Reproduce “Illusion: Safeguarding Topology Confidentiality in LLM Multi-Agent Systems via Behavior-Preserving Obfuscation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiu-Fai Cheung, Hongzhi Guo, Miao Pan, Irwin King, Hao Wang
Illusion protects multi-agent communication topologies by adding low-impact decoy links and task-specific plausible messages while retaining the genuine workflow. It selects decoys through receiver-impact probes and checks output distributions and task accuracy. On MMLU, CIA’s edge-level F1 falls from 0.87 to 0.62 and graph similarity from 0.94 to 0.71; performance preservation has a GPQA star-topology exception.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.