When Collaboration Becomes a Trigger: Collective Evidence-Threshold Backdoors in Multi-Agent Systems

DOI10.48550/arXiv.2608.01085MaRDI QIDQ7284741arXivFDO

Authors Lei Feng, Jia-Hao Xiao, Min-Ling Zhang

Publication date 2 August 2026

Published in arXiv

Tldr We introduce collective evidence-threshold backdoors in LLM-based MAS and Boundary-Conditioned Backdoor Injection (BCBI) to construct counterfactual boundary pairs for benign vs. adversarial behavior before/after a hidden threshold; we also propose clean-only latent-transition defense LATTE that learns benign communication dynamics and quarantines anomalous agent updates.

Full work available at URL https://doi.org/10.48550/arXiv.2608.01085

arXiv classification

cs.LGcs.MA

This page was built for publication: When Collaboration Becomes a Trigger: Collective Evidence-Threshold Backdoors in Multi-Agent Systems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284741)