acceptodds
Withdrawn from ICLR 2027

DASH: Learning Adversarial Distribution Shifts for Transferable Black-Box Attacks on Object Detectors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Object detectors are widely deployed in safety-critical applications. Since deployed models conceal internal information, transferable black-box attacks provide a practical robustness evaluation. Unlike image-classification attacks, detection attacks must suppress multiple objects while transferring across heterogeneous architectures. Our gradient analysis identifies two obstacles: detector gradients are poorly aligned, encouraging surrogate-specific solutions, while object gradients can conflict and interfere with joint suppression. Distribution-level generation provides a natural mechanism for this goal: by aggregating attack supervision over the training distribution and encoding it into generator parameters, it can capture reusable adversarial patterns rather than instance- and surrogate-specific optimization directions, while enabling efficient inference-time generation. Based on this insight, we propose Detector Adversarial Distribution Shift (DASH), which aggregates attack supervision over the training distribution and encodes reusable knowledge into an input-conditioned mapping from natural to adversarial images. DASH combines class-agnostic confidence suppression to coordinate multiple objects with CLIP feature disruption to reduce surrogate dependence. We instantiate this formulation with conditional GAN and Stable Diffusion generators. Across three datasets and thirteen detectors, DASH-GAN and DASH-SD achieve average ASRs of \(83.6%\) and \(80.08%\), outperforming the strongest baseline by up to \(36.14\) percentage points with generation times of \(0.02\) and \(0.75\) seconds per image. DASH-SD also averages \(28.14%\) ASR on two ODDefense-protected models, exceeding the strongest prior baseline by \(10.76\) percentage points. Thus, DASH balances black-box transferability, generation efficiency, and defense robustness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
adversarial attack, object detection, transferable black-box attacks

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_9HJ7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the adversarial robustness of object detectors under transferable black-box attack settings. The authors argue that conventional per-image iterative optimization attacks tend to overfit the surrogate detector and suffer from limited cross-model transferability, primarily due to cross-detector gradient misalignment and inter-object gradient conflict within multi-object scenes. To address these issues, the authors propose DASH, shifting the attack paradigm from per-image pixel optimization to distribution-level adversarial distribution learning via an input-conditioned generative mapping Tθ:Pnat→PadvT_\theta: P_{nat} \to P_{adv}. The method incorporates two core objectives: a class-agnostic confidence suppression loss LCA\mathcal{L}_{CA} to coordinate suppression across multiple objects and mitigate gradient conflict, and an object-region CLIP feature disruption loss DCLIP\mathcal{D}_{CLIP} to provide detector-external semantic supervision and reduce surrogate representation dependence. The framework is instantiated using a conditional GAN (DASH-GAN) and Stable Diffusion with LoRA (DASH-SD). Evaluated after training on VOC2012, experiments assess black-box transferability across 13 detectors on VOC2012, INRIA Person, and COCO-500, along with evaluations on generation efficiency, LPIPS perceptual quality, robustness against defended detectors, threshold sensitivity, and ablation studies.

Critical strengths
  1. Distribution-level transferable attack paradigm with empirical gains: Rather than optimizing perturbations per image independently, the paper trains a shared generator to learn a mapping from the natural to the adversarial distribution. In Table 4, under identical CA+CLIP objectives, black-box average ASR for per-image PGD is only 47.54% / 61.92%, whereas DASH-GAN achieves 85.21% / 86.24% with substantially lower variance across unseen victims. This supports the premise that distribution-level generative modeling can enhance black-box transferability.
  2. Extensive experimental evaluation: The evaluation spans three benchmarks (VOC2012, INRIA Person, COCO-500) across 13 victim detectors covering one-stage, two-stage, end-to-end, and Transformer-based architectures. The authors also report cross-dataset transfer, evaluation against defended detectors (ODDefense), runtime efficiency, LPIPS, threshold sensitivity, and loss ablations. At test time, DASH-GAN and DASH-SD require 0.02 s/image and 0.75 s/image respectively, eliminating per-image gradient optimization.
Critical weaknesses
  1. Severe issues with figures and AI use disclosure: Certain components of Figure 1 appear to be direct outputs from text-to-image generative models without manual curation. For instance, the detector diagram in "Transfer to unseen detectors" lacks coherent structural logic and exhibits blurred boundaries, while meaningless vertical artifacts appear in the "Surrogate gradient misalignment" plot. However, the AI Use Statement only claims verification of code, references, and manuscript text, completely omitting any review of figures. Furthermore, the loss notation in Figure 2 conflicts with or is undefined in the main text equations, and Figures 4 and 5 suffer from low resolution and distortion, hindering readability.
  2. Unclear origin of the spatial gate MM in DASH-SD and threat model inconsistency: Section 3.5 (page 6) defines MM as an "object-region gate obtained from clean image boxes." In a realistic black-box scenario, the attacker receives unlabeled input images. If MM is derived from ground-truth bounding boxes, this constitutes test-time ground-truth leakage. If MM is predicted by the surrogate detector, the surrogate must be queried during inference, contradicting the claim of "one-shot generation" without surrogate dependency at test time. Neither the main text nor Algorithm 1 clarifies how MM is obtained during inference, representing a major logical gap.
Questions for authors
  1. Source of the spatial gate MM during inference in DASH-SD: In Equation (10), the spatial mask MM modulating LoRA is defined as "obtained from clean image boxes." During black-box inference on unannotated images, how is MM acquired? Does this rely on ground-truth boxes? If it relies on predictions from the surrogate detector, does that mean the surrogate must be executed at test time, undermining the claim of one-shot generation without surrogate involvement? Please clarify the inference pipeline and threat model.
  2. Fairness of baseline comparisons and mixture of white-box/black-box evaluations: In Table 1, the reported DASH results are presented across three surrogate detectors and include white-box evaluations where surrogate and victim coincide, whereas baseline methods are reported with only a single column. Could the authors provide a strictly controlled black-box comparison matrix where surrogate detectors, iteration budgets, and evaluation protocols are fully aligned, as well as specify the iteration counts, step sizes, and random restarts used for iterative baselines such as DAG, TOG, and CAT?
  3. Generation provenance of Figure 1 and AI disclosure: Was Figure 1 generated using a text-to-image model? Why did the AI Use Statement omit the disclosure and verification of figures?
  4. Inconsistency between main and ablation metrics: Tables 1–5 report object-level ASR, whereas Table 7 switches to the stricter image-level ASR. What is the justification for switching metrics, and what are the corresponding results for baseline methods under the stricter image-level ASR standard?
Reviewer_fPTfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DASH learns reusable, input-conditioned adversarial transformations for object detectors. It combines class-agnostic confidence suppression with object-region CLIP feature disruption, implemented through a conditional GAN or LoRA-adapted Stable Diffusion. Experiments across thirteen detectors and three datasets report improved transferability, with additional evaluations of defended models, generation time and perceptual similarity.

Critical strengths

The paper addresses the relevant problem of transferable adversarial attacks on object detectors. The evaluation covers multiple detector architectures, datasets and defended models, providing a broad empirical assessment. The ablation study in Table 7 supports the complementary contributions of class-agnostic confidence suppression and CLIP feature disruption.

Critical weaknesses

W1. Limited support for the gradient-based motivation. Figures 3(c) and 6 do not specify the detector, image set or object-pair aggregation, limiting reproducibility. The cross-detector analysis in Section 3.1 uses only DAG, while Figure 6 compares CE, DAG and CA for inter-object alignment. It remains unclear how broadly the cross-detector findings extend to other attack objectives. W2. Incomplete reporting of attack and timing settings. Section 4.1 refers to the baselines' original settings, but the actual iteration budgets, PGD-CA+CLIP optimization settings and timing protocol are not specified. This limits assessment of the attack comparisons in Tables 1, 4 and 9 and the effectiveness-efficiency comparison in Figure 4, supported by the timing results in Table 6. Training time is also unreported, preventing assessment of the learned attacks' total computational cost. W3. Figure readability. Figures 1 and 2 have crowded layouts and small labels that make the method diagrams difficult to follow. Figure 4 appears vertically compressed, reducing the readability of its axes and method labels.

Questions for authors
  1. Which detectors and image samples were used for Figures 3(c) and 6, and how were object pairs selected and aggregated? These choices affect the interpretation and reproducibility of the reported gradient statistics.
  2. For the iterative baselines in Table 1 and PGD-CA+CLIP in Tables 4 and 9, what iteration counts, step sizes and stopping criteria were used? Without these details, the fairness of these attack comparisons is difficult to assess.
  3. What hardware and batch size were used for the runtime measurements in Figure 4 and Table 6, and does the DASH-SD runtime include DDIM inversion? These details are needed to assess the comparability and completeness of the reported generation times.
  4. For Table 8, were LoRA parameters shared across budgets or trained separately, and what training time was required? If each budget requires separate training, evaluating multiple perturbation budgets would incur repeated training costs.
Reviewer_VqM1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DASH (Detector Adversarial Distribution Shift), a framework for generating transferable adversarial images against object detectors in black-box settings. The approach is motivated by two main observations: (1) Detection gradient inconsistency: Different object detectors produce poorly aligned gradients, limiting the transferability of adversarial examples; (2) Objects gradient conflict: Gradients associated with different objects in the same image may interfere, reducing the effectiveness of attacks against multiple objects.

The authors implement the approach using two generative architectures: DASH-GAN and DASH-SD (Stable Diffusion).

The evaluation compares DASH against six existing adversarial attack methods and reports improvements of up to 36.14 percentage points in attack success rate.

Critical strengths
  1. Interesting combination of techniques to improve attack transferability:
  • Distribution-level learning, where a generator learns image-dependent adversarial transformations across a training dataset rather than optimizing each image independently.
  • Class-agnostic confidence suppression to reduce conflicts between objects.
  • CLIP-based feature disruption to reduce dependence on the surrogate detector.
  1. Implementation using two generative architectures: GAN and Stable Diffusion, showing that the general framework can be instantiated using different generation mechanisms.

  2. Extensive experimental evaluation across various detection models, three datasets, and two generative architectures, reporting substantial improvements in attack success rates.

Critical weaknesses
  1. The main concern is about the evaluation setup and fairness of comparisons. The authors report training DASH on 11,987 VOC2012 images and evaluating it on 1,713 images. These appear to be custom subsets rather than the standard VOC2012 detection splits (11,540 training/validation images). Furthermore, the compared methods were originally evaluated under different experimental settings. For example:
  • TOG was evaluated using VOC2007 + VOC2012 configurations (6, 551 training images and 4, 952 testing images)
  • AFOG was evaluated on the VOC2007 test set containing 4,952 images, as well as COCO. It is not sufficiently clear whether all baseline attacks were re-evaluated on exactly the same test images. A more detailed explanation of dataset construction and baseline reproduction is needed to establish that the reported improvements are based on fair comparisons.
  1. The authors compare DASH with existing attack methods, but the criteria for selecting these particular approaches are not sufficiently explained. Moreover, the related work section discusses additional categories of adversarial attacks, while the compared attacks are not explained in details, especially its relevance for comparison.

  2. The system is trained using three surrogate detectors: YOLOv5, YOLO11s, and Faster R-CNN. The resulting attacks are then evaluated on multiple YOLO variants and other detection architectures. The choice of surrogate model can significantly influence transferability. However, the paper does not sufficiently analyze how surrogate architecture, model size, and model generation affect the reported results. For example, YOLO11s generally achieves stronger transferability than YOLOv5 in several experiments. However, without knowing the exact YOLOv5 configuration, it is difficult to determine whether these differences are related to architecture, model capacity, or training.

  3. I identified the following reference issues:

  • G-UAP: The reference incorrectly lists the proceedings editors as paper authors.
  • LoRA: The venue details of the cited as ICLR, 1(2):3, 2022 are incorrect. The original preprint is arXiv:2106.09685 (2021), and the work was published at ICLR 2022.
  • BadPatch: The paper is cited as an ICCV main-conference publication, pages 6244–6254. The correct publication is 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 6303–6313.
  • PASCAL VOC2012: The authors cite the original VOC challenge publication (Everingham et al., 2010). Although this is a valid foundational reference, the 2015 retrospective publication or official VOC2012 benchmark documentation would provide a more specific reference for the dataset used in the evaluation.
Questions for authors
  1. Evaluation aspects: Could the authors clarify how the dataset construction? Were all baseline attacks evaluated on the same test images, with identical victim detector configurations and evaluation metrics? Were the baseline results reproduced or obtained from the original publications?

  2. What criteria were used to select the six compared approaches?

  3. Authors are requested to provide more detailed experimental configurations for reproducibility, particularly for the baseline implementations, dataset construction, training procedures, and model checkpoints?

  4. Please specify the exact variants of YOLOv2, YOLOv5, and YOLOv8 used in the experiments, similar to the explicit designation of YOLO11s.Also relate in evaluation results the aspects of YOLO model sizes.

  5. I identified several inaccuracies in the references, including incorrect publication details for LoRA, BadPatch, and G-UAP.

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

Related papers

Discussion (0)

Sign in to comment.