Illusion Obfuscates LLM Multi-Agent System Topology
Original title: Illusion: Safeguarding Topology Confidentiality in LLM Multi-Agent Systems via Behavior-Preserving Obfuscation
Illusion injects fake edges and messages to conceal topology while preserving output distributions and task accuracy.
The communication topology in Large Language Model Multi-Agent Systems encodes critical intellectual property and is vulnerable to inference attacks. Illusion proactively obfuscates this topology by injecting fake edges and messages that are indistinguishable from genuine ones. Experiments show substantial reductions in edge-level F1 and graph similarity while maintaining performance. The approach resists both similarity-based and LLM-judge filtering.
- Reduces edge-level F1 from over 0.87 to 0.62
- Reduces graph similarity from over 0.94 to 0.71
- Preserves output distributions and task performance
- Fake edges resist filtering under adaptive pruning
From the original article
1 min read at arxiv.orgThe communication topology of Large Language Model Multi-Agent Systems (LLM-MASs) governs how information propagates among agents and encodes critical domain knowledge and expert workflow design. It is valuable intellectual property and plays a decisive role in system utility and safety, yet remains vulnerable to inference attacks that stealthily reconstruct it from agent outputs.
By Yiu-Fai Cheung · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.