Skip to content

Illusion: Protecting the Topology Confidentiality of LLM Multi-Agent Systems via Behavior-Preserving Obfuscation

Illusion: Safeguarding Topology Confidentiality in LLM Multi-Agent Systems via Behavior-Preserving Obfuscation

arXiv· arXiv:2611.00923· Yiu-Fai Cheung, Hongzhi Guo, Miao Pan, Irwin King, Hao WangThe Chinese University of Hong Kong; University of Nebraska-Lincoln; University of Houston; Stevens Institute of TechnologyIndexed by this site · 16 days ago· Originally publishedDefenses

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

AI summaryIllusion is an active topology obfuscation defense for LLM multi-agent systems: it injects fake edges and fake messages to present a deceptive topology to attackers while maintaining output distribution equivalence and no degradation in task accuracy. Experiments show that the edge-level F1 of topology reconstruction attacks drops from above 0.87 to 0.62, and graph similarity drops from above 0.94 to 0.71. Fake edges are difficult to distinguish from real messages under black-box observation, and can resist similarity and LLM-judge filtering as well as adaptive pruning.

Questions
The communication topology of LLM multi-agent systems is important intellectual property, and can be stolen from agent outputs by topology inference attacks; after leakage it also facilitates attacks targeting critical agents; the authors state that existing defenses only perform detection and mitigation, and do not actively hide the topology itself.
Methods
Proposes Illusion: inject fake edges into the MAS and send fake messages produced by generative prompts so that attackers see a deceptive topology; through three stages—candidate generation, probe-based candidate ranking, and iterative injection with verification—it guarantees output distribution equivalence and non-inferior task accuracy.
Experiments and results
Evaluated on MMLU, GSM8K, and GPQA with Llama-3.1-8B and Qwen-3.5-9B as backbones: the authors report that CIA attack F1 drops from above 0.87 to 0.62, and GSim from above 0.94 to 0.71; output distribution equivalence passes in all 18 settings; similarity and LLM-judge filters and adaptive pruning all fail to effectively remove fake messages.
Limitations and future work
In the Limitations section, the authors note: the evaluation covers only three datasets, two open-source backbone models and three representative topologies, and effectiveness in scenarios such as new architectures and dynamic topologies remains to be verified; future directions include dynamic LLM-MAS, resource-aware injection and stronger adaptive attackers.

Related papers

Read originalarxiv.org