acceptodds
Under review as a conference paper at ICLR 2027

HeatGuard: Inspecting Backdoor Poisoning in Driver Attention Heatmap Prediction

Abstract

Driver attention prediction is widely used in driving risk assessment and human-centered driver assistance systems. Existing methods rely heavily on deep neural networks and large-scale training data, which may be obtained from unverified or external sources, exposing the models to backdoor poisoning. By manipulating a small subset of image and heatmap training pairs, an adversary can implant a hidden trigger. The compromised model behaves normally on benign inputs but produces an attacker-specified attention heatmap when the trigger is present, potentially misleading driver attention analysis and downstream driving assistance systems. To address this threat, we propose HeatGuard, a training-data inspection framework. HeatGuard exploits the differences in feature evolution between benign and poisoned samples and constructs attention-conditioned layer-wise feature trajectories to distinguish them. Experimental results demonstrate that HeatGuard effectively detects poisoned samples across multiple backdoor attacks and poisoning-rate configurations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
backdoor attacks, driver attention prediction, heatmap prediction, training data inspection, poisoned sample detection, adversarial machine learning, autonomous driving

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FQvjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies a dataset poisoning threat due to a visual backdoor inserted in some of the images of the dataset. The goal of the attack is to cause the model to misdirect the gaze prediction of the driver when the attention predictor is triggered by a specific visual pattern (the backdoor). The present paper evaluates the attack protocol accross 4 models (MLNet, PGNet, SalM2, SCOUT) and 4 attacks variants (Bad-CAP, Bad-Mis, Blend, SIG). The paper also propose a training-data inspection framework named HeathGuard. The evaluation of HeathGuard is positive (perfect AUROC).

Critical strengths
  • The topic focus on an interesting and important threat in the area of Driver Monitoring Systems (DMS) which is timely and relevant.
Critical weaknesses
  • It is hard to interpret how generalizable the results are. For instance, the paper uses a single dataset. Also, a lot of the tabular results (attack and defense evaluation) lack of statistical meaning (standard deviation, confidence interval, etc.)

  • The attack presented in this paper are efficient but not stealthy Looking at Figure (t-SNE), the adversarial samples can be easily distinguished from genuine sample. Also, looking at table 4, it seems there is no benefit for picking HeatGuard over HCC (which is a state-of-the-art method) because they both achieve similar performance. I would suggest the author to add attacks that are more challenging to detect for the baseline.

  • The positioning of the paper in the literature is unclear. There are extensive litterature in dataset poisoning and mitigation and the authors did a good in presenting the prior art. However, when it comes to the specific use case addressed in this paper (Human Gaze), the paper is vague on the positioning of the paper regarding prior art. For instance, the following reference is missing "L. Du et al. SecureGaze: Defending Gaze Estimation Against Backdoor Attacks. arXiv:2502.20306, 2025." My point would be to provide a clear text or table positioning the current paper against prior art.

Questions for authors
Reviewer_z571OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper focus on a data poisoning attck (backdoor attack) for a new scene "driver attention misleading". The authors formulate and implement two primary attack modes (termed, attention capture & attention misdirection) and propose a training-data inspection framework operating before downstream model training. Specifically, training a lightweight inspection proxy with untrusted dataset and a small trusted clean reference set, to track how intermediate features evolve across different layers when weighted by the predicted attention map. The key results show the effectiveness on poisoned samples identifying.

Critical strengths

A new defense framework for the proposed driver attention prediction heatmap hijakcing scene.

Critical weaknesses
  • The methods is naive with limited innovation
  • Dataset and Experiment is single with questionable generalizability
  • Adaptive attacks may easily escape from the defense & Some stealthiness checking methods as baseline should also be included.
Questions for authors
  1. How does HG fundamentally differe from DBALFA?
  2. How does HG perform under domain shift when D_r is drawn from a different dataset?
  3. Why is the evaluation resticted to static clear-weather frames on TrafficGaze? Can the authors evaluate on more complex benchmarks?
  4. If an adaptive attacker adds a latent feature-alignment regularization term during poisoning crafting to suppress layer-wise deviation, and interlayer breaks, does HG fail?
  5. Do you need to retrain the inspection proxy from scratch everytime the dataset is updated?
Reviewer_uhEwOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies training-data backdoors in driver-attention heatmap prediction. The attacker inserts square, Blend, or SIG triggers into a small fraction of TrafficGaze image–heatmap pairs and replaces the labels with either a trigger-region target or a spatially separate target. HeatGuard trains ProxyAtt on the untrusted pool, uses the proxy-predicted heatmap for attention-conditioned feature pooling at several layers, computes robust deviations of layer-wise cosine trajectories relative to a trusted clean reference set, and combines top-layer deviations with adjacent-layer breaks. A rank-gap rule forms a quarantine set. The paper evaluates MLNet, PGNet, SalM2, and SCOUT under several attack types and rates, comparing HeatGuard with ABL, AC, SPECTRE, AGPD-GCD, DBALFA, and the label-only HCC baseline.

Critical strengths
  1. A clear, underexplored security task. The paper formalizes continuous heatmap hijacking and distinguishes capture near the trigger from spatially separated misdirection.
  2. Reasonably broad defense and baseline coverage. The paper compares feature, gradient, low-loss, and label-only approaches, varies poison rate, trigger style, reference budget, and proxy architecture, and makes clear that rank-gap selection does not use the true poison count.
Critical weaknesses
  1. External validity and statistical reliability are insufficient. The main results use one TrafficGaze dataset, one split, and fixed seeds (proxy-training seed 42 and poison/inspection seed 2026). There is no strict source-video-disjoint evaluation or multi-run confidence interval. Every attack/rate entry in Table 3 has AUROC 1.000, which requires a careful check for near-duplicate frames or overlap between the reference and inspected pools. The authors should report source-video-disjoint splits, repeated runs, and full suspiciousness-score distributions.
  2. The gain over label and output cues has not been established. ProxyAtt is trained on the poisoned pool and its predicted heatmap is used for pooling; when the target heatmap is conspicuous, the method may mainly detect poisoned labels or proxy outputs rather than feature evolution. In Table 4, HCC reaches the same AUROC as HeatGuard in several Bad-Cap settings. The paper also lacks systematic tests of clean-label, dynamic/semantic-target, physical, or proxy/trajectory-adaptive attacks, as well as downstream utility after retraining on the cleaned data.
Questions for authors
  1. Please report exact sample counts for the untrusted pool, trusted reference set, test set, and every poison rate. Are the reference and inspected samples completely separated by source video, scene, and adjacent frames? Please provide source-video-disjoint, multi-seed AUROC/F1 and confidence intervals.
  2. Please compare HCC, simple heatmap statistics, pixel/trigger detectors, and a proxy-output-only baseline under the same attacks. Please also ablate attention-conditioned pooling and separately report label-only and feature-trajectory-only contributions.
  3. Under fixed and variable quarantine budgets, please report precision, recall, F1, clean-removal rate, and the clean utility and attack success after retraining without the quarantine set. How stable is the rank-gap boundary to ties and reference-set changes?
  4. Please add at least one clean-label, dynamic/semantic-target, or adaptive attack against ProxyAtt/trajectory scoring, and evaluate physical triggers and cross-weather/cross-dataset shifts.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.