Towards robust federated learning

Abstract

The proliferation of IoT and edge computing has accelerated the adoption of Federated Learning (FL) as a privacy-preserving framework for collaborative model training. However, FL systems remain vulnerable to Byzantine attacks, where malicious participants disrupt learning through data poisoning or corrupted model updates.

This project explores the integration of Large Language Models (LLMs) into FL to enhance Byzantine fault tolerance. Leveraging LLMs' contextual understanding and analytical capabilities, we aim to improve model robustness against adversarial behaviors while preserving performance.

Our work includes a comprehensive analysis of Byzantine attack strategies and the design and evaluation of LLM-augmented defense mechanisms. Through rigorous experimentation on deep learning tasks in resource-constrained environments, we investigate the trade-offs between robustness, scalability, and model accuracy.

The findings contribute to the development of a secure and resilient FL framework, opening pathways for the reliable deployment of LLMs in distributed, privacy-sensitive applications.