LCEFL: A Lightweight Contribution Evaluation Approach for Federated Learning
Jingjing Guo, Jiaxing Li, Zhiquan Liu, Yupeng Xiong, Yong Ma, Athanasios V. Vasilakos, Xinghua Li, Jianfeng Ma
Abstract
The prerequisite for implementing incentive mechanisms and reliable participant selection schemes in federated learning is to obtain the contribution of each participant. Available evaluation methods for participant contributions require the server to possess a test dataset, often impractical. Additionally, the excessively high complexity of these works is unacceptable when training complex models in large-scale federated learning system. To address these issues, we propose a lightweight contribution evaluation method for federated learning participants, named LCEFL, based on model projection theory, which does not require the server to provide a test dataset. In addition, a model compression method is designed to be used in LCEFL to reduce the computational complexity. Furthermore, a trusted aggregation method based on LCEFL is proposed, where the weight of each participant's local model is determined by its trust level, which can be calculated using its contribution evaluation result. Experimental results show that LCEFL can achieve nearly the same accuracy as schemes based on Shapley Value, while significantly reducing computational overhead by more than 50%. Compared to available aggregation methods, the proposed trusted aggregation scheme is able to accelerate the convergence speed of the global model and improve its accuracy by 2% to 45%.