A node-centralized worker node task allocation method

Abstract

This invention belongs to the technical field of task allocation for worker nodes in federated learning systems, specifically relating to a method for task allocation of worker nodes oriented towards node centralization. This invention can effectively incentivize worker nodes of varying quality to remain in the federated learning system, ensuring a large number of high-quality worker nodes exist within the system over a long period, thus addressing the problem of node centralization hindering the long-term development of the federated learning system. This invention aims to attract more worker nodes of different qualities to join the federated learning system, avoiding node centralization and promoting the long-term development of the federated learning system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of task allocation of worker nodes in a federated learning system, and particularly relates to a worker node task allocation method for node centralization. BACKGROUND

[0002] In a federated learning system, how to allocate appropriate federated learning tasks to worker nodes of different qualities to avoid node centralization has always been a key problem. In order to maximize their own benefits and obtain better model results, task publishers tend to choose high-quality worker nodes to participate in their federated learning tasks. In many models, reputation is an important indicator for evaluating the quality of worker nodes, and the higher the reputation of a worker node, the more likely it is to bring high-quality and reliable training results to federated learning. After each round of federated learning task ends, the reputation of the worker node is updated according to its behavior, which is used as an evaluation indicator for the selection of worker nodes in the next federated learning task.

[0003] Reputation as an indicator for evaluating the quality of worker nodes has good effects, which can help task publishers to select worker nodes that meet the requirements to complete federated learning tasks and improve the quality of models. However, task publishers tend to choose worker nodes with high reputation, which will lead to the fact that other worker nodes in the federated learning system have no opportunity to participate in federated learning tasks, i.e., no opportunity to improve their reputation. Rational worker nodes will not stay in the system in order to obtain benefits. In the long run, the quality distribution of worker nodes in the federated learning system is uneven, and the number of worker nodes cannot be guaranteed, which will damage the interests of the entire system. SUMMARY

[0004] The purpose of the present application is to provide a worker node task allocation method for node centralization, which can attract more worker nodes of different qualities to join federated learning, avoid node centralization problems, and promote the long-term development of the federated learning system.

[0005] A worker node task allocation method for node centralization, comprising the following steps:

[0006] 1) Task initialization stage: the task publisher publishes M federated learning tasks to the platform, and pre-pays a corresponding reward R for each task m m ; the platform publishes these federated learning tasks to the worker nodes in the system and recruits worker nodes to actively participate in the tasks to obtain rewards, wherein the reputation score of each worker node is recorded in the platform;

[0007] 2) Task allocation phase: worker nodes in the federated learning system autonomously select a federated learning task to participate in and inform the platform of the decision; after the platform collects the decisions of the worker nodes in the system, the platform calculates the rewards of each worker node according to the weighted allocation reward method and informs each worker node of the expected reward; after the worker nodes receive the message, they can choose to change their strategies according to the reward they get; if they choose to change the strategy, they will send the new decision to the platform; the platform collects the decisions of the worker nodes at the end of each round and recalculates the rewards of the worker nodes; when there is no worker node in the system to change its strategy, the task selection phase ends;

[0008] 3) Task end phase: the worker nodes in the federated learning system execute the federated learning task according to the strategy in the task selection phase and send the results to the platform; the platform allocates rewards to the worker nodes according to the results calculated in the task selection phase, and sends the training results of the worker nodes to the task publisher; the task publisher aggregates the training results of the worker nodes and feeds back the results to the platform, and the platform updates the reputation scores of the worker nodes according to the feedback

[0009] Further, in the task allocation phase, first, the reputation score is converted into a reputation weight using a sigmoid activation function to filter worker nodes with poor quality and amplify the differences in reputation scores between different worker nodes, so as to allocate rewards according to the quality of the worker nodes after the federated learning task and improve the efficiency and fairness of the system, and the formula of the sigmoid activation function is:

[0010]

[0011] The core idea of the task selection algorithm is to quickly reach Nash equilibrium by multiple rounds of selection and allowing one worker node to change its strategy in each round; at the beginning of each round, the worker nodes in the system inform the platform of the task selection strategy, the platform aggregates the strategies of all the worker nodes, then calculates the sum of the reputation weights of each federated learning task and returns the result to the worker nodes; after receiving the result, the worker nodes calculate the rewards they can get according to the weighted allocation reward method and inform the platform of the strategy in the next round; if there is a worker node in the system that changes the strategy in the next round, the next round of task selection continues; otherwise, the worker nodes in the system reach a consensus on the task selection strategy, and the task selection algorithm ends; in each round of task selection, the strategy is updated through two steps:

[0012] (2.1) The task selection strategy of worker node i in the rth round is represented as After receiving the task selection strategy of each worker node, the platform counts the participation of the worker nodes in each task in the system, calculates the sum of the reputation weights of the worker nodes participating in each federated learning task, and returns the result as a vector group returns to the worker node in the form of a vector group G

[0013] (2.2) The worker node finds the federal learning task that can maximize its own benefit according to the vector group G returned by the platform r The next round of participating in each federal learning task can obtain the benefit of each, that is:

[0014]

[0015] Through the above formula, the worker node i can find the federal learning task that can maximize its own benefit in the next round If It means that the best strategy has been found in the task selection of the rth round, and the platform sends to indicate that the strategy of the worker node i does not change; otherwise, the worker node i will update its task selection strategy in the next round to obtain higher benefit; when the platform receives the update request of all worker nodes, it randomly selects an updated worker node strategy and sends a confirmation response to it; next, the platform re-aggregates the worker node strategy of the next round and returns the calculation result G r+1 to the worker nodes in the system; repeat the above process until , that is, all worker nodes select the federal learning task that maximizes their own benefit, and do not change the strategy in the next round; this means that the non-cooperative game of task selection reaches Nash equilibrium, and the worker nodes in the system confirm and then execute the federal learning task according to their own strategy;

[0016] Use θ min and θ max to represent the minimum and maximum values of the reputation weight of the worker node respectively, use R min and R max to represent the minimum reward and maximum reward of the federal learning task in the system respectively, and c min represents the minimum cost of the worker node, then the number of rounds that the task selection algorithm can perform at most is:

[0017]

[0018] The beneficial effects of the present application are:

[0019] The present application can effectively encourage worker nodes of different qualities to stay in the federal learning system, so that a large number of high-quality worker nodes exist in the system for a long time, solving the problem of node centralization in the federal learning system that damages the long-term development of the system. The present application can attract more worker nodes of different qualities to join the purpose of federal learning, avoid the problem of node centralization, and promote the long-term development of the federal learning system. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a structural diagram of a federal learning system.

[0021] Figure 2 is the reputation weight conversion effect diagram in the application.

[0022] Figure 3 is the task selection algorithm flow chart in the application. DETAILED DESCRIPTION

[0023] The application will be further described below with reference to the drawings.

[0024] The application proposes a worker node incentive mechanism based on reputation weight to solve the problem of worker node task allocation strategy facing node centralization. In order to ensure the fairness of reward distribution, the incentive mechanism needs to ensure that worker nodes with different reputations in the federated learning system have the opportunity to participate in federated learning tasks, maintain the number of worker nodes in the federated learning system to a certain scale, and eliminate the head effect and bystander effect. In addition, the incentive mechanism also needs to have the ability to eliminate worker nodes with poor quality, and ensure the efficiency of federated learning tasks in the system.

[0025] 1) In the federated learning system, the task publisher publishes training tasks and rewards corresponding to each task to the platform, and the platform publishes these information to the worker nodes in the system.

[0026] 2) The worker nodes in the federated learning system autonomously select a federated learning task to participate in and inform the platform of the decision.

[0027] 3) After the platform collects the decisions of the worker nodes in the system, it calculates the expected reward of each worker node using reputation weight, and returns the result to the worker nodes.

[0028] 4) After receiving the message, the worker nodes can choose whether to change their strategy according to the reward they get. If they choose to change the strategy, they will send the new decision to the platform.

[0029] 5) After each round, the platform collects the decisions of the worker nodes and recalculates the rewards of the worker nodes. When there is no worker node in the system to change its strategy, the task selection stage ends.

[0030] 6) The worker nodes execute the federated learning task according to the strategy and send the result to the platform. The platform allocates rewards to the worker nodes according to the reputation weight, and sends the training results of the worker nodes to the task publisher. The task publisher aggregates the training results of the worker nodes and feeds back the results to the platform, and the platform updates the reputation score of the worker nodes according to the feedback.

[0031] As Figure 1As shown, the federated learning system is composed of a task platform, a task publisher and workers. The system execution process can be divided into three stages according to the cycle, namely the task initialization stage, the task selection stage and the task end stage.

[0032] Task initialization stage: the task publisher publishes M federated learning tasks to the platform, and pre-pays the corresponding rewards R for each task m m The platform publishes these federated learning tasks to the worker nodes in the system and recruits the worker nodes to actively participate in the tasks to obtain rewards, wherein the reputation score of each worker node is recorded in the platform.

[0033] Task allocation stage: the worker nodes in the federated learning system autonomously select a federated learning task to participate in and inform the platform of the decision. After the platform collects the decisions of the worker nodes in the system, the expected rewards of each worker node are calculated according to the weighted allocation method, and each worker node is informed of the expected rewards. After receiving the message, the worker node can choose to change its strategy according to the rewards it gets. If the strategy is changed, the new decision is sent to the platform. After each round, the platform collects the decisions of the worker nodes and recalculates the rewards of the worker nodes. When there is no worker node in the system to change its strategy, the task selection stage ends.

[0034] Task end stage: the worker nodes in the federated learning system execute the federated learning task according to the strategy in the task selection stage and send the results to the platform. The platform allocates rewards to the worker nodes according to the results calculated in the task selection stage, and sends the training results of the worker nodes to the task publisher. The task publisher aggregates the training results of the worker nodes and feeds back the results to the platform, and the platform updates the reputation score of the worker nodes according to the feedback.

[0035] In the task allocation stage, the reputation score needs to be converted into reputation weight using the sigmoid activation function first to filter out worker nodes with poor quality, and to amplify the differences in reputation scores between different worker nodes, so as to allocate the benefits according to the quality of the worker nodes after the federated learning task, and improve the running efficiency and fairness of the system. The formula of the sigmoid activation function is as follows:

[0036]

[0037] Figure 2 The influence of different parameters on the above formula image is shown.

[0038] The core idea of the task selection algorithm is to quickly reach Nash equilibrium by multiple rounds of selection and allowing one worker node to change its strategy in each round. At the beginning of each round, the worker nodes in the system inform the platform of their task selection strategies. After the platform aggregates all the strategies of the worker nodes, it calculates the sum of the reputation weights of each federated learning task and returns the result to the worker nodes. After receiving the result, the worker nodes calculate the obtainable revenue according to the weighted revenue allocation method, and inform the platform of the strategy in the next round. If there is a worker node in the system that changes the strategy in the next round, the next round of task selection continues; otherwise, the worker nodes in the system reach a consensus on the task selection strategy, and the task selection algorithm ends. The overall process of the algorithm Figure 3 In each round of task selection, strategy updating is achieved through two steps.

[0039] (1) The task selection strategy of worker node i in the rth round is denoted as After receiving the task selection strategy of each worker node, the platform counts the participation of the worker nodes in each task in the system, and calculates the sum of the reputation weights of the worker nodes participating in each federated learning task, and returns the result to the worker nodes in the form of vector group .

[0040] (2) The worker node calculates the obtainable revenue of participating in each federated learning task in the next round according to the vector group G r returned by the platform, that is,

[0041]

[0042] Through the above formula, worker node i can find the federated learning task in the next round that can maximize its own revenue If , it means that the best strategy has been found in the rth round of task selection, and is sent to the platform to indicate that the strategy will not be changed. Otherwise, the worker node i will update its task selection strategy in the next round to obtain higher revenue. When the platform receives the update request of all worker nodes, it randomly selects to update the strategy of one worker node and sends a confirmation response to it. Next, the platform re-aggregates the strategies of the worker nodes in the new round and returns the calculation result G r+1 to the worker nodes in the system. Repeat the above process until , that is, all worker nodes have selected the federated learning task that maximizes their own revenue, and will not change the strategy in the next round. This means that the non-cooperative game of task selection has reached Nash equilibrium, and the worker nodes in the system confirm to start performing federated learning tasks according to their own strategies.

[0043] Use θ min and θ maxRmin and Rmax represent the minimum and maximum values of the reputation weight of the worker node, respectively, and R min and R max Rmin and Rmax represent the minimum and maximum values of the reputation weight of the worker node, respectively, and R min Rmin and Rmax represent the minimum and maximum values of the reputation weight of the worker node, respectively, and R

[0044]

[0045] The worker node task allocation method for node centralization provided by the present application can effectively encourage worker nodes of different qualities to stay in the federated learning system, so that a large number of high-quality worker nodes exist in the system for a long time, and the problem of node centralization in the federated learning system damaging the long-term development of the system is solved.

[0046] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for assigning worker node tasks in a node-centralized manner, characterized in that, Includes the following steps: 1) Task initialization phase: The task publisher releases M federated learning tasks to the platform and prepays a corresponding reward R for each task m. m The platform publishes these federated learning tasks to worker nodes in the system and recruits worker nodes to actively participate in the tasks to earn rewards. The reputation score of each worker node is recorded in the platform. 2) Task allocation phase: Worker nodes in the federated learning system autonomously select a federated learning task to participate in and inform the platform of their decision; after collecting the decisions of the worker nodes in the system, the platform calculates the income of each worker node according to the weighted distribution method and informs each worker node of the expected reward; after receiving the message, the worker nodes can choose whether to change their strategy based on the reward they have received. If a change of strategy is chosen, the new decision is sent to the platform; after each round, the platform collects the worker nodes' decisions and recalculates their rewards; when no worker node in the system wants to change its strategy, the task selection phase ends. 3) Task completion phase: Worker nodes in the federated learning system execute federated learning tasks according to the strategy of the task selection phase and send the results to the platform; the platform distributes rewards to worker nodes according to the results calculated in the task selection phase and sends the training results of worker nodes to the task issuer. The task publisher aggregates the training results of worker nodes and feeds the results back to the platform, which then updates the reputation scores of the worker nodes based on the feedback.

2. The worker node task allocation method for node centralization according to claim 1, characterized in that: In the task allocation phase, the sigmoid activation function is first used to convert reputation scores into reputation weights to filter out worker nodes with poor quality and amplify the differences in reputation scores between different worker nodes. This facilitates the allocation of rewards according to the quality of worker nodes after the federated learning task, improving the system's operational efficiency and fairness. The formula for the sigmoid activation function is: The core idea of ​​the task selection algorithm is to quickly reach Nash equilibrium through multiple rounds of selection, allowing one worker node to change its strategy in each round. At the beginning of each round, the worker nodes in the system inform the platform of their task selection strategies. After the platform aggregates the strategies of all worker nodes, it calculates the sum of the reputation weights for each federated learning task and returns the results to the worker nodes. After receiving the results, the worker nodes calculate the potential rewards based on a weighted distribution method and inform the platform of their strategy for the next round. If any worker node in the system changes its strategy in the next round, the task selection process continues; otherwise, the worker nodes in the system reach a consensus on the task selection strategy, and the task selection algorithm ends. In each round of task selection, the strategy is updated through two steps: (2.1) The task selection strategy of worker node i in the rth round is expressed as: After receiving the task selection strategy from each worker node, the platform statistically analyzes the participation of worker nodes in each task within the system, calculates the sum of reputation weights of worker nodes participating in each federated learning task, and presents the results as a vector group. The form is returned to the worker node; (2.2) The worker node uses the vector group G returned by the platform. r Calculate the benefits that can be obtained by participating in each federated learning task in the next round, namely: Based on the above formula, worker node i can find the federated learning task that will maximize its benefits in the next round. like This indicates that the optimal strategy has been found in the task selection of the r-th round, and it is sent to the platform. If worker node i indicates that it will not change its strategy, then it will update its task selection strategy in the next round to obtain higher returns. After the platform receives update requests from all worker nodes, it randomly selects one worker node to update its strategy and sends it a confirmation response. Next, the platform re-aggregates the new round of worker node strategies and calculates the result G. r+1 Return to the worker nodes within the system; repeat the above process until... That is, all worker nodes have chosen the federated learning task that maximizes their own benefits and do not change their strategies in the next round; this means that the non-cooperative game of task selection has reached Nash equilibrium, and after the worker nodes in the system confirm their choices, they begin to execute the federated learning task according to their own strategies. Using θ min and θ max Let R represent the minimum and maximum values ​​of the reputation weight of the worker node, respectively. min and R max Let c represent the minimum and maximum rewards for the federated learning task within the system, respectively. min Let represent the minimum cost of a worker node. Then, the maximum number of rounds that this task selection algorithm will perform is:

Images