Megadose The AI curve, compressed.

Securing Multi-Agent Systems by Design

· ArXiv · AI/CL/LG ·
SecMAS resisted every attack evaluated in the paper while preserving utility comparable to existing aggregation rules, the authors report.

Their response manipulation attack, ReMA, refines malicious agents’ answers using feedback from a surrogate multi-agent system. Existing aggregation rules proved vulnerable in the evaluations. SecMAS instead selects an answer whose meaning is consistent with a sufficiently large subset of other agents’ responses. The authors establish theoretical security guarantees against targeted and untargeted attacks and test adaptive ReMA variants.

ArXiv · AI/CL/LG's note

score 5

Categories: Research