IDEA Reports 0.972 Average TPR for Trigger-Free GUI Agent Backdoor Detection at 1% FPR

Found first: a primary source the press has not covered yet.

Researchers at HKUST, Peking University, and Huawei Noah's Ark Lab have published IDEA (Intent-Dominance Evidence Audit), a runtime method for detecting backdoored GUI agents that requires neither trigger specification nor attack training data. Across five attack families on the AndroidControl benchmark, IDEA achieves 0.972 average true-positive rate at a controlled 1% false-positive rate, outperforming the next-best detector (PCA residual at 0.893).

What the source says

The method identifies "intent-evidence dominance inversion": in a backdoored step, screen regions the user's instruction does not reference exert more influence on the agent's decision than the instruction itself. Intent Counterfactual Probing (ICP) measures how much instruction perturbation shifts the agent's decision state; Region Causal Audit (RCA) measures how much ablating candidate screen regions does the same. A gated conformal scorer compares these shifts against a clean-step calibration set and produces a finite-sample false-positive guarantee. The primary victim model is Qwen2.5-VL-3B on AndroidControl; a held-out test on OS-Atlas-Base-7B returns 0.883 TPR at 0.008 clean FPR. Per-step overhead is 9-10%, versus 12.9-15.4 seconds per step for the nearest comparable method (TeCo), and IDEA also localizes trigger regions for flagged steps.

Why it matters

Enterprises increasingly deploy open-weight GUI agents downloaded from public model hubs; a backdoor passes clean-data acceptance testing and activates only when a crafted visual trigger appears in a live screen. Prior inference-time detectors require knowledge of the trigger or attack examples, and none provides a formal bound on false positives. Input perturbation methods returned below 0.10 TPR across all five attack families in this evaluation. IDEA needs only a calibration set of clean agent steps and provides a formal guarantee on the false-alarm rate. The localization output is also actionable: when a step is flagged, operators can see which screen region drove the decision.