Adaptive trimmed mean-based aggregation for semi-centralized federated learning

Security issues in Federated Learning (FL) have consistently attracted research attention, particularly concerning system stability under Byzantine client attacks. To address this critical problem, this paper proposes ATmean, an aggregation method based on adaptive trimmed mean. Additionally, a verifiable authentication mechanism is integrated on top of the traditional Byzantine consensus protocol to ensure secure updates of client gradient information. The experiments are conducted on a semi-centralized federated secure learning chain, and a detailed analysis of this learning framework is presented. Experimental results demonstrate that the adaptive trimmed mean (ATmean) aggregation rule exhibits superior anti-Byzantine capability compared to other statistical aggregation methods.