Downloaded GUI agents can ship with a backdoor: IDEA audits each step before it runs and averages 0.972 detection across five trigger families
Synopsis
Open-weight GUI agents circulating on public model hubs can be backdoored before release, passing acceptance tests on clean inputs and switching to an attacker-chosen action only when a visual trigger appears on screen. IDEA treats the clean-step relation, in which the instruction selects the action and the evidence lies where the instruction points, as its baseline, measures decision-state displacement under two interventions, instruction redirection and candidate-region ablation, and calibrates a threshold against clean steps at the same state distance. On Qwen2.5-VL-3B across five trigger families at a 1% target false-positive rate, IDEA averages 0.972 detection with a worst family of 0.898 and returns a trigger region with every alarm.
Interpretation
IDEA turns backdoor detection for GUI agents from a whole-step anomaly judgment into a per-step audit of whether the instruction still controls the action, and names the screen region that took over the decision when it alarms. Earlier inference-time detectors target image classifiers and CLIP-style encoders and assume a classification output and global embedding shifts; according to the authors, none addresses GUI agents that map a screenshot and an instruction to a coordinate, and none provides per-step detection with a calibrated false-positive rate that also returns the controlling region. On a strict app- and episode-disjoint AndroidControl test split at a 1% target false-positive rate, IDEA averages 0.972 detection over the five trigger families against 0.894 for the PCA residual, and its worst family reaches 0.898 against 0.580; input-perturbation and output-based detectors stay below 0.10 on every family.
The audit does not read the agent's own attention; it intervenes on the input twice, redirecting the instruction to see how far the decision state moves and ablating a candidate region to see how far it moves again. Redirects are built from other visible elements in the accessibility tree and OCR as click instructions, and only those the agent actually follows are kept; candidate regions come from accessibility-tree boxes, OCR boxes, connected components of gradient attribution and a fixed grid, and ablation replaces their visual tokens with tokens from a 3-pixel Gaussian-blurred screenshot. Both displacements grow with the decision state's distance from the clean distribution, so each is compared with a fitted clean law at the same distance to give an excess-evidence and an intent-deficit term, and a Mahalanobis gate screens out most clean steps, passing 0.047–0.051 of clean test steps and 98–100% of triggered ones.
An alarm hands the step to a human operator rather than repairing the action, because a compromised checkpoint could repeat the backdoor on re-query. The gate and the statistic combine into one score whose threshold is a split-conformal quantile on clean calibration steps, giving a finite-sample false-positive guarantee over clean steps. Timed online on the Patch and Grounding victims, the gate adds 0.034 s to a 0.65 s agent step, and the audit runs only on the 2.7–3.0% of clean steps that pass the gate at 0.79 s, so a clean step takes 0.67–0.68 s, about 9–10% overhead.
Perspective
The demonstrated setting is Qwen2.5-VL-3B victims backdoored on AndroidControl with five trigger families at 10% poisoned training steps, with detection evaluated on strict app- and episode-disjoint test steps at 1% and 0.5% target false-positive rates; the method needs the checkpoint weights and a small clean auxiliary set and uses no trigger location, target action or attack data.
The statistic's form was finalized after observing the five development trigger families, so the main table is development evidence and the blind test covers only two held-out victims; the split-conformal guarantee is marginal over clean steps, and after resampling apps the intervals for Popup, Grounding and Adaptive-Blend exceed 1%, with the worst single seed reaching 0.023; on the full-screen Adaptive-Blend trigger IDEA is on par with the Mahalanobis screen and falls below it at the 0.5% target; and the five-family evaluation centers on a single 3B backbone, with larger models and other model families not yet tested on all triggers.
