MIRAGE: showing the attacker a fake organisational chart of staff to protect the real one

In a multi-agent system, it is possible to work out who is communicating with whom by analysing their reasoning traces. MIRAGE generates a credible phantom topology to thwart such attacks without compromising the actual work.

In a multi-agent system based on LLMs, the way in which agents are connected — who passes what on to whom — is often a trade secret. However, recent research has shown that an attacker can reconstruct this black-box topology simply by identifying the directional dependencies between the visible reasoning traces. This gives them both a competitive advantage and a map of the system’s weaknesses.

The authors propose MIRAGE, which allows the true topology to function as normal whilst steering the visible clues towards a decoy structure. The method proceeds in three stages: synthesising a decoy topology that is distinctly different from the real one; creating plausible semantic dependencies to materialise its links; and then running the system in a protected manner by removing any clues that would betray the real links absent from the decoy.

Tested on three topology optimisation frameworks and four datasets, MIRAGE significantly reduces the effectiveness of inference attacks whilst retaining most of the system’s performance on the tasks.