FedLSG: LLM-Enhanced Semantic Calibration for Federated Graph Backdoor Defense

FedLSG defends Federated Graph Neural Networks against backdoor attacks by converting local graph structures and client updates into natural language via a graph-to-text grounding scheme. A server-side teacher LLM evaluates global context during aggregation, while clients utilize a LoRA-based student model to perform semantic reasoning and mitigate poisoning.