Client incentive optimization method and system for semi-asynchronous federated learning

Abstract

The invention discloses a client incentive optimization method and system for semi-asynchronous federated learning. The method comprises the following steps: classifying multi-dimensional private information of a client according to the multi-dimensional private information of the client and determining a contract form; determining a utility function of the client and the server; mathematical modeling is carried out on an excitation optimization problem, wherein the mathematical modeling comprises excitation compatibility and individual rationality constraint; a discontinuous network architecture learning method comprising a linear component and a nonlinear bias component is utilized to optimize a contract structure so as to determine an optimal contract, and discontinuity possibly occurring on a feasible region boundary of a server utility function is effectively processed; the server issues a contract, and the client selects and executes the contract; the server issues rewards and aggregates the models according to client execution conditions, wherein aggregation weights consider model obsolessness and data quality. Through the innovative learning type contract design, the real participation of the client is effectively stimulated, the training efficiency and stability of the system are improved, and challenges caused by the discontinuity of multi-dimensional information and utility functions are overcome.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a client incentive optimization method and system for semi-asynchronous federated learning. Background Art

[0002] Federated Learning (FL), an emerging distributed machine learning paradigm, allows multiple clients (such as mobile devices and institutions) to collaboratively train a global model while protecting user data privacy. Traditional synchronous federated learning requires all participants to synchronize updates during each round of communication. This is susceptible to the "straggler effect," where clients with slower computing speeds or poor network conditions can slow down the overall training process.

[0003] To alleviate this problem, Semi-Asynchronous Federated Learning (SAFL) was proposed. SAFL combines the characteristics of synchronous and asynchronous updates, allowing the server to aggregate updates after collecting a certain number of client updates, thereby improving training efficiency and system throughput. Despite SAFL's advantages, incentivizing continuous and high-quality client participation remains a core challenge in practical deployments. Clients participating in federated learning consume their own computing resources and communication bandwidth, and may face the risk of data privacy leakage. Without an effective incentive mechanism, rational and selfish clients may be unwilling to contribute their valuable resources and data, resulting in poor model training results or system crashes.

[0004] Research on incentive mechanisms for federated learning has made some progress, but most efforts focus on synchronous federated learning scenarios or modeling client-side private information (e.g., data size, CPU frequency) in a single or two-dimensional manner. However, in SAFL, client-side private information is often multidimensional, such as data quality (which influences model accuracy), computing power (which impacts local training time), and privacy preferences (which influence the privacy risk they are willing to tolerate and the strength of the privacy protection techniques they employ). This multidimensional information is known only to the client itself but is crucial for the server to formulate optimal incentive strategies.

[0005] Existing research attempting to process multidimensional information often relaxes incentive compatibility (IC) requirements to simplify computation. This means there's no guarantee that clients will truthfully report their types or select contracts that exactly match their types, potentially leading to suboptimal resource allocation. More critically, these studies often overlook a crucial fact: when a server maximizes its own utility (e.g., improved model accuracy minus client costs), its utility function may exhibit discontinuities at the boundaries of the feasible contract parameter space. For example, when the reward just meets a client's participation threshold, the server's utility may jump. Traditional gradient-based optimization methods or those that rely on function continuity struggle to effectively solve such problems, limiting the performance of the designed contracts.

[0006] Therefore, in the semi-asynchronous federated learning scenario, designing an incentive mechanism that can fully consider multi-dimensional private information such as client data quality, computing power, privacy preferences, and effectively handle the discontinuity of the server utility function while ensuring incentive compatibility and individual rationality is of great significance to improving the overall performance and execution efficiency of the SAFL system. Summary of the Invention

[0007] In response to the problems existing in the above-mentioned background technology, the main purpose of the present invention is to provide a client incentive optimization method and system for semi-asynchronous federated learning. This method and system can effectively motivate clients with multi-dimensional private information to actively and truly participate in the federated learning process, while overcoming the difficulties brought about by the discontinuity of the server utility function to contract optimization, thereby improving the training efficiency, model performance and collaborative stability of the overall system.

[0008] To achieve the above object, the present invention adopts the following technical solutions: A client incentive optimization method for semi-asynchronous federated learning includes the following steps: S1: Classify clients based on their multi-dimensional private information (such as data quality, computing resources, and privacy preferences) and determine contract forms for different categories of clients. The contract form usually defines the resources that the client needs to contribute (such as the number of local iterations). ) and the corresponding reward R paid by the server for this.

[0009] S2: Determine the server's utility function and the client's utility function. The server's utility function aims to maximize the difference between the value obtained from the client (such as improved model accuracy) and the cost paid, while also taking time into account. The client's utility function aims to maximize the net benefit obtained while meeting its costs.

[0010] S3: Mathematically model the incentive optimization problem. The goal is to maximize the server's total utility. Constraints include incentive compatibility constraints (ensuring that clients choose contracts that truly reflect their type to maximize their own utility) and individual rationality constraints (ensuring that the utility of participating in federated learning is not negative).

[0011] S4: Adopt a learning-based approach, in particular, a discontinuous network architecture, to optimize the contract structure, i.e., determine the number of local iterations The corresponding optimal reward By combining linear components and bias components for discontinuities, the proposed network architecture can effectively fit and optimize the server utility function that may have discontinuities at the boundary of the feasible region.

[0012] S5: The server will design the optimal contract (including , The contract is sent to clients that are ready to participate in semi-asynchronous federated learning. Based on their own type (multi-dimensional private information) and the principle of utility maximization, the client selects the most suitable terms from the contract menu and reports their selection to the server.

[0013] S6: The client strictly follows the requirements of the contract terms (such as the number of local iterations , data usage specifications, etc.) to train the local model and upload the trained local model parameters to the server within the specified time.

[0014] S7: After receiving the model uploaded by the client, the server first verifies whether the client has complied with the contract (such as the number of iterations and timely upload). For clients that fulfill their obligations, the server issues the contractually agreed-upon rewards. The server then performs a weighted semi-asynchronous aggregation of the collected local model parameters from different clients. The aggregation weights take into account the staleness of the model updates and the quality of the client's data to generate a new global model.

[0015] Furthermore, in step S1, the client's data quality is divided into X levels, and the computing resources are divided into The privacy budget corresponding to the privacy preference is divided into Z levels, and the combination type of the client is ,in The contract form is expressed as ,in is the number of local iterations, Contributing to clients of type (x,y,z) The total reward obtained after local iterations.

[0016] Furthermore, in step S2, the utility function of the server

[0017] in is the local precision function, is the data quality parameter, is the privacy preference parameter, The maximum tolerance time for the server to submit the local model to the client. To calculate the time, is the communication time, and is the weight factor, Indicates the server's satisfaction with the client's participation in the global aggregation delay; the client's utility function includes the reward from the server, model training cost, data leakage cost, and model upload cost,

[0018] in For type The total cost to the client, Represent its computing energy consumption and communication energy consumption respectively, represents the unit cost per joule, represents the cost of privacy leakage; therefore, the utility function of the client is:

[0019] Furthermore, in step S4, the learning-based method for optimizing the contract structure includes: Based on the marginal cost of the client participating in federated learning Reclassify the client categories, ; Will target category Client rewards Rewritten to match the local iteration count , marginal cost and unit rewards The relevant functional form; The discontinuous network architecture is used to fit the server's utility function and optimize the unit reward and local iteration count The discontinuous network architecture is able to capture the discontinuity of the server utility function at the boundary of the feasible unit reward domain.

[0020] Furthermore, the discontinuous network architecture includes a first subnetwork and a second subnetwork; the first subnetwork is a traditional network based on the rectified linear unit (ReLU) activation function, which is used to capture the linear part of the server utility function; the second subnetwork is a bias network based on the hyperbolic tangent (Tanh) activation function, whose input depends on the activation mode of the first subnetwork, and is used to capture the discontinuity in the server utility function; the output of the discontinuous network architecture is the sum of the output of the first subnetwork and the output of the second subnetwork.

[0021] Furthermore, in step S7, the weight of the weighted semi-asynchronous aggregation operation According to the staleness of the model uploaded by the client and client data quality Calculated, specifically ,in Represents the staleness effect of client-side latency, represents the data quality impact of the client, and B is the client set of the current aggregation batch.

[0022] The present invention also provides a client incentive optimization system for semi-asynchronous federated learning, which includes functional modules corresponding to the above method steps: The client classification and contract form determination module; the utility function determination module; the mathematical modeling and constraint setting module; the learning-based contract optimization module (with a discontinuous network optimization unit at its core); the contract management and delivery module; the client interaction and execution module (typically located on the client device); and the verification and aggregation module work together to implement the incentive optimization method described.

[0023] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: This invention discloses a smart contract optimization method based on federated learning. This method obtains multidimensional private information from clients, determines the client type combination and contract clause form, and constructs utility functions for both the server and client. The method then constructs an optimization problem to maximize server utility, reclassifies client types using client marginal costs, and approximates the server utility function using a discontinuous network architecture. By training this network architecture, the method solves the optimization problem, obtains optimal contract clauses, and distributes them to the clients. Finally, the method receives local models uploaded by clients that meet the clauses and performs a semi-asynchronous aggregation operation. This method incorporates multidimensional private information (data quality, computing power, and privacy preferences) into the design of the incentive mechanism for semi-asynchronous federated learning, achieving more refined and personalized incentives and filling a gap in research on multidimensional incentives in the SAFL scenario. The method also optimizes contract design using a learning method based on a discontinuous network architecture, effectively addressing the difficulty traditional optimization methods have in handling discontinuities in the server utility function at the boundary of the feasible region, thereby enabling the solution of a more optimal contract structure. Through carefully designed incentive compatibility and individual rationality constraints, we ensure that clients will act truthfully and choose the contract terms that best suit them, thereby incentivizing them to actively participate in federated learning and overcoming client selfishness. Overall, this invention maximizes server utility while protecting client privacy, improving the efficiency and effectiveness of federated learning and significantly enhancing the participation, training efficiency, model quality, and collaborative stability and fairness of semi-asynchronous federated learning systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the overall process of the method of the present invention.

[0025] Figure 2 It is a typical architectural diagram of the system of the present invention.

[0026] Figure 3 It is a schematic diagram of the discontinuous network architecture of the present invention.

[0027] Figure 4 This is an exemplary diagram of the client's utility changes under different contract terms, demonstrating incentive compatibility.

[0028] Figure 5 This is an example comparison chart showing how the client's utility changes when they comply with or violate the terms of the contract.

[0029] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0030] Example 1 like Figure 1 In this embodiment, a method for designing an incentive mechanism for semi-asynchronous federated learning may specifically include: S101. Obtain multidimensional private information of the client, where the multidimensional private information includes the client's data quality, computing resources, and privacy preferences, and determine the type combination of the client based on the multidimensional private information; determine the form of the contract terms, where the contract terms represent the resource combination that the client is willing to contribute, the number of local iterations, and the reward obtained.

[0031] During the server initialization phase, basic identification information of participating clients is collected to generate a preliminary client registration dataset. Based on this preliminary registration dataset, a classification algorithm is used to preliminarily group clients and determine their potential categories. Exploratory contract clause drafts are sent to clients to obtain their responses to different clauses and determine their preferences for privacy protection and resource allocation. Based on these responses, a pre-defined multi-dimensional information evaluation model is used to analyze the client's data quality, computing resources, and privacy preferences, resulting in a eigenvalue of the client's multi-dimensional private information. If data for any dimension in the eigenvalue is missing or abnormal, supplementary information is inferred from historical participation records and behavioral patterns to determine a complete dataset of the client's multi-dimensional private information. Based on this complete dataset, a clustering algorithm is used to conduct in-depth analysis of the client's multi-dimensional private information, identifying specific client type combinations. By matching these type combinations with pre-defined contract clause templates, customized contract clause drafts are generated for each client type to determine whether they meet the client's utility maximization criteria. If the matching results indicate that the utility of some type combinations does not meet expectations, the contract structure is optimized using a discontinuous network architecture, and clause parameters are adjusted to determine the final optimal contract terms. Based on the optimal contract terms, a formal contract is issued to the client and feedback data is collected to obtain the client's selection results and determine whether the client has signed terms that match its type combination.

[0032] Specifically, the client type is determined based on the client's data quality. Divided into Level, computing resources Divided into Level, privacy budget Divided into Level, and then determine the client's combination type The contract terms are determined in the form of Indicates the terms of the contract, meaning that the client is willing to contribute type The resource combination and local iteration number are The total reward received is .

[0033] Specifically, the server collects relevant information of the client during the initialization phase, assuming that clients, each client With 3D private information: Data quality (For example, the comprehensive indicators of dataset size and label accuracy are divided into levels: high, medium, low), computing resources (For example, CPU / GPU processing speed is divided into levels: fast, medium, slow), privacy preferences (For example, the differential privacy budget that one is willing to accept is divided into Z=2 levels: strict and loose). Therefore, the client type There are 3x3x2=18 options in total.

[0034] The server generates a local iteration count for each possible (m=1, ..., M, for example, M=5, ∈{10, 20, 30, 40, 50} epochs) design a reward placeholder). The contract form is .

[0035] S102: Determine a utility function of the server and a utility function of the client, where the utility function is related to a type combination of the client and the contract terms.

[0036] Based on the client's multi-dimensional private information, the system classifies and processes the client's type combinations, taking into account data quality, computing resources, and privacy preferences. Feature extraction is performed on the classified client type combinations, and the behavioral response data of each client type under different contract terms is analyzed to determine the initial model of the client's utility function. The system uses historical server interaction data and global model training records, combined with the characteristics of the client type combinations, to construct a preliminary framework for the server utility function and obtain its mathematical expression. Based on the correlation data between the client type combinations and contract terms, the initial client utility function model is mapped and matched with the contract term parameters to determine the utility trend of each client type under different terms.

[0037] Among them, the determination of the server utility function, the server utility includes the test accuracy of the local model, time consumption and distributed rewards, type The benefits to the server are:

[0038] in Represents the local accuracy function, which means that the local model has been The expected accuracy improvement after rounds of iteration is related to data quality Positively correlated with the strength of privacy protection (usually Negative correlation, that is The smaller the value, the stronger the protection, but the greater the loss of accuracy may be). Indicates the server's satisfaction with the client's participation in the global aggregation delay. Indicates the maximum tolerance time for the server to submit a local model to the client. represents the calculation time, Indicates the communication time, and Denotes the weight factor. The client utility function is determined, which includes the reward from the server, model training cost, data leakage cost, and model upload cost. The total cost is:

[0039] in represent computing energy consumption and communication energy consumption respectively, represents the unit cost per joule, represents the privacy leakage cost, represents the communication cost, represents the computational cost. Therefore, the utility function of the client is:

[0040] S103. Construct a mathematical model for the incentive optimization problem, wherein the mathematical model aims to maximize the total utility of the server and satisfies incentive compatibility constraints and individual rationality constraints.

[0041] Using mathematical modeling, we take maximizing server utility as the optimization objective and, combined with incentive compatibility constraints and individual rationality constraints, formulate a contract design optimization problem. To ensure that clients sign their own contract terms, the following two constraints must be met.

[0042] Incentive compatibility constraint: type is The client of must choose the contract terms that match its type to maximize its utility, i.e.

[0043] Individual rationality constraint: Clients will only participate if their utility is non-negative, i.e.

[0044] Note that incentive compatibility is very important in practice because it can encourage clients to truthfully disclose their types. Under the premise of satisfying incentive compatibility and individual rationality constraints, the server aims to maximize its own utility and solve the optimal contract. The mathematical optimization problem formula for its contract design is:

[0045] The constraints include incentive compatibility constraints, individual rationality constraints and time constraints. .

[0046] By classifying the multidimensional private information of clients, different client types are identified based on data quality, computing resources, and privacy preferences, and the basic characteristic parameters of each client type are determined. Based on the classification results, the server's utility function and the client's utility function are constructed, respectively expressing the server's profit objectives in contract design and the client's profit objectives in training. Using mathematical modeling, maximizing server utility is the optimization objective, combined with incentive compatibility constraints and individual rationality constraints to formulate a contract design optimization problem. A discontinuous network architecture is used to optimize the possible discontinuities of the server's utility function within the feasible domain, obtaining the initial contract data required for network training. The discontinuous network architecture is trained on random contract data, and after multiple rounds of iterative calculations, the mean squared error between the network output and the server's actual utility value is minimized. The trained network results are solved using the interior point method to determine the specific parameters and structure of the optimal contract terms. The server distributes the optimal contract terms and the initialized global model to each client, obtaining feedback on the client's selection of contract terms based on their category. If the client selects a contract clause that matches its category, local training is performed using the corresponding resource combination and the model update is uploaded on schedule to determine whether the client complies with the contract. If the client selects a mismatched clause, its selection is recorded and marked as a potential breach of contract. Based on the client's upload behavior and contract compliance, a reward allocation or deduction mechanism is implemented, resulting in the final client utility value and the server's semi-asynchronous aggregate update of the global model.

[0047] S104. Optimizing the correspondence between the reward and the number of local iterations in the contract form using a learning-based method to determine an optimal contract. The learning-based method can handle possible discontinuities in the server utility function.

[0048] As the number of client types increases, incentive compatibility constraints make traditional methods computationally infeasible. Therefore, we propose to use learning methods to optimize the contract structure. First, we use the marginal cost of the client relative to the number of local iterations to reclassify the categories. The calculation formula is:

[0049] in Representation type The marginal cost of a client in local iteration. If a client is of type , then there is the following corresponding relationship: , and Due to the reward With the number of local iterations At the same time, as the number of local iterations increases, the utility obtained by the server shows a decreasing trend. , we rewrite the reward as: ,in Representation type Unit reward. Feasible domain analysis of unit reward:

[0050]

[0051] in Therefore, the feasible domain of unit reward can be rewritten as:

[0052] The analysis leads to the following conclusions: the utility function of the server exist The global optimal contract May appear in the piecewise linear region at the border.

[0053] Based on the above discussion, let’s reframe the original contract design problem:

[0054] constraint: , , incentive compatibility constraints and individual rationality constraints. Indicates in the fragment Based on this problem, we adopt Figure 3 The discontinuous network architecture shown is used to optimize the contract structure and obtain the optimal contract. The network consists of two sub-networks, a traditional network using the ReLU activation function (the network has linear properties), and a bias network using the Tanh activation function , used to capture discontinuities in functions. Specifically, when constructing a discontinuous network architecture, a biased network is used. and traditional networks The dual-branch structure of the network Responsible for handling activation mode changes, Maintaining the continuity characteristics of the traditional network, the weights are initialized using the He normal distribution to adapt to the ReLU activation function.

[0055] The discontinuous network architecture is initialized and predefined parameters for the server and client utility functions are loaded to obtain an initialized network model. Based on the initialized network model, multiple contract data sets are randomly generated as training samples, and the input structure of the training dataset is determined. The discontinuous network architecture is iteratively trained using the training dataset, and network parameters are adjusted to minimize the mean squared error between predicted and actual utility values, resulting in a trained network model. The trained network model is used to simulate the server utility function's response under different contract terms, obtaining data on the feasible domain distribution of the utility function. Based on this feasible domain distribution data, the interior point method is used to solve the optimization problem of maximizing server utility and determine a preliminary solution set for the optimal contract terms. If the contract terms in the preliminary solution set satisfy incentive compatibility and individual rationality constraints, the solution set is retained as a candidate solution and a candidate contract term combination is determined. If the constraints are not met, the optimization parameters are adjusted and the solution is repeated to obtain a new preliminary solution set. A matching calculation is performed on the candidate contract term combinations, combined with the multi-dimensional private information categories of the clients, to determine the optimal contract terms for each client type. Based on the optimal contract terms for each client type, customized incentive contract data is generated to determine the final contract content issued to the client. Based on the final contract content, the client's selection feedback data is recorded and stored in the server database, obtaining the contract execution basis for the client's participation in semi-asynchronous federated learning.

[0056] S105. The server sends the optimal contract to each client that is ready to participate in semi-asynchronous federated learning. The client selects the contract terms that maximize utility based on its own type and feeds back to the server.

[0057] The server will design the optimal contract The contract is sent to the client, and the client selects the clause that maximizes its own utility by comparing each contract clause and notifies the server. Note that due to the existence of incentive compatibility constraints, the client can only maximize its own utility by signing the contract clause of its own type, which prevents the client from misrepresenting its own type. Figure 3 As shown, the client can only maximize its utility by signing contract terms that match its own category, and it is guaranteed that the client's utility is non-negative.

[0058] The server calculates the optimal contract terms and uses the interior point method to optimize the training results of the discontinuous network architecture to obtain the specific content of the optimal contract. Based on the calculated optimal contract terms, the server encodes and distributes them to all clients participating in semi-asynchronous federated learning, ensuring the integrity of the distribution process. The client receives the distributed contract terms and performs a utility evaluation based on its own data quality, computing resources, and privacy preferences, determining the contract terms that maximize its own utility. If a client selects a contract term after evaluation, it uploads its selection to the server, confirming its selection record. Based on the list of signed clients, the server sends a local training start command to eligible clients and determines the allocation status of the training task. Eligible clients receive the start command, execute local model training, and use differential privacy techniques to protect their data, obtaining the trained local model update data. The client uploads the trained local model update data to the server buffer, ensuring the integrity and timeliness of the uploaded data.

[0059] S106. The client performs local model training according to the selected contract terms and uploads the trained local model parameters.

[0060] Each contract clause specifies the resources and time delay that the client needs to contribute. The client needs to perform local training and upload in strict accordance with the contract terms.

[0061] S107. The server verifies whether the client meets the contract terms, issues corresponding rewards to the client that meets the requirements, and performs weighted semi-asynchronous aggregation operations on the local model parameters uploaded by the client to update the global model.

[0062] For clients that have completed training and upload, the server will strictly check according to the terms of the contract they signed. Clients that are verified to be correct will receive corresponding rewards, otherwise rewards will be deducted. When the buffer reaches the set threshold, the aggregation operation begins. The server verifies the client Whether timely upload or use of specified resources has been completed. If fulfilled, rewards will be paid. The server maintains a buffer. When a sufficient number (for example, B) of client model updates are collected, aggregation is performed. Specifically, the server obtains the local model and delay information uploaded by the client from the buffer and assigns weights to them. The aggregation weight is calculated as follows:

[0063] Used to measure the model staleness, where Represents the staleness effect of client-side latency, Represents the data quality impact of the client.

[0064] Example 2 As attached Figure 2 As shown, the system model used by the above method of the present invention is composed of The system consists of 8 types of clients and one server. To ensure that there are enough clients participating in semi-asynchronous federated learning, the server needs to design corresponding contract terms for each type of client and initialize the global model before the global model training begins.

[0065] Example 3 Figure 3 is a discontinuous network architecture diagram of the present invention, specifically, Generated activation mode will serve as When the network When the activation pattern of the discontinuous network remains unchanged, Consistent with traditional networks, continuity can be captured. When the activation pattern of the network changes, The dynamic response of the network (output change) can accurately capture the discontinuity. For training, first randomly generate Contract data , and use these data to train the network for T rounds, with the goal of minimizing the predicted utility value output by the network and the actual utility value of the server The goal is to enable this network architecture to approximate the server utility function. After training, the interior point method is used to solve the optimal contract terms.

[0066] The server will send the designed contract terms and global model to the participating clients, and the clients will select the contract terms according to their own circumstances. Figure 3 As shown, the client can only maximize its utility by signing contract terms that match its own category, and it is guaranteed that the client's utility is non-negative.

[0067] If the client signs the type The contract terms need to use the type The resource combination is used for local training and also needs to be uploaded on time. Figure 4 As shown in Figure 2, if the client violates the terms of the contract or delays uploading due to network instability, the server will deduct the corresponding reward, and the client's utility will become negative.

[0068] After the client completes local training and upload, the server will check it. Clients that have not violated the terms of the contract will receive the corresponding rewards. Otherwise, the server will deduct their rewards.

[0069] Figure 4The utility curves of 8 clients under different contract terms are displayed, showing that they only achieve maximum utility on contracts that match their own type. Figure 5 It is possible to show the changes in utility caused by different behaviors of the client, showing that if the contract is not completed, its final benefits will be negative or much lower than expected.

[0070] Example 4 This example applies the method to a semi-asynchronous federated learning scenario in smart healthcare. Scenario Description: Multiple hospitals (clients) wish to jointly train an early prediction model for a disease (such as diabetes) without sharing their private patient data. A central research institution acts as the server.

[0071] 1. Client (hospital) characteristics (S1): Data quality: Hospital A has a large amount of high-quality annotated case data (high), Hospital B has less data but of acceptable quality (medium), and Hospital C has noisy data (low).

[0072] Computing resources: Hospital A has dedicated servers (fast), Hospital B uses ordinary workstations (medium), and Hospital C has limited computing resources (slow).

[0073] Privacy preference: Hospital A requires extremely high privacy protection for data. The value is small (strict); Hospitals B and C are relatively loose. Larger values ​​(lenient).

[0074] 2. Utility Function and Contract (S2, S3): Server: The goal is to obtain a high-precision global prediction model while minimizing the incentive costs and waiting time costs paid to the hospital.

[0075] Hospitals: Participation yields benefits that cover their computational, human, and deployment costs of additional technologies (e.g., differential privacy) to meet privacy requirements.

[0076] The mathematical model was established as in Example 1.

[0077] 3. Contract Optimization and Learning (S4): The server discovered that for Hospital A, which has high data quality and strict privacy requirements, if the reward is below a certain threshold, it may choose not to participate or only provide a very small number of iterations. When the reward is slightly above the threshold, Hospital A's willingness to participate and the quality of its contribution increase significantly, resulting in a step-change in the server's utility.

[0078] The discontinuous network architecture ψ described in Example 2 is used. The network can capture this kind of utility function by learning a large amount of simulated contract data. exist The final optimized contract will accurately give Hospital A a reward slightly higher than the initial reward and optimal for the server, while also formulating appropriate contract terms for Hospitals B and C.

[0079] 4. Contract execution and model aggregation (S5, S6, S7): Each hospital chooses a contract. For example, Hospital A may choose a contract with more iterations and higher rewards, while Hospital C may choose a contract with fewer iterations and lower rewards.

[0080] The hospital conducts local model training according to the contract (for example, using local electronic health records to train an LSTM-based risk prediction model) and uploads the model parameters.

[0081] The server verifies fulfillment and pays rewards. During aggregation, Hospital A's model receives a higher weight due to its high data quality; if a hospital uploads its model later, its weight will be lowered accordingly.

[0082] In this way, even in the face of extremely strong heterogeneity among hospitals, complex private information and discontinuity in server utility, the method of the present invention can design an effective incentive contract to encourage each hospital to actively and high-quality participate, and ultimately obtain a global disease prediction model with excellent performance.

[0083] Example 5 This embodiment provides a client incentive optimization system for semi-asynchronous federated learning, which includes functional modules corresponding to the above method steps: The client classification and contract form determination module; the utility function determination module; the mathematical modeling and constraint setting module; the learning-based contract optimization module (with a discontinuous network optimization unit at its core); the contract management and delivery module; the client interaction and execution module (typically located on the client device); and the verification and aggregation module work together to implement the incentive optimization method described.

[0084] In summary, this paper proposes a multi-dimensional incentive mechanism design method suitable for semi-asynchronous federated learning scenarios. This method fully considers the multi-dimensional private information of clients (such as data quality, computing power, and privacy preferences) and the discontinuous nature of the server's utility function. This method more accurately characterizes the true cost of participating in federated learning for heterogeneous clients, effectively incentivizing clients to accept and sign contract terms that meet the server's expectations, and thus improving overall system participation and efficiency.

[0085] Compared to existing technologies: 1. This invention fills a gap in incentive mechanism design in semi-asynchronous federated learning environments. Addressing the inadequate consideration of key factors in existing research, it proposes a learning-based contractual incentive mechanism. This mechanism integrates multidimensional private information, such as data quality, computing resources, and privacy preferences, enabling more refined and personalized incentive design. To our knowledge, this is the first contract design solution to incorporate multidimensional private information in a semi-asynchronous federated learning scenario.

[0086] 2. This invention overcomes the limitations of traditional incentive mechanisms in contract design. The proposed learning-based contract mechanism fully accounts for the discontinuity of the server's utility function. By introducing a discontinuous network optimization method, it successfully solves for the optimal contract structure. Furthermore, this mechanism exhibits excellent scalability, allowing for flexible adjustments to contract structures based on client needs, including the addition of key factors or the removal of irrelevant ones, to further enhance the system's adaptability and practicality.

[0087] 3. Incentivize clients through contract design, so that every client participating in semi-asynchronous federated learning will receive a profit of no less than zero. This will motivate client participation and satisfy individual rationality. 4. The design of this incentive mechanism can not only effectively prevent malicious clients from misrepresenting their own types to obtain higher profits, but also incentivize clients to maintain a more stable network environment, thereby improving the security and reliability of the overall system.

[0088] The above disclosure is only a preferred embodiment of the present invention, and it is certainly not intended to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A client incentive optimization method for semi-asynchronous federated learning, characterized in that: The following steps are involved: S1: Classify clients based on their private information in at least three dimensions: data quality, computing resources, and privacy preferences. Determine contract forms for different categories of clients, including the number of local iterations the client must contribute and the corresponding rewards. S2: Determine a utility function of the server and a utility function of the client, wherein the utility function of the server is related to the local model accuracy contributed by the client, the time consumed to complete training and communication, and the reward paid to the client, and the utility function of the client is related to the reward obtained from the server and the cost of the client participating in learning; S3: Construct a mathematical model for the incentive optimization problem, wherein the mathematical model aims to maximize the total utility of the server and satisfies incentive compatibility constraints and individual rationality constraints; S4: optimizing the relationship between the reward and the number of local iterations in the contract form using a learning-based method to determine an optimal contract, wherein the learning-based method can handle possible discontinuities in the server utility function; S5: The server sends the optimal contract to each client that is ready to participate in semi-asynchronous federated learning. The client selects the contract terms that maximize utility based on its own type and feeds them back to the server. S6: The client trains the local model according to the selected contract terms and uploads the trained local model parameters; S7: The server verifies whether the client meets the contract terms, issues corresponding rewards to the client that meets the requirements, and performs weighted semi-asynchronous aggregation operations on the local model parameters uploaded by the client to update the global model.

2. The method according to claim 1, characterized in that In step S1, the data quality of the client is divided into The computing resources are divided into The privacy budget corresponding to the privacy preference is divided into The client's combination type is ,in The contract form is expressed as ,in is the number of local iterations, For type Clients are conducting The total reward obtained after local iterations.

3. The method according to claim 1, characterized in that In step S2, the utility function of the server in, is the local precision function, is the data quality parameter, is the privacy preference parameter, The maximum tolerance time for the server to submit the local model to the client. To calculate the time, is the communication time, and is the weight factor, Indicates the server's satisfaction with the client's participation in the global aggregation delay; the client's utility function includes the reward from the server, model training cost, data leakage cost, and model upload cost, ; in, For type The total cost to the client, Represent its computing energy consumption and communication energy consumption respectively, represents the unit cost per joule, represents the cost of privacy leakage; therefore, the utility function of the client is: 。 4. The method according to claim 1, wherein In step S4, the learning-based method for optimizing the contract structure includes: Based on the marginal cost of the client participating in federated learning Reclassify the client categories, ; Will target category Client rewards Rewritten to match the local iteration count , marginal cost and unit rewards The relevant functional form; The discontinuous network architecture is used to fit the server's utility function and optimize the unit reward and local iteration count The discontinuous network architecture is able to capture the discontinuity of the server utility function at the boundary of the feasible unit reward domain.

5. The method according to claim 4, characterized in that The discontinuous network architecture includes a first subnetwork and a second subnetwork; the first subnetwork is a traditional network based on the rectified linear unit (ReLU) activation function, which is used to capture the linear part of the server utility function; the second subnetwork is a bias network based on the hyperbolic tangent (Tanh) activation function, whose input depends on the activation mode of the first subnetwork, and is used to capture the discontinuity in the server utility function; the output of the discontinuous network architecture is the sum of the output of the first subnetwork and the output of the second subnetwork.

6. The method according to claim 1, characterized in that In step S7, the weight of the weighted semi-asynchronous aggregation operation According to the staleness of the model uploaded by the client and client data quality Calculated, specifically ,in Represents the staleness effect of client-side latency, represents the data quality impact of the client, and B is the client set of the current aggregation batch.

7. A client incentive optimization system for semi-asynchronous federated learning, characterized in that: include: a client classification and contract form determination module, configured to classify clients according to their multi-dimensional private information and determine contract forms for different categories of clients, wherein the contract forms include the number of local iterations required by the client and the corresponding rewards; a utility function determination module configured to determine a utility function of the server and a utility function of the client, wherein the utility function of the server is related to the local model accuracy, time consumption, and payment reward contributed by the client, and the utility function of the client is related to the reward obtained and the participation cost; A mathematical modeling and constraint setting module is configured to construct a mathematical model of an incentive optimization problem with the goal of maximizing the total utility of the server, and to set incentive compatibility constraints and individual rationality constraints; a learning-based contract optimization module configured to optimize the relationship between rewards and local iteration counts in the contract form using a learning-based method to determine an optimal contract, wherein the learning-based method is capable of handling possible discontinuities in the server utility function; A contract management and issuance module configured to issue the optimal contract to each client ready to participate and receive the contract terms selected by the client; A client interaction and execution module configured to enable the client to perform local model training according to the selected contract terms and upload the trained local model parameters; The verification and aggregation module is configured to verify whether the client meets the contract terms, issue corresponding rewards to the client that meets the requirements, and perform weighted semi-asynchronous aggregation operations on the local model parameters uploaded by the client to update the global model.

8. The system according to claim 7, characterized in that The learning-based contract optimization module is further configured to: Based on the marginal cost of the client participating in federated learning Reclassify the client categories, ; The reward for clients in category g will be Rewritten to match the local iteration count , marginal cost and unit rewards The relevant functional form; A continuous network architecture is used to fit the server's utility function and optimize the unit reward and local iteration count The optimal contract is solved by a combination of,the discontinuous network architecture includes a first sub-network based on the ReLU activation function for capturing the linear part of the server utility function, and a second sub-network based on the Tanh activation function, whose input depends on the activation mode of the first sub-network and is used to capture the discontinuity of the server utility function.

9. The system according to claim 7, wherein: The verification and aggregation module is further configured to, when performing the weighted semi-asynchronous aggregation operation, weight the weighted semi-asynchronous aggregation operation According to the staleness of the model uploaded by the client and client data quality Calculated, specifically ,in Represents the staleness effect of client-side latency, represents the data quality impact of the client, and B is the client set of the current aggregation batch.

10. The system according to claim 7, wherein: The client classification and contract form determination module is further configured to classify the client's data quality into X levels, the computing resources into Y levels, the privacy budget corresponding to the privacy preference into Z levels, and the client combination type is (x, y, z), where 1≤x≤X, 1≤y≤Y, 1≤z≤Z; the contract form is expressed as ,in is the number of local iterations, Contributing to clients of type (x,y,z) The total reward obtained after local iterations.

Images