Scholarlink

<p>Federated learning (FL) enables the decentralized training of machine learning models across multiple clients while preserving data privacy. However, the exchange of local models among untrusted participants introduces significant security vulnerabilities, particularly model poisoning and Byzantine attacks, which can severely degrade the performance of the global model. To address these challenges, we propose a lightweight and robust Explainable Federated Learning (XFL) framework that leverages Explainable AI (XAI) techniques to detect and filter out malicious model updates. Specifically, we employ Class Activation Maps (CAM) and introduce two novel detection methods: the Cosine Similarity Method (CSM) and the Similarity Difference and Uniqueness (SDU) approach, along with an enhanced variant, Improved-SDU (I-SDU), which analyzes multiple convolutional layers for greater robustness. Experimental evaluations on the MNIST and ImageNet datasets demonstrate that the proposed framework achieves an average malicious model detection accuracy of 95%, even in scenarios involving up to 100% adversarial clients. Furthermore, the XFL framework effectively restores the accuracy of the global model from 20% under a full-scale attack to 80% using CSM and up to 88% with I-SDU. These results underscore the effectiveness of our explainability-driven approach in improving the security and reliability of FL systems, particularly in privacy-sensitive domains such as healthcare and finance.</p>

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

Unique explainable federated learning approach for detecting model poisoning attacks to secure systems

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Muhmmad Fizan,
  • Alaa Awad Abdellatif,
  • Muhammad Asad,
  • Muhammad Zubair,
  • Bin Ni,
  • Bin Xu,
  • Xuefeng Liu,
  • Jichuan Xiong,
  • Muhammad Qasim Mehmood

摘要

Federated learning (FL) enables the decentralized training of machine learning models across multiple clients while preserving data privacy. However, the exchange of local models among untrusted participants introduces significant security vulnerabilities, particularly model poisoning and Byzantine attacks, which can severely degrade the performance of the global model. To address these challenges, we propose a lightweight and robust Explainable Federated Learning (XFL) framework that leverages Explainable AI (XAI) techniques to detect and filter out malicious model updates. Specifically, we employ Class Activation Maps (CAM) and introduce two novel detection methods: the Cosine Similarity Method (CSM) and the Similarity Difference and Uniqueness (SDU) approach, along with an enhanced variant, Improved-SDU (I-SDU), which analyzes multiple convolutional layers for greater robustness. Experimental evaluations on the MNIST and ImageNet datasets demonstrate that the proposed framework achieves an average malicious model detection accuracy of 95%, even in scenarios involving up to 100% adversarial clients. Furthermore, the XFL framework effectively restores the accuracy of the global model from 20% under a full-scale attack to 80% using CSM and up to 88% with I-SDU. These results underscore the effectiveness of our explainability-driven approach in improving the security and reliability of FL systems, particularly in privacy-sensitive domains such as healthcare and finance.