Reputation-Based Client Selection for Federated Learning Under Unreliable Communication
Federated learning is a distributed machine learning framework that has received widespread attention for its data protection properties.However,malicious clients and unreliable communication seri-ously affect its performance and efficiency.To solve the above problems,a reputation-based federated learning client selection mechanism for multi-task publishers in unreliable communication is proposed.Firstly,the communication reliability is evaluated using the uplink model transmission success probability and its impact on the performance of the aggregation model is also considered.Secondly,a comprehensive client reputation evaluation method is proposed and a client selection mechanism with the optimization objective of maximizing the performance of the aggregation model of the task publisher as well as the reputation-price ratio of the selected client is constructed.To solve this optimization problem,it is mod-eled as a Markov decision process and the curiosity driven deep Q-learning network algorithm is used to achieve optimization.The result shows that the proposed algorithm outperforms the baselines,leading to a significant improvement in the performance of federated learning.
federated learningunreliable communicationreputationclient selectioncuriosity driven deep Q-learning network(CDQN)
国家科技期刊平台
NSTL
万方数据
