MADGDF: Multi-Agent Distributed Graph-State Deception Framework

Résumé fourni par la source

Reference implementation and experimental pipeline for MADGDF, a defensive architecture for autonomous multi-agent systems combining graph neural network state estimation with a deterministic memory-canary mechanism to detect coordinated adversarial behavior and unauthorized persistent-memory retrieval. Evaluated on a controlled synthetic corpus (SwarmBench-2026, 32- and 250-agent configurations) and on real agent-execution trajectories from ToolBench, tau-bench, and SWE-agent.