity coefficient, represents a preset maximum client delay.

6. The federated learning based medical data privacy computing method of claim 5, wherein, The model parameter update is performed based on the plurality of neighbor medical client clusters and the global privacy model to be updated to obtain a plurality of model update parameter group sets, which includes: The plurality of neighbor medical client clusters are recorded as a plurality of federal training client sets, and the following operations are performed on each federal training client set in the plurality of federal training client sets: extracting federal training clients in the federal training client set in turn; obtaining a client local data set based on the extracted federal training client, and training the global privacy model to be updated by using the client local data set to obtain a model update parameter group; aggregate the model update parameter group corresponding to each federal training client to obtain a model update parameter group set; aggregate the model update parameter group set corresponding to each neighbor medical client cluster to obtain a plurality of model update parameter group sets.

7. The federated learning based medical data privacy computing method of claim 6, wherein, The plurality of model update parameter group sets are aggregated by using the plurality of target center clients to obtain a plurality of aggregated update parameter groups, which includes: The following operations are performed on each model update parameter group set in the plurality of model update parameter group sets: determine a data processing center client in the plurality of target center clients based on the federal training client set corresponding to the model update parameter group set; upload the model update parameter group set to the data processing center client to obtain an aggregation center client; obtain the training data quantity of each federal training client in the federal training client set corresponding to the model update parameter group based on the aggregation center client to obtain a training data quantity set; normalize the training data quantity set to obtain a normalized data quantity set; statistical model update index group in the model update parameter group set, and the following operations are performed on each model update index in the model update index group: identify the parameter set to be aggregated in the model update parameter group set based on the model update index, and perform weighted summation on the parameter set to be aggregated according to the normalized data quantity set to obtain an aggregated update parameter; aggregate the aggregated update parameters corresponding to each model update index to obtain an aggregated update parameter group; Aggregate the aggregated update parameter groups corresponding to each model update parameter group set to obtain a plurality of aggregated update parameter groups. 8.The federated learning based medical data privacy computing method of claim 7, wherein, The plurality of consensus medicine clients are used to perform consensus screening on the plurality of aggregated update parameter groups to obtain a plurality of qualified update parameter groups. The aggregated update parameter groups corresponding to each model update parameter group set are aggregated to obtain a plurality of aggregated update parameter groups. The plurality of consensus medicine clients are used to perform data integrity verification on the to-be-consensus update parameter group to obtain a plurality of consensus pass values. Based on the plurality of consensus pass values, the validity of the aggregated update parameter group is judged to obtain a data validity result, wherein the data validity result is data valid or data invalid. If the data validity result is data valid, the to-be-consensus update parameter group is recorded as an effective update parameter group. The effective update parameter groups are aggregated to obtain a plurality of effective update parameter groups. The plurality of consensus medicine clients are used to perform model update quality evaluation on the plurality of effective update parameter groups to obtain a plurality of model update qualities, wherein the model update quality corresponds to the effective update parameter group in a one-to-one manner. Based on the plurality of model update qualities and a preset update quality threshold, the plurality of effective update parameter groups are screened for data eligibility to obtain a plurality of qualified update parameter groups. 9.The federated learning based medical data privacy computing method of claim 8, wherein, The plurality of consensus medicine clients are used to perform model update quality evaluation on the plurality of effective update parameter groups to obtain a plurality of model update qualities, wherein the model update quality corresponds to the effective update parameter group in a one-to-one manner. The plurality of consensus medicine clients are used to perform model update quality evaluation on the plurality of effective update parameter groups to obtain a plurality of model update qualities, wherein the model update quality corresponds to the effective update parameter group in a one-to-one manner. The system comprises: The learning instruction receiving module is configured to receive a federated learning instruction, and confirm a distributed medicine client set based on the federated learning instruction, wherein the distributed medicine client set comprises a plurality of distributed medicine clients. The medicine client clustering module is configured to perform client clustering based on the distributed medicine client set to obtain a plurality of near neighbor medicine client clusters and a plurality of target center clients, wherein the target center client corresponds to the near neighbor medicine client cluster in a one-to-one manner, and the near neighbor medicine client cluster contains the target center client. The update parameter aggregation module is configured to receive a to-be-updated global privacy model, perform model parameter updating based on the plurality of near neighbor medicine client clusters and the to-be-updated global privacy model to obtain a plurality of model update parameter group sets, and aggregate the plurality of model update parameter group sets by using the plurality of target center clients to obtain a plurality of aggregated update parameter groups. ​ ​ 10.A medical data privacy computing system based on federated learning, characterized in that, ​ ​ ​ ​ The global model updating module is configured to select a plurality of consensus medical clients from the distributed medical client set, perform consensus screening on the plurality of aggregated update parameter groups based on the plurality of consensus medical clients, obtain a plurality of qualified update parameter groups, aggregate the plurality of qualified update parameter groups to obtain a target update parameter group, perform model parameter updating on the global private model to be updated by using the target update parameter group, and obtain a target global private model.

Images