Abstract: The federated learning paradigm is designed to preserve privacy by enabling multiple clients to collaboratively train a global model without compromising the original training data.However,due to the lack of direct access to local training data and monitoring capabilities during the training process,federated learning is vulnerable to various Byzantine attacks,including data poisoning and model tampering attacks.These malicious activities aim at disrupting the federated learning model training process and degrading its performance.While several studies have proposed various aggregation algorithms to address this issue,they predominantly concentrate on single Byzantine attack scenarios,often overlooking the threats associated with hybrid Byzantine attacks that can manifest in real-world environments.To address this issue,inspired by the principle of water purifiers,we propose an innovative multi-feature detection and adaptive dynamic weighting allocation algorithm called FL-Sieve for identifying Byzantine clients,aiming to filter out malicious clients through multi-level screening.Firstly,the algorithm assesses feature similarity between clients through angular range similarity and model boundary metric,generates a similarity matrix and calculates the similarity score.Then,it performs clustering to ensure that nodes with similar features are grouped together.Subsequently,it employs predefined rules to filter potential benign clients.Finally,it intelligently allocates weights based on the trustworthiness of each client,further enhancing the defense mechanisms and system robustness.To evaluate the performance of the FL-Sieve algorithm,experiments are conducted using three datasets:MNIST,Fashion-MNIST,and CIFAR-10.The experiments consider scenarios with both non-IID data distribution and hybrid Byzantine attack situations.The number of hybrid Byzantine clients increases from 20% to 49% to simulate large-scale hybrid Byzantine client attacks.Additionally,the performance of the FL-Sieve algorithm is tested in both IID and non-IID data distribution,as well as in single attack scenarios.The experimental results demonstrate that FL-Sieve effectively withstands Byzantine attacks in various scenarios,maintaining high main task accuracy even under the challenging condition of 49% hybrid Byzantine client attacks.In comparison,several existing classical algorithms exhibit varying degrees of failure,underscoring the significant advantages of the FL-Sieve algorithm.

Key words: Federated learning, Hybrid Byzantine attack, Multi-feature detection, Dynamic weight allocation, Robust aggregation algorithm