Skip to content
InfoResearchPreprint

IDEA: Intent-Dominance Evidence Audit for Inference-Time Backdoor Detection in GUI Agents

Published
Record updated
View JSON

Summary

The paper proposes IDEA (Intent-Dominance Evidence Audit), a method for detecting backdoor activation in open-weight GUI agents at inference time, per step, without trigger knowledge or attack data. It combines Intent Counterfactual Probing (ICP) and Region Causal Audit (RCA) in a gated conformal score with a finite-sample false-positive guarantee. Experiments on Qwen2.5-VL-3B across five attack families show it detects backdoors under a controlled false-positive rate, outperforms prior detectors, and localizes the trigger region for each alarm.