Client Selection for Federated Learning Against Poisoning Attacks on Non-IID Data

Federated Learning (FL) trains a global model using local models uploaded by clients rather than raw data, leading to the protection on client privacy. However, FL is vulnerable to malicious attacks, such as poisoning attacks, which can degrade the performance of the global model. Especially in non-Independent and Identically Distributed (non-IID) scenarios, traditional malicious client detection methods are largely ineffective. To address these challenges, we propose an algorithm termed Federated Learning with Client Selection Based on Contribution and Credibility (FL-CSBC<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>). Specifically, we design a client contribution model based on marginal contribution and a client credibility model utilizing the Rank-Sum Ratio (RSR) in the FL training process. By adopting the concept of evolutionary clustering, we utilize the current round's contribution and the previous round's credibility to quantify the credibility of each client by computing a client score, based on which we select clients that will participate in aggregation phase. At the same time, we incorporate the root dataset on the server side and utilize the server model trained by the root dataset as the initial aggregation model for each round. We evaluate the performance of our algorithm on two public datasets with different proportions of malicious clients under non-IID scenario. Extensive experimental results demonstrate the superiority of FL-CSBC<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> over a variety of baseline algorithms in terms of test accuracy.