FedWM-Guard: defending against imagination poisoning in autonomous driving based on federated world models

The study proposes FedWM-Guard to defend against imagination poisoning in autonomous driving based on federated world models: a compromised vehicle submits bounded WM updates that maintain benign short-term predictions but corrupt long-term rollouts, thereby misleading the downstream planner. The defense screens certified updates in hidden canary scenarios, audits by comparing predicted futures with subsequent observations, and enables a safety guardrail independent of the WM upon persistent inconsistency; the study plans to evaluate under non-IID data, adaptive attacks and benign distribution shifts.