acceptodds
Under review as a conference paper at ICLR 2027

CtrlCamo: Controllable Adversarial Camouflage Editing for Vehicle Detection

Abstract

Deep neural networks (DNNs) have achieved remarkable success in computer vision but remain vulnerable to adversarial attacks. For vehicle detection, many existing attacks rely on imperceptible perturbations or localized adversarial patches, which can limit attack effectiveness or introduce visually conspicuous patterns. In this paper, we introduce , a controllable framework for adversarial camouflage editing, which edits vehicle appearance to reduce detector performance while maintaining visual consistency with the environment. CtrlCamo supports two stylization strategies: image-level, which adapts vehicles to their surroundings, and scene-level, which aligns them with semantic concepts in the scene. A two-stage ControlNet fine-tuning pipeline jointly enforces structural preservation, style consistency, and adversarial effectiveness. Experiments on COCO and LINZ demonstrate at least 38% reduction over prior methods while improving structural preservation and perceptual consistency. The resulting adversarial effects also transfer to unseen black-box detectors. Code and data will be released upon paper acceptance to facilitate further research.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Digital Attack, Adversarial Camouflage Editing, Diffusion Models, Object Detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xYfZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an image editing method that uses ControlNet to generate camouflage on images, with both image-level and scene-level generation approaches, achieving better attack effects while preserving the vehicle structure and improving visual concealment for the human eye.

Critical strengths
  1. For the first time, the camouflage attacks on real images are formalized as a way to fine-tune a control net, which is very novel, and two specific camouflage strategies are proposed.
  2. The attack is highly effective, capable of reducing the detector's AP 50 by more than 38%.
  3. A complex loss combination is designed to ensure that the generated paint patterns meet expectations.
  4. Significant improvement compared to other baselines.
Critical weaknesses
  1. Ablation experiments on the effects of the two stages should include in the main text and need quantitively result.
  2. Lack the human experiments details, like the number of participants.
Questions for authors

Same as weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer xYfZ

W1. Quantitative ablation of the two-stage design.

Table 1. Ablation Study (Image-Level, Faster-RCNN). We report AP50\mathrm{AP}_{50}, SSIM, LPIPS, and SWD on COCO and LINZ. SSIM and LPIPS are computed on grayscale crops around each vehicle following reviewer QR3f Q3/Q5. SSIM values therefore differ from those in the original submission. Bold = best, italic = second best.

DatasetMethodAP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓SWD ↓
COCOOurs15.00.8020.2580.121
COCOOne-stage23.80.7900.2650.131
COCOW/O LbL_b29.00.7640.2810.139
COCOW/O LcL_c25.60.7880.2650.126
COCOW/O LstructL_{struct}22.30.6960.3190.158
COCOW/O LsL_s27.90.8230.2610.383
LINZOurs18.30.9160.2300.074
LINZOne-stage36.40.9250.2180.090
LINZW/O LbL_b38.10.9170.2120.087
LINZW/O LcL_c29.10.9270.2140.095
LINZW/O LsL_s27.80.9440.2180.090

 

Table 2. Ablation Study (Scene-Level, Faster-RCNN).

DatasetMethodAP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓SWD ↓
COCOOurs16.60.8020.2670.107
COCOOne-stage8.800.7860.2760.129
COCOW/O LcL_{c}21.00.8010.2680.133
LINZOurs27.50.9120.2220.109
LINZOne-stage59.70.9180.2230.138
LINZW/O LcL_{c}31.20.9340.2070.162

 

We thank the reviewer for this suggestion. We added a quantitative ablation of the two-stage design under the same total training budget and hyperparameters (Table 1 and Table 2 in this response), comparing one-stage joint training, two-stage training without the color-consistency loss LcL_c, and the complete model, using AP50\mathrm{AP}_{50} for attack effectiveness, SSIM and LPIPS for structural preservation, and reference-style SWD for appearance consistency. The results support our motivation that jointly optimizing all objectives is unstable: one-stage training sacrifices attack effectiveness, structural preservation, or both in every setting, and its AP50\mathrm{AP}_{50} varies widely across settings, from 8.8% on COCO scene-level to 59.7% on LINZ scene-level, whereas the complete model remains between 15.0% and 27.5%. For example, on LINZ image-level, two-stage training without LcL_c reduces AP50\mathrm{AP}_{50} from 36.4% to 29.1% compared with one-stage training, and the complete model further reduces it to 18.3% while lowering reference-style SWD from 0.090 to 0.074. Overall, the complete model achieves lower SWD than one-stage training in all four settings and lower AP50\mathrm{AP}_{50} in three of the four; the exception is COCO scene-level, where one-stage training yields a stronger attack (8.8% vs. 16.6%) at the cost of lower structural and style consistency (SSIM 0.786 vs. 0.802; SWD 0.129 vs. 0.107).

As the reviewer suggests, we moved this quantitative ablation into Section 4.3 (page 9) of the revised manuscript, alongside the qualitative examples.

 

W2. Human evaluation details.

We thank the reviewer for pointing this out and we provide the participant statistics below. In the comparison with existing methods (Appendix D.2), we recorded 758 responses from 26 participants, 22 of whom completed all 30 questions (15 image-level and 15 scene-level), drawn from a pool of 30 images covering 10 scene types with three images each. In the ablation study (Appendix D.3), 15 participants each answered 10 questions randomly drawn from image-level groups on COCO and LINZ. We received 144 responses. The complete model was preferred over its ablation variants in 94.6% (95% CI: 86.9–97.9%) of COCO responses and 71.4% (95% CI: 60.0–80.7%) of LINZ responses, well above the chance rate of 33.3%. All participants were unpaid volunteers, and in both studies the options were unlabeled and presented in randomized order, which means participants were blind to the method. Confidence intervals are 95% Wilson score intervals.

We added participant and response counts in Appendix D.2 (page 26) and Appendix D.3 (page 27), including the COCO/LINZ response breakdown for the ablation study. We hope these additions address the reviewer's concerns, and we thank the reviewer again for the positive assessment of our work.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer xYfZ,

Thank you again for your valuable feedback and for taking the time to review our paper.

As the discussion period is approaching its end, with approximately three days remaining, we would like to kindly follow up on our response submitted three days ago. We hope our responses have adequately addressed your questions and concerns.

If there are any remaining concerns or points that require further clarification, we would greatly appreciate your feedback and would be happy to provide additional explanations before the discussion deadline.

Thank you again for your time and consideration.

Best regards, The Authors

Thank the authors for the detailed response. My concerns have been addressed with the revised claim and additional experiments. I will keep my previous score.

AuthorsOpenReview ↗

Thank you for your feedback and for supporting our previous rebuttal.

Dear Reviewer xYfZ,
Thank you for your feedback and for supporting our work and rebuttal.

Reviewer_QR3fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CtrlCamo for adversarial vehicle camouflage editing. The method first learns structure-preserving stylization and then introduces detector supervision to improve attack effectiveness. It supports references from either the local surroundings or scene-level concepts. Experiments cover COCO, LINZ, black-box transfer, and projection recapture tests.

Critical strengths
  1. The two-stage design is well motivated. Learning to reconstruct and stylize vehicles before applying the adversarial loss helps address the object distortion observed in the one-stage variant. The two reference strategies also offer useful control over camouflage appearance.
  2. The detection results are strong. On COCO, the image-level attack reduces Faster R-CNN AP50 from 85.6% to 15.0%. The transfer results on YOLOv5 and YOLOv8 further strengthen the empirical contribution.
Critical weaknesses
  1. Diff-PGD already separates style transfer from adversarial optimization to improve stability. CtrlCamo instead learns a conditional generator, but the paper does not sufficiently isolate the contribution of generator training and contextual reference selection beyond the existing two-stage design.
  2. The two-stage ablation is confounded by the color-consistency loss. The one-stage variant described in Appendix C.5, while the complete two-stage model includes it. The improvement cannot be attributed to staged training without controlling this difference and the total training budget.
Questions for authors
  1. Do the baseline methods optimize over the same vehicle masks as CtrlCamo? Are their camouflage styles subject to the same constraints?
  2. Diff-PGD already separates stylization from adversarial optimization. What is the main methodological advance over this design.
  3. Could you add LPIPS on vehicle crops as a complementary perceptual measure, alongside SSIM and the human study, to strengthen the analysis of perceptual quality and naturalness?
  4. Can you compare the one-stage variant with a two-stage variant without Lc, using the same total training budget, and evaluate their attack effectiveness and appearance quality.
  5. Is SSIM computed within the vehicle mask or over a crop containing background?
7 comments
AuthorsOpenReview ↗

Response to Reviewer QR3f

Q3/Q5. Clarification of SSIM computation and additional LPIPS evaluation.

We really appreciate the reviewer for pointing these out, which led us to re-examine and refine our evaluation protocol. Section 4.1 of the original submission (page 6) states that images were cropped before computing SSIM. Upon re-examining our pipeline, we found that this description was inaccurate, and SSIM was in fact computed at full-image resolution. Because many vehicles, particularly in LINZ, occupy only a small part of the image, large unchanged regions then dominate the score and compress the differences between methods. In the revised protocol, we extract a crop around each vehicle, defined by the bounding box of its ground-truth mask extended by 10 pixels, and resize it to 224×224 so that the vehicle occupies most of the evaluated region. The crop therefore contains a thin band of background around the vehicle rather than the vehicle mask alone. Because all methods composite the edited vehicle onto the same real background, these pixels are identical across methods and do not favor any of them. In addition, we found that color differences strongly affect these metrics. This is because camouflage intentionally changes vehicle color, comparing RGB values penalizes the intended edit rather than measuring structural preservation. We therefore compute SSIM on grayscale (luminance) crops, averaged over the crops. Following the reviewer's suggestion, we also report LPIPS (VGG) on the same grayscale crops, replicating the luminance channel across three channels. All methods and ablation variants are evaluated with this protocol.

Under this protocol, as shown in Table 1 and 2 in this response (due to the space we put them to the next comment), CtrlCamo achieves the highest SSIM and lowest LPIPS in all COCO settings. For example, on COCO with Faster R-CNN (image-level), CtrlCamo obtains an SSIM of 0.802 and an LPIPS of 0.258, compared with 0.657 and 0.413 for Diff-PGD and 0.593 and 0.426 for AdvCAM. On LINZ, CtrlCamo achieves the best SSIM and LPIPS under scene-level stylization. Under image-level stylization, AdvCAM obtains higher similarity (SSIM 0.940 vs. 0.916; LPIPS 0.181 vs. 0.230), but it yields a substantially weaker attack. Since grayscale SSIM measures luminance structure, which our structure-preservation loss explicitly constrains, we report it as a measure of structural preservation and LPIPS as a complementary perceptual measure.

We have corrected and expanded the evaluation description of SSIM and LPIPS in Section 4.1 (page 7). We have replaced the SSIM results and added grayscale LPIPS results in Tables 1–2 (page 7) of the revised manuscript. We also updated the metric results in Table 5 (page 10), Table 6 (page 16), and Table 8 (page 20).

AuthorsOpenReview ↗

Response to Reviewer QR3f (Continuation on Q3/Q5)

Table 1. Quantitative comparison on the COCO dataset for stylization-based digital attacks. We report AP, AP50\mathrm{AP}_{50}, SSIM, LPIPS, and average sampling time. Bold = best.

DetectorLevelMethodAP (%) ↓AP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓Inf. Latency (s) ↓
Faster-RCNNimage-levelNormal63.385.6---
Faster-RCNNimage-levelAdvCAM45.064.70.5930.42621.8
Faster-RCNNimage-levelDiff-PGD45.061.10.6570.41333.2
Faster-RCNNimage-levelCtrlCamo8.9015.00.8020.2587.32
Faster-RCNNscene-levelNormal63.385.6---
Faster-RCNNscene-levelAdvCAM44.766.40.5760.44721.8
Faster-RCNNscene-levelDiff-PGD47.164.60.6410.43433.2
Faster-RCNNscene-levelCtrlCamo10.716.60.8020.2677.32
ViTDetimage-levelNormal72.591.4---
ViTDetimage-levelAdvCAM48.569.70.5950.42421.8
ViTDetimage-levelDiff-PGD53.070.40.6520.41334.2
ViTDetimage-levelCtrlCamo12.419.20.8040.2498.15
ViTDetscene-levelNormal72.591.4---
ViTDetscene-levelAdvCAM47.667.90.5750.44721.8
ViTDetscene-levelDiff-PGD53.170.80.6360.43334.2
ViTDetscene-levelCtrlCamo6.6012.50.8070.2638.15

 

Table 2. Quantitative comparison on the LINZ dataset for stylization-based digital attacks.

DetectorLevelMethodAP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓Inf. Latency (s) ↓
Faster-RCNNimage-levelNormal98.3---
Faster-RCNNimage-levelAdvCAM90.50.9400.18121.8
Faster-RCNNimage-levelDiff-PGD73.20.8990.23432.0
Faster-RCNNimage-levelCtrlCamo18.30.9160.2307.00
Faster-RCNNscene-levelNormal98.3---
Faster-RCNNscene-levelAdvCAM88.60.9060.23121.8
Faster-RCNNscene-levelDiff-PGD66.40.8980.23932.0
Faster-RCNNscene-levelCtrlCamo27.50.9120.2227.00
ViTDetimage-levelNormal97.8---
ViTDetimage-levelAdvCAM84.10.9410.18121.8
ViTDetimage-levelDiff-PGD62.30.8990.23732.4
ViTDetimage-levelCtrlCamo13.70.9150.2277.67
ViTDetscene-levelNormal97.8---
ViTDetscene-levelAdvCAM80.80.9040.23021.8
ViTDetscene-levelDiff-PGD59.60.8980.23932.4
ViTDetscene-levelCtrlCamo11.10.9150.2167.67
AuthorsOpenReview ↗

Response to Reviewer QR3f

W1/Q2. Methodological distinction from Diff-PGD.

We thank the reviewer for these comments. We agree that both methods separate stylization from adversarial optimization, but they differ in both how they stylize and how they treat the stylized result. For stylization, Diff-PGD applies per-image neural style transfer from a square reference patch, whereas CtrlCamo matches region-level feature statistics (Eq. 5), which allows irregularly shaped references such as the ring surrounding the vehicle or a scene concept taken from a different, synthesized image. For the stylized result, Diff-PGD keeps the stylized image fixed and confines the attack to a norm-bounded perturbation that per-image PGD adds on top of it. In CtrlCamo, the stylization is not fixed: the second stage continues to update the generator with the adversarial loss together with the style and color-consistency losses, so the camouflage appearance itself becomes adversarial while remaining consistent with the reference. Our reason for staging also differs: when all objectives are trained jointly, the adversarial loss can be lowered by failed reconstructions rather than by camouflage (Section 4.3, page 9 of the revised manuscript), and staging prevents this shortcut. Because the result is a learned generator rather than a per-image optimization, inference requires a single sampling pass, taking 7.3 s per image compared with 33.2 s for Diff-PGD on the same GPU.

We assess the main components through complementary ablations. Comparing the first-stage model with the complete model evaluates the benefit of second-stage adversarial fine-tuning with its associated regularization. On COCO with Faster R-CNN, AP50\mathrm{AP}_{50} decreases from 67.2% to 15.0% for image-level stylization and from 65.1% to 16.6% for scene-level stylization (Table 10, revised manuscript). Removing LsL_s evaluates the contribution of explicit reference-style supervision: reference-style SWD increases from 0.121 to 0.383 on COCO image-level stylization as shown in Table 3 in this response. Results with alternative scene–concept pairings further show that the method remains effective under the tested reference choices (Table 8, revised manuscript). Finally, we compare one-stage training with two-stage training without LcL_c to evaluate staging, and then compare the latter with the complete model to evaluate the additional effect of LcL_c as shown in Table 3 and Table 4 in this response (due to the space we put them to the next comment).

We clarified the methodological differences from prior stylization-based attacks in Section 2 (page 3) and revised the contribution statement in Section 1 (page 2). We also added quantitative ablations in Table 5 (page 10) and reference-style consistency ablations in Table 12 (page 25), with corresponding analyses in Section 4.3 (pages 9–10) and Appendix D.1 (page 26).

Reviewer_5zi8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CtrlCamo, a diffusion-based framework for generating adversarial camouflage against vehicle detectors. The method introduces two stylization strategies that adapt vehicle appearance to either its immediate surroundings or a representative visual concept within the scene. A two-stage ControlNet fine-tuning procedure is employed to preserve vehicle structure and style consistency while optimizing adversarial effectiveness. Experiments on COCO and LINZ demonstrate substantial reductions in detection performance compared with existing attack methods, together with improved structural preservation, visual consistency, and transferability to unseen detectors.

Critical strengths

The paper addresses an interesting problem by combining adversarial effectiveness with context-aware appearance editing. Unlike conventional perturbation-based attacks, the proposed framework attempts to generate visually coherent camouflage while preserving the structure of the target vehicle. The distinction between image-level and scene-level stylization is reasonable and provides flexibility for different environments.

The experimental results are encouraging. CtrlCamo achieves substantial performance improvements over the compared methods on both datasets and across different detection architectures. The evaluations also cover black-box transferability, perceptual consistency, inference efficiency, and several ablation studies. The additional analyses in the appendix, including alternative generative backbones, cross-view transferability, and imaging variations, further strengthen the empirical evaluation.

Critical weaknesses

The methodological novelty is somewhat limited. The framework primarily combines existing components, including ControlNet, latent perceptual features, structure-preserving constraints, and detector-oriented adversarial optimization. The two-stage training procedure is a reasonable design choice, but its underlying motivation is relatively straightforward. It remains unclear whether the proposed formulation offers a sufficiently substantial methodological contribution beyond integrating established techniques for a specific application.

The comparison with existing methods raises some concerns about fairness. CtrlCamo allows appearance modifications across the vehicle region, whereas several competing methods operate under more restrictive perturbation or patch-based settings. Moreover, the stylization-based baselines are adapted from classification attacks and use square reference patches, which may limit their ability to exploit contextual information. Consequently, the large improvements in attack performance cannot be attributed entirely to the proposed optimization framework. A more controlled comparison under comparable appearance-modification constraints would make the empirical conclusions more convincing.

The evaluation setting is relatively constrained despite the extensive experiments. The COCO evaluation uses only 400 selected test images containing a single vehicle, and the scene-level strategy relies on a small number of predefined scene–concept pairings. The method also depends on accurate vehicle segmentation masks, while the procedure for obtaining and selecting these masks at test time is not sufficiently clear. These assumptions leave some uncertainty about the applicability of the framework to more complex scenes with multiple objects and diverse environmental conditions.

The evidence supporting perceptual stealthiness could be more convincing. The main human evaluation measures how closely the generated camouflage matches a reference area, which is not necessarily equivalent to assessing whether the edited vehicle appears natural or inconspicuous in the original scene. Although the appendix includes a separate naturalness study, it primarily compares different variants of the proposed method. The relatively limited image pool and insufficient reporting of participant statistics also make it difficult to assess the reliability of these perceptual conclusions.

Questions for authors

See weakness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 5zi8

W1. Novelty.

Table 1. Ablation Study (Image-Level, Faster-RCNN). We report AP50\mathrm{AP}_{50}, SSIM, LPIPS, and SWD on COCO and LINZ. SSIM and LPIPS are computed on grayscale crops around each vehicle following reviewer QR3f Q3/Q5. SSIM values therefore differ from those in the original submission. Bold = best, italic = second best.

DatasetMethodAP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓SWD ↓
COCOOurs15.00.8020.2580.121
COCOOne-stage23.80.7900.2650.131
COCOW/O LbL_b29.00.7640.2810.139
COCOW/O LcL_c25.60.7880.2650.126
COCOW/O LstructL_{struct}22.30.6960.3190.158
COCOW/O LsL_s27.90.8230.2610.383
LINZOurs18.30.9160.2300.074
LINZOne-stage36.40.9250.2180.090
LINZW/O LbL_b38.10.9170.2120.087
LINZW/O LcL_c29.10.9270.2140.095
LINZW/O LsL_s27.80.9440.2180.090

 

Table 2. Ablation Study (Scene-Level, Faster-RCNN).

DatasetMethodAP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓SWD ↓
COCOOurs16.60.8020.2670.107
COCOOne-stage8.800.7860.2760.129
COCOW/O LcL_{c}21.00.8010.2680.133
LINZOurs27.50.9120.2220.109
LINZOne-stage59.70.9180.2230.138
LINZW/O LcL_{c}31.20.9340.2070.162

 

We thank the reviewer for this insightful question. We agree that CtrlCamo builds on established components, including ControlNet, LatentLPIPS features, one-step latent estimation, and a detector cross-entropy loss, and we do not claim novelty for these components individually. Our contribution lies in the problem formulation and in two design choices required to make it work. First, we formulate the task as amortized, context-conditioned adversarial editing. Camouflage-generation methods such as LAKE-RED blend objects into a scene but have no adversarial objective, while stylization attacks such as AdvCAM and Diff-PGD were designed for image classifiers and optimize each image separately using a square style patch. CtrlCamo instead learns a single conditional editor that takes a real image, a vehicle mask, and a context reference, and produces an adversarial edit in one sampling run. After a one-time training cost of 22k–35k iterations per setting, no per-image optimization is needed: generating an edit takes 7.3 s, compared with 33.2 s for Diff-PGD on the same GPU.

Second, our style loss (Eq. 5) builds on the classical insight that style can be represented by spatially invariant feature statistics, but applies it in a setting those losses do not address: it is computed on the one-step clean-latent estimate during diffusion or flow-matching training, in latent rather than pixel space, between a vehicle and a reference that differ in shape and position and may come from a different, synthesized image. Our formulation directly supports irregular, spatially unaligned reference regions, while the evaluated baseline implementations use square reference patches. Removing the loss raises reference-style SWD from 0.121 to 0.383 on COCO (Table 1 in this response).

Third, we found that when all objectives are trained jointly, the adversarial loss rewards failed reconstructions. Early in training, the one-step estimate often fails to reconstruct the vehicle, and the detector naturally assigns low confidence to such non-vehicle content, so the adversarial loss can be lowered by destroying the object rather than camouflaging it. Our two-stage design is intended to mitigate this failure mode. The first stage learns faithful reconstruction and stylization, and the second introduces the detector loss, with the color-consistency loss LcL_c using the frozen first-stage model as its appearance reference. In the new ablation under the same training iterations (Table 1 and Table 2 in this response), the full two-stage model achieves lower reference-style SWD than one-stage training in all four settings, and lower AP50\mathrm{AP}_{50} in three of four.

We revised the contribution statement in Section 1 (page 2) and a more detailed discussion of related methods in Section 2 (page 3). We have also added quantitative ablations in Table 5 (page 10) and reference-style consistency ablations in Table 12 (page 25), with corresponding analyses in Section 4.3 and Appendix D.1.

AuthorsOpenReview ↗

Response to Reviewer 5zi8

W2. Fairness of baseline comparisons.

Table 3. Square-reference comparison under Faster R-CNN.

MethodCOCO AP50\mathrm{AP}_{50} (%) ↓COCO SSIM ↑COCO LPIPS ↓LINZ AP50\mathrm{AP}_{50} (%) ↓LINZ SSIM ↑LINZ LPIPS ↓
AdvCAM64.70.5930.42690.50.9400.134
Diff-PGD61.10.6570.41373.20.8990.234
CtrlCamo15.00.8020.25818.30.9160.230
CtrlCamo (square reference)13.10.7660.28723.20.9010.239

 

We thank the reviewer for this comment and would like to address the three aspects of the concern in turn.

First, regarding editing freedom, we agree that the perturbation- and patch-based methods in Table 3 (page 7 of the revised manuscript) operate under different threat models, and we include them to motivate the problem rather than as like-for-like comparisons. The controlled comparison is instead provided by the stylization-based baselines in Tables 1 and 2 (page 7 of the revised manuscript): AdvCAM and Diff-PGD edit exactly the same vehicle masks as CtrlCamo, with the background unchanged. The editable spatial region is therefore identical, while each method retains its original stylization and optimization constraints.

Second, regarding the square reference patches used by AdvCAM and Diff-PGD, we examine whether CtrlCamo's advantage in attack strength comes from its richer access to context. Without the style loss LsL_s, as shown in Table 1 (please refer to the response to W1), CtrlCamo still reaches 27.9% AP50\mathrm{AP}_{50} on COCO and 27.8% on LINZ (Faster R-CNN, image-level), compared with 61.1% and 73.2% for Diff-PGD and 64.7% and 90.5% for AdvCAM. These results show that the attack advantage persists without explicit style-loss supervision. Meanwhile, removing LsL_s increases reference-style SWD from 0.121 to 0.383 on COCO, demonstrating its contribution to reference-style alignment.

Additionally, we evaluate a square-reference variant of CtrlCamo, matching the reference patch sampling format used by these baselines. As shown in Table 3, this variant achieves 13.1% AP50 on COCO and 23.2% on LINZ, compared with 61.1% and 73.2% for Diff-PGD and 64.7% and 90.5% for AdvCAM, respectively. This indicates the attack effectiveness persists when it uses square references.

We clarified the baseline implementation in Appendix C.1 (page 18), specifying that square patches define the style references rather than the editable regions. The stylization-based comparisons are reported in Tables 1 and 2 (page 7).

AuthorsOpenReview ↗

Response to Reviewer 5zi8

W3. Evaluation scope and segmentation masks.

Table 4. Quantitative comparison on COCO dataset containing multiple vehicles. We report AP\mathrm{AP}, AP50\mathrm{AP}_{50}, SSIM, and LPIPS.

DetectorLevelMethodAP (%) ↓AP50\mathrm{AP}_{50} (%) ↓SSIM ↑LPIPS ↓
Faster-RCNNimage-levelNormal42.868.8--
Faster-RCNNimage-levelCtrlCamo11.721.50.7500.237
Faster-RCNNscene-levelNormal42.868.8--
Faster-RCNNscene-levelCtrlCamo4.908.700.7480.256
ViTDetimage-levelNormal49.072.3--
ViTDetimage-levelCtrlCamo4.007.800.7380.252
ViTDetscene-levelNormal49.072.3--
ViTDetscene-levelCtrlCamo4.808.900.7490.255

 

We thank the reviewer for these important questions and experiments suggestions and would like to address each aspect below. First, regarding scenes with multiple objects, we additionally evaluate CtrlCamo on 300 randomly selected COCO images containing multiple vehicles, comprising 798 vehicle instances in total. We use the model trained on single-vehicle images without any retraining, generate camouflage for every vehicle in each image independently using its segmentation mask, and composite all edited vehicles into a single image for evaluation. As shown in Table 4 in this response, CtrlCamo reduces Faster R-CNN AP50\mathrm{AP}_{50} from 68.8% to 21.5% (image-level) and 8.7% (scene-level), and ViTDet AP50\mathrm{AP}_{50} from 72.3% to 7.8% and 8.9%, respectively. These results show that CtrlCamo generalizes to multi-vehicle scenes it was not trained on. We added this experiment to Appendix E.1 (pages 27–28), with quantitative results in Table 13 (page 27) and visualizations in Figure 20 (page 28).

Second, regarding the evaluation set and scene–concept pairings, the 400 COCO test images were randomly sampled from test images containing a single vehicle. The scene–concept pairings are not fixed assumptions of the method but a controllable input: scene types are assigned automatically by an MLLM, and each concept is chosen from the most frequent objects detected in that scene type (Appendix A), so the pairing can be changed to suit the deployment environment. Across both datasets, our pairings cover ten scene types, and an alternative set of pairings (Appendix C.3, Table 8, page 20 of the revised manuscript) yields comparable results, with AP50\mathrm{AP}_{50} of 21.1% and 20.2% on COCO and 24.8% and 14.5% on LINZ for Faster R-CNN and ViTDet, respectively, compared with clean baselines of 85.6%, 91.4%, 98.3%, and 97.8%.

Third, regarding diverse environmental conditions, Appendix E.2 (revised manuscript) places the same camouflaged vehicles, with unchanged appearance, into new backgrounds within the same scene category, covering different weather (cloudy, rainy, foggy) and seasons (summer, winter). Faster R-CNN AP50\mathrm{AP}_{50} drops from 99.5% on the clean composites to 38.2% on the camouflaged ones. In addition, the attack remains effective after denoising and smoothing (Appendix C.4, revised manuscript) and after projection and camera recapture under changed illumination (Appendix E.4, revised manuscript).

Fourth, regarding test-time segmentation, we prompt SAM 2 with a single point near the center of each target vehicle and manually select the candidate mask that best covers the vehicle. This procedure requires the attacker to identify the target vehicle and select its segmentation mask. CtrlCamo, AdvCAM, and Diff-PGD use identical test-time masks to define their editable regions. We clarified this human-assisted mask-selection procedure in Section 4.1 (page 6).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Reviewers and Area Chair,
We sincerely thank you for your time and constructive feedback. We have revised the manuscript and provided additional experiments and clarifications to address the reviewers’ questions and concerns. Changes are highlighted in red in the revised manuscript. The main revisions are summarized below.

1. Methodological distinction and baseline comparison (5zi8 W1–W2; QR3f W1, Q1–Q2; page 2–3, 18)

We clarified the distinction between CtrlCamo and stylization-based per-image optimization methods (AdvCAM, Diff-PGD) and highlight our contributions (page 2–3). We also specified that all methods edit identical vehicle masks with unchanged backgrounds, while retaining method-specific appearance constraints (page 18). We additionally evaluated our method using square-patch style references. Under Faster R-CNN evaluation, this variant achieves 13.1% AP50\mathrm{AP}_{50} on COCO and 23.2% on LINZ, remaining stronger than both baselines (please refer to Table 3 in our response to 5zi8). These results show that the attack effectiveness of our pipeline persists regardless of the style loss.

2. Test-time mask acquisition (5zi8 W3; page 6)

We clarified that SAM 2 is prompted with a single point near the center of each target vehicle, followed by manual selection of the candidate mask that best covers the vehicle.

3. SSIM protocol and additional LPIPS evaluation (QR3f Q3/Q5; page 7, 16, 20)

We corrected the SSIM evaluation description and recomputed SSIM on 224 ×\times 224 grayscale vehicle crops (page 7). We also added LPIPS (VGG) on the same crops and updated the corresponding results in Tables 1–2, 6, and 8.

4. Controlled training-setting Ablation (5zi8 W1-W2; QR3f W2/Q4; xYfZ W1; page 9–10, 25–26).

We moved the discussion of one-stage ablation study from the appendix to the main paper under Section 4.3, and added additional quantitative comparisons of one-stage training, two-stage training without LcL_c, and the complete model under identical total training iterations (Table 5). Corresponding reference-style color-distribution similarity SWD results are provided in Table 12. Our strategy achieves lower SWD in all four settings and lower AP50\mathrm{AP}_{50} in three of four settings than one-stage training. We also discussed the improvement from the first-stage model to the complete model after adversarial fine-tuning (Table 10). Removing LsL_s preserves a substantial attack advantage over the baselines, while increasing reference-style SWD from 0.121 to 0.383 on COCO (please refer to Table 1 in our response to 5zi8).

5. Human-study reporting and scope (5zi8 W4; xYfZ W2; page 26–27).

We added participant and response counts for the baseline study. We clarified that the studies assess reference-style agreement across methods and relative whole-scene naturalness among ablation variants, respectively.

6. Generalization to multi-vehicle scenes (5zi8 W3; page 27–28).

We added an evaluation on 300 COCO images containing 798 vehicles, using models trained on single-vehicle images without additional training (Table 13, Figure 20).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.