Dynamic committee consensus method for decentralized federated learning

Abstract

The invention discloses a decentralized federated learning-oriented dynamic committee consensus method, which belongs to the technical field of machine learning, and comprises the following steps of: firstly, establishing a node utility function according to node training income, consensus income and corresponding cost; each node selects whether to join a committee according to a utility maximization principle; the game process is subjected to scale and reward adjustment, strategy updating and member selection stage iteration until the committee composition meets the utility maximization and Nash equilibrium conditions; after the committee is dynamically elected, nodes in the committee become a committee leader through aggregation local model updating competition, block-out right and block-out rewards are obtained, improvement of the global model performance is evaluated according to the nodes, the node with the best improvement effect on the global model accuracy becomes the leader, and the node with the best improvement effect on the global model accuracy becomes the leader. While the interests of the nodes are ensured, the utility of the committee is maximized, the nodes are motivated to participate in the federated learning process for a long time, the global model training quality is improved, and the federated learning process is accelerated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning, and particularly relates to a dynamic committee consensus method for decentralized federated learning. BACKGROUND

[0002] As a distributed machine learning paradigm for protecting data privacy, federated learning avoids direct transmission of raw data by decentralizing model training tasks to local nodes, ensuring data security while enabling cross-institutional collaborative modeling. However, traditional federated learning frameworks rely on a central server to coordinate training and aggregation between nodes, and their centralized architecture not only has a single point of failure risk, but also can cause performance bottlenecks and system instability.

[0003] The development of blockchain technology provides new technical support for solving the problems in traditional federated learning. The decentralization, transparency, and security of blockchain effectively address some security challenges in federated learning. Through the blockchain consensus mechanism, local model updates and global models are verified and recorded, freeing the system from dependence on a central server and improving the robustness and security of the system.

[0004] However, traditional blockchain consensus mechanisms such as proof of work and proof of stake often require significant communication costs to reach consensus between nodes, especially when there are a large number of participating nodes that need to handle a large number of local model updates, resulting in significant resource waste and efficiency bottlenecks.

[0005] To this end, the committee consensus mechanism is proposed, which consists of a committee composed of a part of high-performance nodes elected, and the model verification and record are only conducted within the committee, which greatly reduces the resource overhead and computational complexity of the whole network consensus, while maintaining a certain degree of decentralization. First, the traditional method usually adopts random selection or fixed index evaluation when selecting committee members. Random selection cannot guarantee the resource performance of committee members, thereby affecting the consensus efficiency; fixed index evaluation to elect committee members may lead to nodes with resource advantages participating in the consensus process for a long time, thereby reducing the decentralization of the system, leading to low-performance nodes or potential malicious nodes entering the committee, thereby reducing the overall efficiency of the system. Second, the existing committee consensus method lacks a flexible dynamic adjustment mechanism, which cannot adapt to the changes in the network environment in real time and the problem of node heterogeneity network environment or node resource changes. Although the regular replacement of committee members can achieve a certain degree of dynamics, this updating method is usually preset and lacks the ability to make immediate adjustments according to the current system state or node performance changes, which may not effectively cope with complex dynamic network environments. Finally, the existing committee consensus mechanism tends to focus on short-term contributions of nodes in incentive design, while ignoring the establishment of long-term incentive mechanisms, which may lead to an increase in the frequency of node exit from the committee, affecting the stability and sustainability of the system. SUMMARY

[0006] To achieve the above purpose, the technical scheme of the present application is as follows: a dynamic committee consensus method for decentralized federated learning, in the committee selection stage, the node selection is optimized through utility analysis and game theory, improving the overall performance and utility of the committee. In the consensus stage, the global model is optimized and evaluated based on the competitive aggregation method, ensuring that the selection of the leader node can guarantee the accuracy and robustness of the global model, thereby effectively improving the overall performance of the system. Therefore, the present application provides a reliable and efficient solution for the application of federated learning in complex dynamic environments, which helps to promote the further development and landing of federated learning technology. The method comprises the following steps:

[0007] First, a node utility function is established according to the node training income, consensus income and corresponding cost; each node selects whether to join the committee according to the utility maximization principle; the game process is iterated through scale and reward adjustment, strategy update and member selection stage until the committee composition meets the utility maximization and Nash equilibrium conditions;

[0008] After the dynamic election of the committee, the nodes in the committee update the competition by aggregating the local model to become the committee leader, obtain the block right and block reward, and evaluate the performance of the node according to the performance of the global model, and the node with the best improvement effect on the accuracy of the global model becomes the leader.

[0009] As an improvement to this invention, the utility of a node when it does not join the committee is defined as the difference between the training reward and the training cost. The utility of a node when it joins the committee is defined as the difference between its reward and its cost, wherein the reward includes both training reward and consensus reward, and the cost includes both training cost and consensus cost.

[0010] (1),

[0011] (2),

[0012] in, Training rewards in federal mission transactions. For nodes Number of data samples available This represents the total number of data samples owned by all participating nodes in the federated training. For nodes Resource weight, For nodes CPU cycle frequency, For nodes The number of CPU cycles required to train unit data is typically... Represented as nodes Cost of training unit data , Consensus rewards in federal mission transactions. The pre-defined committee size for federal mission transactions. For nodes The unit communication cost.

[0013] As an improvement to this invention, committee utility is defined as the difference between committee satisfaction and committee consensus cost. Committee consensus satisfaction is affected by two factors: the overall computing resources available to the committee and consensus latency.

[0014] (3),

[0015] (4),

[0016] in, To satisfy the consensus of the committee, For the committee consensus cost, For nodes The computing resources available For nodes Communication delay, For the sake of the overall computing resources owned by the committee, For the committee consensus delay, the more computing resources the committee has, the smaller the consensus delay, and the higher the committee satisfaction.

[0017] As an improvement of the application, before the game, all nodes participating in the federal training are comprehensively scored, only the nodes with a comprehensive score higher than the average value can obtain the qualification to join the committee to participate in the game process,

[0018] (5),

[0019] Among them, The communication delay of the node The first The quality of the local training model, The marginal contribution of the node The first The communication delay of the node The communication delay of the node The communication delay of the node The weight of each part respectively.

[0020] As an improvement of the application, according to the preset committee size m, update frequency n and reward in the federal task transaction, the game process is started, including the following three stages:

[0021] Stage 1, game selection stage, all nodes participating in this federal learning task calculate the utility of joining the committee and not joining the committee according to the formula and compare them, make the corresponding selection strategy, and record the number of nodes selected out and joined the committee;

[0022] Stage 2, size adjustment stage, when the number of selected nodes is less than half of the committee size, randomly select nodes from the committee to move out of the committee until the number of exit nodes is equal to half of the committee size;

[0023] Stage 3, member selection stage, the nodes selected to join the committee are added to the committee candidate set, the committee utility gain brought by each node in the committee candidate set after joining the committee is calculated according to the committee utility formula and ranked in descending order, the node with the largest committee utility gain is selected to join the committee, until all nodes do not change the decision and the number of committee members reaches m, the game ends.

[0024] As an improvement of the application, when the number of nodes selected to join the committee is less than the number of nodes selected to exit the committee, according to the game definition, When

[0025] (6),

[0026] (7),

[0027] When the number of nodes choosing to join the committee is less than the number of committee nodes choosing to leave, the committee size will be adjusted according to the formula below.

[0028] (8),

[0029] in, for The committee size is pre-set for each round. The adjustment range is dynamically adjusted according to the needs of the federated mission. The final adjustment value is stipulated to not exceed one-fifth of the preset size in the mission transaction. Then, the node game selection process is carried out again according to the adjusted size. If the required number is still not reached after the adjustment, the consensus reward will be increased in this round to meet the needs of the federated mission.

[0030] As an improvement to this invention, the steps for committee members to complete the on-chain and update of the global model through committee consensus based on competitive aggregation include:

[0031] (9),

[0032] in, For competing nodes No. Global model gain of round-competitive aggregation. For the first wheel node The accuracy of the aggregated global model. For the first The accuracy of the global model.

[0033] In each round, competing nodes aggregate the global model based on their local model updates. Different nodes use different aggregation methods. For each competing node, the gain of its aggregated model on the global model is calculated according to the formula. Based on the gain index, the node with the largest gain is selected as the leader of the committee for this round. The leader node is responsible for submitting the final global model, packaging the block, and broadcasting it to other committee members for verification.

[0034] As an improvement of the present invention, during the committee consensus phase, each member node independently verifies and votes on the aggregated model submitted by the candidate leader node. The verification includes model performance indicators, data integrity, and the degree of matching between the model and the task objective. Through the distributed voting process of member nodes, a consensus is reached. The aggregated model that passes the consensus will be released as a global model and used in the next round of federated learning tasks.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1) The application designs a committee node selection mechanism based on repeated game. A repeated game model based on utility is constructed, and nodes dynamically adjust whether to join the committee for game according to training rewards, consensus rewards, communication and calculation costs, realize committee composition under Nash equilibrium, and ensure system stability and node enthusiasm. And support dynamic adjustment of committee size, adapt to different task requirements and network environment.

[0037] 2) The application designs a dynamic committee consensus mechanism based on competitive aggregation. Based on dynamic selection of committee members, the global model accuracy improvement is introduced as a competition index in the committee, and nodes compete with different aggregation strategies to select the leader node to execute model aggregation and block packaging, so as to directly link the consensus process with the federal learning goal, and significantly improve the model accuracy and convergence speed. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the overall framework diagram of the application;

[0039] Figure 2 is the overall flowchart of the application;

[0040] Figure 3 is the committee election flowchart of the application. DETAILED DESCRIPTION

[0041] The application will be further illustrated below in combination with the drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the application and not to limit the scope of the application.

[0042] EMBODIMENT

[0043] The embodiment proposes a dynamic committee consensus method for decentralized federated learning, which maximizes committee utility while ensuring node self-interest, long-term incentives for nodes to participate in federated learning process, improves global model training quality and speeds up federated learning process. Dynamic committee consensus based on repeated game mainly includes two parts, one part is to select committee members from participants through committee election method based on repeated game, and the other part is to complete the on-chain and update of global model by committee members through committee consensus based on competitive aggregation.

[0044] The decentralized federated learning architecture proposed by the application is shown in Figure 1 . It mainly consists of four parts: task publisher, participant, committee and blockchain. Specifically, the task publisher starts the entire federated learning process by publishing a federated learning task transaction, then the participants perform local training according to the federated learning task published by the task publisher, and submit local model updates to the blockchain, the committee is responsible for aggregating all local updates recorded on the chain, and submitting the global model to the blockchain. Each part will be described in detail below.

[0045] (1) Task publisher: Publishes a federated learning task transaction according to the demand, provides key information of the task, including an initial global model of federated learning, a federated learning task budget, wherein the budget of the training process and the consensus process is respectively paid to the participants participating in the training process and the consensus process, and the committee size and the update frequency are set.

[0046] (2) Participants: The participating nodes are the same as the client nodes in the traditional federated learning, and the participating nodes have their own local data, which are usually highly sensitive and private. In the federated learning process, the participating nodes do not need to upload the local data, but use the local data to perform local model training according to the global model, calculate the local gradient or parameter update and upload.

[0047] (3) Committee: According to the committee size set in the federated learning task transaction and the performance of the participants, the committee is dynamically elected to participate in the consensus, and the leader node is elected in the committee through the competitive aggregation, which is responsible for the model aggregation and uploading. In addition, the committee is updated according to the committee update frequency set in the federated learning task transaction and the node training performance.

[0048] (4) Blockchain: The blockchain replaces the central server in the traditional federated learning to store information such as local model update and global model update. All participants jointly maintain the blockchain ledger, automatically reach an agreement through a consensus algorithm, remove the trust demand for the central server, increase the transparency and reliability of the system. In addition, the non-tamperability of the blockchain ensures that the model parameters and transaction records cannot be maliciously modified or forged, and the encrypted transactions cannot be stolen by third parties, ensuring the privacy and security of the data.

[0049] The overall process of the decentralized federated learning proposed in the application is as follows: Figure 2The task publisher publishes a federated learning task transaction through the blockchain network first, and the participants receive the task and use the local data set for local training and submit the local model update to the blockchain. After the participants submit the local model to the blockchain, the nodes are first comprehensively scored to select nodes qualified to join the committee, and then a committee election method based on repeated game is used to elect committee nodes from the qualified nodes. According to the committee size set in the federated learning task transaction and the performance of the participating nodes, the nodes are selected to participate in consensus through a dynamic election method based on repeated game. Game theory provides a framework for analyzing and optimizing participant behavior, especially in the face of multi-party interest conflicts, heterogeneous participants and potential malicious nodes. It can effectively solve the problems of committee dynamics and consensus efficiency in decentralized federated learning. Moreover, nodes can adjust their strategies and resource allocation according to their actual utility in the early stage, which effectively provides long-term incentives and encourages nodes to adopt more reasonable strategies in multiple rounds of game to maintain the stability and efficiency of federated learning. In the process of committee members submitting the global model to the blockchain, a competitive aggregation-based method is used to achieve committee consensus. Within the committee, the leader node is elected by competitive aggregation of the global model, which is responsible for model aggregation and uploading to the blockchain. At the same time, the computing resources of the committee members are reasonably utilized, and the committee consensus is reached after verification, effectively improving the performance of federated learning.

[0050] Further, the method comprises:

[0051] I. Dynamic committee election based on repeated game. The committee election flowchart of the present application is shown in Figure 3 .

[0052] First, the node utility function is established according to the node training income, consensus income and corresponding cost; each node selects whether to join the committee according to the utility maximization principle; the game process is iterated through the scale and reward adjustment, strategy update and member selection stage until the committee composition meets the utility maximization and Nash equilibrium conditions. The node utility and committee utility are maximized, and the node participation enthusiasm, system stability and consensus efficiency are improved.

[0053] 1. The utility analysis step includes:

[0054] 1) Node utility analysis:

[0055] In federated learning, participating nodes can choose to join a committee or not. Nodes that do not join a committee only receive training rewards and bear the corresponding training costs. Nodes that choose to join a committee receive additional consensus rewards and may compete to become a leader node and block producer, gaining the right to record transactions. However, nodes also incur consensus costs when joining a committee and participating in consensus. Therefore, to maximize their utility, nodes need to comprehensively consider training rewards, consensus rewards, their training costs, and consensus costs to make a reasonable choice.

[0056] The training reward a node receives in federated learning is primarily determined by the amount of data it possesses, while its training overhead is mainly determined by its own performance. A node's utility when not in a committee is defined as the difference between its training reward and training cost. The consensus reward a node receives is primarily determined by the pre-defined committee size and consensus reward in the federated task transaction, while the node's consensus overhead mainly considers its communication costs. A node's utility when in a committee is defined as the difference between its reward and its overhead. The reward includes both training and consensus rewards, while the overhead includes both training and consensus costs.

[0057] (1),

[0058] (2),

[0059] in, Training rewards in federal mission transactions. For nodes Number of data samples available This represents the total number of data samples owned by all participating nodes in the federated training. For nodes Resource weight, For nodes CPU cycle frequency, For nodes The number of CPU cycles required to train unit data is typically... Represented as nodes Cost of training unit data . Consensus rewards in federal mission transactions. The pre-defined committee size for federal mission transactions. For nodes The unit communication cost.

[0060] 2) Committee utility analysis:

[0061] In decentralized federated learning, consensus speed directly impacts the efficiency and convergence speed of model training. In conventional consensus mechanisms, all nodes participate, resulting in high computational and communication overhead and latency. Committee consensus mechanisms, however, involve only committee nodes in the consensus process, improving overall system efficiency and throughput. However, the resources available to the committee as a whole directly determine its processing power and consensus efficiency. Therefore, committee utility is defined as the difference between committee satisfaction and committee consensus cost. Committee consensus satisfaction is primarily influenced by two factors: the computational resources available to the committee as a whole and consensus latency.

[0062] (3),

[0063] (4),

[0064] in, To satisfy the consensus of the committee, The cost of committee consensus is mainly considered in terms of communication costs between nodes. For nodes The computing resources available For nodes Communication delay, For the sake of the overall computing resources owned by the committee, This refers to the consensus latency of the committee. Generally, the more computing resources a committee has, the lower the consensus latency, and the higher the committee's satisfaction level.

[0065] 2. Utility-based dynamic alliance formation game:

[0066] In this embodiment, the selection process of committee members is modeled as a utility-based dynamic alliance formation game.

[0067] To ensure the overall performance of federated learning and guarantee its efficiency and stability, all nodes participating in the federated training are comprehensively evaluated before the game begins, thus ensuring the performance of committee members. Only nodes with a comprehensive score above the average are eligible to join the committee and participate in the game process. The comprehensive evaluation of nodes mainly considers factors such as communication latency, marginal contribution, and the quality of local model training.

[0068] (5),

[0069] in, For nodes No. Regarding the quality of locally trained models, For nodes No. Wheel marginal contribution, For nodes communication delay, The weight of each part is shown in the table.

[0070] The game of dynamic alliance formation based on utility for each federated learning task is defined as follows:

[0071] Player: the node whose comprehensive score is higher than the average of the comprehensive scores of all nodes participating in this federated learning task.

[0072] Utility: the utility of the node when joining the committee and not joining the committee is calculated by the corresponding formula.

[0073] Strategy: if the node does not join the committee, the node's optional strategies are to join the committee or not to join the committee; if the node has joined the committee, the node's optional strategies are to quit the committee or to stay in the committee.

[0074] According to the committee size m, the update frequency n and the reward in the federated task transaction, the game process is started. It includes the following three stages:

[0075] Stage 1, game selection stage. All nodes participating in this federated learning task calculate their utility of joining the committee and not joining the committee according to the formula and compare them, make the corresponding selection strategy. And record the number of nodes that choose to exit and join the committee.

[0076] Stage 2, size adjustment stage. In order to ensure the fairness of the committee election and prevent the node from monopolizing the committee, this method provides that if the number of nodes that choose to exit is less than half of the committee size, a certain number of nodes in the committee will be randomly selected to move out of the committee to ensure that half of the nodes exit the committee.

[0077] However, at this time, the number of nodes that choose to join the committee may be less than the number of committee nodes that choose to exit. According to the game definition, when , the node chooses to join the committee.

[0078] (6),

[0079] (7),

[0080] Therefore, the strategy of the node is mainly related to the consensus reward proposed by the federated task and the committee size. Therefore, when the number of nodes that choose to join the committee is less than the number of committee nodes that choose to exit, the committee size will be adjusted according to the following formula.

[0081] (8),

[0082] wherein, is Round preset committee size, To adjust the amplitude, it can be adjusted dynamically according to the federal task demand. The final adjustment value is not more than one fifth of the preset size in the task transaction. Then the node game selection process is carried out again according to the adjusted size. If the adjustment size still cannot meet the specified number, the consensus reward of this round will be appropriately increased to meet the demand of federal task.

[0083] Stage 3, member selection stage. The nodes selected to join the committee are added to the committee candidate set. According to the committee utility formula, the committee utility gain brought by each node in the committee candidate set after joining the committee is calculated and ranked in descending order. The node that brings the greatest utility gain to the committee is selected to join the committee, until all nodes do not change the decision and the number of committee members reaches m, and the game ends.

[0084] II. Committee consensus based on competition aggregation:

[0085] After the dynamic election of the committee, the nodes in the committee become the committee leaders by aggregating the local model updates, obtain the block right and block reward. According to the evaluation of the node to the global model performance, the node with the best improvement effect on the global model accuracy becomes the leader, so as to ensure that the selected leader node can maximize the improvement of the global model quality.

[0086] (9),

[0087] Among them, The competition node The first The global model gain of the competition aggregation in the round, The global model accuracy aggregated by the node in the round, The global model accuracy in the round. In each round, the competition node aggregates the global model according to the local model update. Different nodes use different aggregation methods, such as FedAvg, FedProx, adaptive aggregation, etc. For each competition node, according to the formula, the gain of its aggregated model on the global model is calculated, based on the gain index, the node with the greatest gain is selected as the committee leader in this round. The leader node is responsible for submitting the final global model, packaging the block, and broadcasting it to other members of the committee for verification.

[0088] In each round, the competition node aggregates the global model according to the local model update. Different nodes use different aggregation methods, such as FedAvg, FedProx, adaptive aggregation, etc. For each competition node, according to the formula, the gain of its aggregated model on the global model is calculated, based on the gain index, the node with the greatest gain is selected as the committee leader in this round. The leader node is responsible for submitting the final global model, packaging the block, and broadcasting it to other members of the committee for verification.

[0089] ​​To prevent resource waste and avoid all committee nodes participating in leader election, the method limits the nodes using different aggregation methods to participate in leader election to control resource consumption. Meanwhile, to improve the efficiency of federated learning and ensure the performance of federated learning, the application introduces the concept of competition rounds. According to the federated task demand and model performance, the competition rounds can be appropriately adjusted. In the competition rounds, nodes use different aggregation methods for competitive aggregation. After multiple rounds of competition, the aggregation method with the best performance in the early competition is selected to be used in the subsequent global model update, until the federated training is completed.

[0090] In the committee consensus phase, each member node independently verifies and votes on the aggregation model submitted by the candidate leader node. The verification content includes model performance indicators, data integrity, and the matching degree of the model and the task target. Through the distributed voting process of the member nodes, a consensus is reached to ensure that the final selected aggregation model has high performance and robustness. The aggregation model that passes the consensus will be published as the global model and used for the next round of federated learning task. This process can effectively avoid the introduction of low-quality models and the potential risk of malicious aggregation attacks, thereby ensuring the integrity of the data and the accuracy of the global model.

[0091] The embodiment can better adapt to changes in network environment and reduce consensus delay by dynamically selecting high-performance nodes to form a committee through a game model. In the committee, competitive aggregation is performed to select the optimal leader node to be responsible for the aggregation and update of the global model, which significantly improves the overall performance and stability of federated learning.

[0092] Training phase:

[0093] Step 1: The task publisher publishes a federated learning task transaction through the blockchain network, including the initial global model, committee size, training and consensus rewards, etc.

[0094] Step 2: Local nodes register in the blockchain to enter the federated learning framework and join the blockchain network.

[0095] Step 3: Nodes use local data sets for training, and after training, broadcast model parameters, self-resource and delay information through the blockchain.

[0096] Consensus phase:

[0097] Step 1: The smart contract in the blockchain calculates the comprehensive score according to the model update and performance of the nodes, and the nodes with a comprehensive score greater than the average value have the opportunity to join the committee.

[0098] Step 2: Nodes with the opportunity to join the committee calculate the utility according to the reward and their own cost to decide whether to join the committee

[0099] Step 3: If the game does not reach equilibrium, the nodes make decisions again after adjusting the committee size or consensus reward according to the node decision.

[0100] Step 4: The members in the committee compete for the leader node by aggregating the global model in different ways, and the node with the highest accuracy of the global model becomes the block node in this round, and the global model is packaged into the candidate block and broadcast to other committee members for verification.

[0101] Step 5: The member nodes independently verify and vote on the aggregated model submitted by the candidate leader node, and a consensus is reached when more than 2 / 3 of the members vote.

[0102] Then repeat the training and consensus until the global model meets the requirements of the task publisher.

[0103] It should be noted that the above content only illustrates the technical idea of the present application and cannot limit the protection scope of the present application. For ordinary skilled persons in the technical field, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements fall within the protection scope of the claims of the present application.

Claims

1. A dynamic committee consensus method for decentralized federated learning, characterized in that, The method comprises the following steps: First, according to the node training income, consensus income and corresponding cost, the node utility function is established; each node selects whether to join the committee according to the utility maximization principle; the game process is iterated through scale and reward adjustment, strategy updating and member selection stage until the committee composition meets the utility maximization and Nash equilibrium conditions; After the dynamic election of the committee, the nodes in the committee compete to become the committee leader by aggregating the local model updates, obtain the block generation right and block generation reward, and evaluate the performance of the global model according to the performance of the global model, and the node with the best improvement effect on the global model accuracy becomes the leader.

2. The method of claim 1, wherein, The utility of the node when not joining the committee is defined as the difference between the training reward and the training cost, and the utility of the node when joining the committee is defined as the difference between the reward obtained and the consumption cost, wherein the reward obtained includes training reward and consensus reward, and the consumption cost includes training cost and consensus cost, （1）， （2）， wherein, is a node utility when not joining the committee, is a node training reward obtained, is a node training cost, is a training reward in a federated task transaction, is a node number of data samples owned, represents the total sum of the amount of data samples owned by all participating federated training nodes, is a node resource weight, is a node CPU cycle frequency, is a node number of CPU cycles required for training unit data, is a node utility when joining the committee, is a node consensus reward obtained by participating in consensus, is a node consensus cost paid for participating in the consensus process, is a consensus reward in a federated task transaction, is a preset committee size in a federated task transaction, is a node unit communication cost.

3. The method of claim 2, wherein, The committee utility is defined as the difference between the committee satisfaction and the committee consensus cost, and the committee consensus satisfaction is affected by the computing resources owned by the committee and the consensus time delay, （3）， （4）， wherein, is the committee utility, is the committee consensus satisfaction, is the committee consensus cost, is the number of committee members, is the node owns computing resources, is the node communication delay, is the total sum of computing resources owned by the committee as a whole, is the consensus latency of the committee, the more computing resources the committee owns, the smaller the consensus latency, and the higher the committee satisfaction.

4. The method of claim 3, wherein, Before the game, all nodes participating in the federal training are scored comprehensively, only the nodes with a comprehensive score higher than the average value can obtain the qualification to join the committee and participate in the game process, （5）， wherein, is a comprehensive score of the node , is a comprehensive score of the node , is a local training model quality of the node in the first round, is a marginal contribution of the node in the first round, is a communication latency of the node , is a weight of each part, respectively.

5. The method of claim 4, wherein, According to the preset committee size m, update frequency n and reward in the federal task transaction, the game process is started, which comprises the following three stages: Stage 1, game selection stage, all nodes participating in the federal learning task calculate the utility of joining the committee and not joining the committee according to the formula and compare them, make the corresponding selection strategy, and record the number of nodes selected out of the committee and the number of nodes selected into the committee; Stage 2, size adjustment stage, when the number of selected out nodes is less than half of the committee size, randomly select nodes from the committee to move out of the committee until the number of selected out nodes is equal to half of the committee size; Stage 3, member selection stage, the nodes selected to join the committee are added to the committee candidate set, the committee utility gain brought by each node in the committee candidate set after joining the committee is calculated according to the committee utility formula and ranked in descending order, the node bringing the maximum utility gain to the committee is selected to join the committee, and the game ends when all nodes do not change the decision and the number of committee members reaches m.

6. The method of claim 5, wherein, In the scale adjustment phase, when the number of nodes that choose to join the committee is less than the number of committee nodes that choose to exit, according to the game definition, when the node chooses to join the committee, （6）， （7）， When the number of nodes selected to join the committee is less than the number of nodes selected to exit the committee, the committee size is adjusted according to the following formula, （8）， Wherein, is The round preset committee size, is the adjustment range, dynamically adjusted according to the federal task demand, and the final adjustment value is not more than one fifth of the preset size in the task transaction. Then, according to the adjusted size, the node game selection process is carried out again. If the adjusted size still cannot meet the specified number, the consensus reward of this round will be increased to meet the demand of the federal task.

7. The method of claim 6, wherein, The committee members complete the steps of uploading and updating the global model through the committee consensus based on competition aggregation, which comprises: （9）， wherein, is a competing node the round aggregated global model gain, is the round node accuracy of the aggregated global model, is the accuracy of the global model of the round; In each round, the competing nodes aggregate the global model according to the local model updates, and the aggregation methods used by different nodes are different. For each competing node, the gain of the aggregated model on the global model is calculated according to the formula, and based on the gain index, the node with the maximum gain is selected as the committee leader in this round. The leader node is responsible for submitting the final global model, packaging the block, and broadcasting it to other committee members for verification.

8. The method of claim 7, wherein, In the committee consensus phase, each member node independently verifies and votes on the aggregation model submitted by the candidate leader node, including model performance indicators, data integrity, and the degree of matching between the model and the task target. Through the distributed voting process of the member nodes, a consensus is reached, and the aggregation model that passes the consensus will be published as a global model and used for the next round of federated learning tasks.

Images