Safe Multi-Agent Reinforcement Learning (SMARL) aims to learn cooperative control policies while respecting predefined safety constraints. Existing SMARL approaches mainly enforce safety during policy optimization, leaving inference‑time vulnerabilities largely unexplored. This work investigates observation attacks, where an adversary perturbs the observations of a subset of agents, thereby altering the actions generated by their policies.
We introduce the MinMax Attack framework. It jointly selects adversarial target actions for compromised agents and accounts for the Cooperative Resistance of benign agents, effectively steering the system toward a worst‑case behavior.
Our theoretical analysis derives an upper bound on the induced safety cost $C_{s}$:$$C_{s} \le f(B_{c}, B_{p})$@@@MATH_BLOCK2@@@B{c}@@@MATH_BLOCK3@@@B{p}$ the perturbation budget; the function $f$ captures their interaction. This bound highlights how budget allocation governs safety risk.
Experiments on several safety‑critical multi‑agent benchmarks and state‑of‑the‑art SMARL algorithms show that MinMax Attack yields substantially higher safety costs than existing baselines while preserving task reward, making the attack both effective and stealthy.
Review