When to Intervene? State-Aware Sparse Manipulation in Federated Reinforcement Learning
The paper formalizes Byzantine Federated Reinforcement Learning by separating *when* to intervene from *how* to construct a malicious update. It introduces the Viability-Constrained Behavioral Steering Attack (V-BSA), which selects specific trajectory states for manipulation based on policy entropy. Findings show that changing these selected states materially degrades downstream robustness, revealing task and aggregation-dependent vulnerabilities.