Scholarlink

Federated learning (FL) has emerged as a decentralized machine learning paradigm that enables collaborative model training across distributed devices while preserving data privacy. However, FL is highly susceptible to poisoning attacks, where adversaries manipulate data or model updates to degrade the overall performance. This paper comprehensively analyzes poisoning attacks in FL, categorizing them into data poisoning and model poisoning. We implement various poisoning attack strategies and evaluate their impact on FL systems across different scenarios. To counteract these threats, FL relies on robust aggregation techniques. We examine the resilience of standard aggregation methods, such as Federated Averaging (FedAvg) and Federated Proximal (FedProx), against poisoning attacks and compare them with robust approaches, including Krum, trimmed mean, and median-based aggregation. The results for model and data poisoning attacks highlight the effectiveness of the Krum aggregation technique, which maintains robust performance even with up to 50% malicious nodes.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Assessing the Robustness of Federated Learning Aggregation Techniques Against Diverse Poisoning Attack Strengths

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aya Nabil Sayed,
  • Md. Mosarrof Hossen,
  • Faycal Bensaali,
  • Armstrong Nhlabatsi

摘要

Federated learning (FL) has emerged as a decentralized machine learning paradigm that enables collaborative model training across distributed devices while preserving data privacy. However, FL is highly susceptible to poisoning attacks, where adversaries manipulate data or model updates to degrade the overall performance. This paper comprehensively analyzes poisoning attacks in FL, categorizing them into data poisoning and model poisoning. We implement various poisoning attack strategies and evaluate their impact on FL systems across different scenarios. To counteract these threats, FL relies on robust aggregation techniques. We examine the resilience of standard aggregation methods, such as Federated Averaging (FedAvg) and Federated Proximal (FedProx), against poisoning attacks and compare them with robust approaches, including Krum, trimmed mean, and median-based aggregation. The results for model and data poisoning attacks highlight the effectiveness of the Krum aggregation technique, which maintains robust performance even with up to 50% malicious nodes.