Safety-Contract Graph Multi-Agent Reinforcement Learning for Autonomous Network Security Response

The article introduces a safety-contract graph multi-agent reinforcement learning (MARL) framework called ACD$^3$-GAT, which integrates a Graph Attention Network encoder for enhanced operational discipline in network security response. Evaluated in the CAGE Challenge 4, ACD$^3$-GAT significantly improves performance by reducing mean downtime costs to 48.2 with a 13.8% violation rate, compared to 355.4 and 100% for the standard MAPPO-GAT. This framework emphasizes the importance of constrained optimization and budget-aware decision-making, making it a critical advancement for practitioners aiming to deploy effective autonomous security systems.