DistilGuard - Large Language Models for Poisoning Detection in Federated Learning

Federated Learning (FL) systems face significant risks from poisoning attacks. Existing detection methods, relying on numerical or statistical analysis, often struggle against sophisticated threats. This paper introduces DistilGuard, a novel Large Language Model (LLM)-based defense mechanism for poisoning detection in FL - representing the first known application of LLMs for this specific purpose in this domain. Unlike conventional approaches, our novel defense architecture considers poisoning detection as a natural language understanding problem. This is achieved by transforming statistical and behavioral signatures from client model updates into structured natural language representations for the LLM to process. This allows the fine-tuned LLM to analyse complex patterns and nuances beyond simple statistical comparisons. The server-side poisoning filtering mechanism utilises the LLM's classifications to detect and exclude malicious updates before aggregation. We demonstrate high detection accuracy across various datasets and both data and model poisoning attack types, assessing a new direction for intelligent FL security with LLMs.