acceptodds
Under review as a conference paper at ICLR 2027

CamouPurifier: Zero-shot and Camouflage-Agnostic Defense Against Adversarial Camouflage Attacks on Monocular Depth Estimation

Abstract

Adversarial camouflage attacks pose a significant threat to monocular depth estimation (MDE) systems in autonomous driving due to their physical realizability and robustness across diverse viewpoints. Existing adversarial defenses primarily focus on imperceptible perturbations or localized adversarial patches, rendering them ineffective against full-surface adversarial camouflage textures and potentially introducing additional geometric distortions. To address these limitations, we propose CamouPurifier, a zero-shot and camouflage-agnostic adversarial purification framework that incorporates adversarial camouflage localization and texture reconstruction. Specifically, we first employ a low-confidence-threshold detector to identify candidate vehicles and then integrate semantic-aware localization (SAL) with multi-domain texture anomaly localization (MTAL) to generate precise adversarial camouflage masks. Subsequently, we remove the localized adversarial patterns and reconstruct the missing vehicle textures using a geometry-guided diffusion model conditioned on vehicle edge information. To integrate the reconstructed textures into the original image, we further apply boundary-aware blending to maintain visual consistency with the surrounding scene. Comprehensive experiments demonstrate that CamouPurifier reduces the vehicle-region depth discrepancy by 84.5% compared with the undefended setting and by 51.2% compared with the strongest baseline. Its fidelity-oriented configuration further reduces global and background discrepancies by 53.7% and 41.2%, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Attack, Adversarial Defense, Purification, Camouflage Attack

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_iCcGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies defense against physically realizable, full-surface adversarial camouflage attacks on monocular depth estimation. The authors propose CamouPurifier, a zero-shot preprocessing framework consisting of camouflage localization and texture reconstruction. It first detects candidate vehicle regions and combines DINO-based semantic-aware localization with several handcrafted texture-anomaly cues to identify suspicious camouflage pixels. The localized regions are then removed and optionally reconstructed using Stable Diffusion inpainting guided by Canny edges through ControlNet, followed by Gaussian boundary blending. Experiments in CARLA cover four attacks and six monocular depth estimation models, with additional evaluations across weather conditions, vehicle types, an adaptive attack, and a small-scale physical case study. The reported results show substantial reductions in vehicle-region depth discrepancy compared with no defense and the Jedi, PAD, and APEX baselines. The localization-only and reconstruction-based variants provide different trade-offs between attack suppression, image fidelity, and computational cost.

Critical strengths
  1. Clear and easy-to-follow presentation. The paper is generally well organized, and the motivation, threat model, and overall defense pipeline are presented clearly. The decomposition into camouflage localization and texture reconstruction makes the proposed approach intuitive and easy to understand.

  2. Meaningful and practically relevant problem setting. Full-surface adversarial camouflage poses a realistic physical threat to monocular depth estimation systems, particularly in autonomous-driving scenarios. Compared with conventional imperceptible perturbations or localized patches, this threat remains relatively underexplored from the defense perspective.

  3. Strong empirical performance. CamouPurifier achieves substantial reductions in vehicle-region depth discrepancy across multiple attacks and MDE architectures. The results are generally consistent across different weather conditions and vehicle types, and the method compares favorably with the evaluated Jedi, PAD, and APEX baselines.

  4. Broad experimental coverage. The evaluation includes four attack variants, six MDE models, diverse environmental conditions, an adaptive attack, and a physical case study. The ablation study also helps clarify the contributions of the semantic localization, texture anomaly localization, and reconstruction components.

Critical weaknesses
  1. Limited practical efficiency. Although the paper motivates the defense in the context of autonomous driving, its runtime is far from real-time. Even the localization-only variant runs at only 2.705 FPS on an A100 GPU, while the complete reconstruction pipeline runs at 0.098 FPS, corresponding to approximately 10 seconds per image. It is also unclear whether these measurements include the downstream MDE inference. Therefore, the claim that the method is suitable for latency-sensitive scenarios is not sufficiently supported.

  2. Potential limitations of the evaluation protocol. For each attack–model pair, the authors rank 1,043 samples according to their undefended vehicle-region error and retain only the strongest 500 samples. Since this selection is performed before applying any defense, it does not directly bias the relative comparison among defense methods. However, the resulting evaluation measures recovery conditional on highly successful attacks and does not characterize performance over the original data distribution. Reporting results over all samples, together with attack success rates and per-condition statistics, would provide a more complete evaluation.

  3. Several claims are not fully supported by the current evaluation. First, although the paper reports clean-image depth discrepancy, it does not report the localization false-positive rate or predicted mask area on benign vehicles. Because both localization branches use percentile-based thresholds, it would be useful to evaluate whether naturally complex vehicle appearances, such as decals, reflections, or patterned paint, trigger unnecessary purification. Second, the adaptive attack is only briefly described, and the DINO percentile threshold is changed specifically for this experiment rather than fixed beforehand. More details about the adaptive objective and optimization procedure are needed. Finally, while the aggregate results are strong, the defense can degrade performance for some individual attack–model pairs. For example, under MIMDepth and 3D²Fool, the vehicle-region discrepancy increases from 2.860 without defense to 3.771 and 3.209 for the two variants. A clearer discussion of such failure cases would help qualify the overall conclusions.

Questions for authors

Please see the questions raised in the Critical Weaknesses section.

6 comments
AuthorsOpenReview ↗

Author Response to Reviewer iCcG (Part 1/3)

1. Response to W1

W1: Limited practical efficiency. Although the paper motivates the defense in the context of autonomous driving, its runtime is far from real-time. Even the localization-only variant runs at only 2.705 FPS on an A100 GPU, while the CamouPurifier$^\ddagger$ pipeline runs at 0.098 FPS, corresponding to approximately 10 seconds per image. It is also unclear whether these measurements include the downstream MDE inference. Therefore, the claim that the method is suitable for latency-sensitive scenarios is not sufficiently supported.

We thank the reviewer for highlighting the computational overhead of CamouPurifier. Our main contribution lies in proposing train-free, adversarial camouflage-agnostic and effective localization and purification of full-surface adversarial camouflage without retraining or modifying downstream MDE models. We acknowledge that the current implementation does not yet meet the stringent real-time requirements of autonomous driving, particularly when diffusion-based reconstruction is enabled.

CamouPurifier provides two configurations with different computational costs. The localization-only variant achieves 2.705 FPS, bringing a 27.6× speedup compared with the CamouPurifier$^\ddagger$ pipeline (0.098 FPS). The CamouPurifier$^\ddagger$ additionally aims to preserve visual and structural consistency through generative restoration. Notably, even the CamouPurifier$^\ddagger$ variant runs faster than the evaluated PAD (0.020 FPS) and APEX (0.026 FPS) implementations, showing its favorable computational efficiency among reconstruction-based defenses. All reported FPS measurements include only the defense purification stage and exclude downstream MDE inference.

We agree that further acceleration is necessary for real-time deployment. Potential directions include diffusion model distillation to reduce iterative sampling overhead and keyframe-based selective purification to avoid redundant processing of every video frame.

AuthorsOpenReview ↗

Author Response to Reviewer iCcG (Part 2/3)

2. Response to W2

W2: Potential limitations of the evaluation protocol. For each attack–model pair, the authors rank 1,043 samples according to their undefended vehicle-region error and retain only the strongest 500 samples. Since this selection is performed before applying any defense, it does not directly bias the relative comparison among defense methods. However, the resulting evaluation measures recovery conditional on highly successful attacks and does not characterize performance over the original data distribution. Reporting results over all samples, together with attack success rates and per-condition statistics, would provide a more complete evaluation.

We sincerely thank the reviewer for this constructive suggestion. We agree that evaluating only the 500 samples with the largest undefended vehicle-region discrepancies primarily characterizes defense effectiveness against highly successful attacks, rather than performance over the complete data distribution.

Motivation for the original selection. Our original Top-500 evaluation was intended to focus on challenging adversarial scenarios in which camouflage attacks induce substantial depth prediction deviations. As the reviewer correctly noted, the selection was performed before applying any defense and therefore did not directly favor a particular defense method.

Additional evaluation on all scenes. To address this concern, we have extended the evaluation to all 1,043 scenes for Monodepth2 across four adversarial camouflage attacks: 3D2Fool, 3D-APA, 3D-SSAP, and RobCamou. All baseline defenses are evaluated on the same complete set of scenes.

Table 1. Full-distribution evaluation on Monodepth2. Lower is better.

3D2Fool3D-APA3D-SSAPRobCamou
MethodEI\mathcal{E}_\mathbf{I} ↓\downarrowEv\mathcal{E}_\mathbf{v} ↓\downarrowEb\mathcal{E}_\mathbf{b} ↓\downarrowEI\mathcal{E}_\mathbf{I} ↓\downarrowEv\mathcal{E}_\mathbf{v} ↓\downarrowEb\mathcal{E}_\mathbf{b} ↓\downarrowEI\mathcal{E}_\mathbf{I} ↓\downarrowEv\mathcal{E}_\mathbf{v} ↓\downarrowEb\mathcal{E}_\mathbf{b} ↓\downarrowEI\mathcal{E}_\mathbf{I} ↓\downarrowEv\mathcal{E}_\mathbf{v} ↓\downarrowEb\mathcal{E}_\mathbf{b} ↓\downarrow
No Defense1.3002.8651.0432.05713.8861.1842.19213.6841.1742.72419.8281.393
Jedi1.2692.0251.1171.2202.5641.1431.3042.9101.1451.4024.0411.194
PAD1.5821.3891.5811.6232.8451.5601.6893.9391.5711.9816.9091.674
APEX1.4882.5521.3291.9408.1241.3872.1789.3931.4072.24110.2311.521
CamouPurifier$^\dagger$1.2601.4711.2691.3661.5261.3791.3451.5121.3581.3991.6391.399
CamouPurifier$^\ddagger$0.9500.8160.9570.9921.1361.0010.9861.1800.9881.1482.3931.048

Key findings.

  1. Consistent effectiveness on the full distribution. CamouPurifier achieves lower vehicle-region discrepancies than all evaluated baseline defenses across all four attacks.
  2. Substantial attack suppression. Averaged across the four attacks, the vehicle-region discrepancy decreases from 12.566 without defense to 1.537 for masking-only purification and 1.381 for CamouPurifier$^\ddagger$, corresponding to reductions of 87.8% and 89.0%, respectively.
  3. Improved CamouPurifier$^\ddagger$ performance on the full set. The CamouPurifier$^\ddagger$ achieves the lowest average global and background discrepancies of 1.019 and 0.998, respectively. Its average vehicle-region discrepancy is also lower than that of masking-only purification on the complete evaluation set.
AuthorsOpenReview ↗

Author Response to Reviewer iCcG (Part 3/3)

3. Response to W3

Several claims are not fully supported by the current evaluation. First, although the paper reports clean-image depth discrepancy, it does not report the localization false-positive rate or predicted mask area on benign vehicles. Because both localization branches use percentile-based thresholds, it would be useful to evaluate whether naturally complex vehicle appearances, such as decals, reflections, or patterned paint, trigger unnecessary purification. Second, the adaptive attack is only briefly described, and the DINO percentile threshold is changed specifically for this experiment rather than fixed beforehand. More details about the adaptive objective and optimization procedure are needed. Finally, while the aggregate results are strong, the defense can degrade performance for some individual attack–model pairs. For example, under MIMDepth and 3D²Fool, the vehicle-region discrepancy increases from 2.860 without defense to 3.771 and 3.209 for the two variants. A clearer discussion of such failure cases would help qualify the overall conclusions.

We sincerely thank the reviewer for these constructive suggestions. We agree that a comprehensive evaluation should address false positives on benign vehicles, adaptive robustness, and individual cases where purification may degrade performance.

  • (1) False-positive mask rate (FPMR) We agree that clean-image depth discrepancy alone does not directly quantify unnecessary purification. Although the intersection of SAL and MTAL is intended to suppress spurious localization, percentile-based thresholding may still select regions on benign vehicles.

    Table 2: FPMR on 1,043 clean vehicles. Precision, Recall, F1, and IoU are localization scores on the attacked-image.

    FPMRPrecisionRecallF1IoU
    Jedi0.0280.5980.4410.5010.322
    PAD0.1280.3460.6920.4470.316
    APEX0.0600.4110.2200.2070.119
    ours0.1480.6300.7080.6520.495

    Within the vehicle texture region, Jedi, PAD, APEX, and CamouPurifier mark 2.8%, 12.8%, 6.0%, and 14.8% of the pixels on average. On attacked images, the same CamouPurifier masks still give the highest Recall (0.708), F1 (0.652), and IoU (0.495).

    In addition, we evaluated three types of normal graffiti: Graffiti-1 consists of localized colored graffiti; Graffiti-2 consists of global colored graffiti covering the entire vehicle body; and Graffiti-3 consists of global black-and-white graffiti covering the entire vehicle body. For each category, 1,024 images with random distances and angles were collected for further evaluation. Ev\mathcal{E}_\mathbf{v} is compared with the ground truth depth. CamouPurifier achieves strong adversarial camouflage localization, but its texture-sensitive design can also respond to visually complex benign vehicle appearances. PAD achieves the lowest error rate, but it completely blackens the vehicle region, thereby destroying the scene's semantic information.

Table 3: FPMR and Ev\mathcal{E}_\mathbf{v} on graffiti vehicles.

Graffiti-1Graffiti-2Graffiti-3
FPMREv\mathcal{E}_\mathbf{v}FPMREv\mathcal{E}_\mathbf{v}FPMREv\mathcal{E}_\mathbf{v}
no defense-1.346-1.087-1.185
Jedi0.0821.920.3742.0450.2271.613
PAD0.3531.520.9521.1440.6731.26
APEX0.0772.0820.1082.6810.1492.694
ours0.4841.7750.8051.8340.4291.493
  • (2) Failure cases and limitations. We thank the reviewer for highlighting the importance of analyzing individual attack–model combinations. However, we would like to clarify a minor misinterpretation of the results in Table 1. The reported values of 3.771 and 3.209 correspond to the full-image discrepancy EI\mathcal{E}_\mathbf{I}, rather than the vehicle-region discrepancy Ev\mathcal{E}_\mathbf{v}. Specifically, under MIMDepth and 3D²Fool, the undefended vehicle-region discrepancy is 2.860. CamouPurifier† increases this value to 3.164, whereas CamouPurifier‡ reduces it to 2.573, achieving a 10.0% improvement over the undefended setting. Therefore, the complete reconstruction-based defense does improve vehicle-region depth consistency in this case. Nevertheless, we acknowledge that the full-image discrepancy increases from 3.134 without defense to 3.771 and 3.209 for the two variants, respectively. Our observations suggest that MIMDepth is particularly sensitive to localized appearance modifications, which can induce depth-prediction changes beyond the modified vehicle region.
Reviewer_YTYbOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CamouPurifier, a zero-shot defense framework against adversarial camouflage attacks in monocular depth estimation. It combines DINOv3 semantic features with multi-domain texture anomaly detection to localize camouflaged regions, then reconstructs vehicle textures using a geometry-guided diffusion model. Experiments demonstrate its defense effectiveness across different attacks, models, and environmental conditions.

Critical strengths
  1. The work targets full-surface vehicle camouflage attacks, which conventional localized patch defenses struggle to handle. The application scenario is clear and has substantial real-world relevance.
  2. The framework is relatively comprehensive, combining semantic and texture anomaly localization while attempting to reduce structural damage introduced by the defense through geometry-guided reconstruction.
  3. The evaluation is thorough, covering four attacks, six depth-estimation models, physical-scene tests, weather variations, and adaptive attacks. Compared with existing methods, it markedly reduces vehicle-region depth errors, while ablation studies support the effectiveness of the localization module.
Critical weaknesses
  1. The approach is similar to existing zero-shot localization and structure-guided restoration methods, such as APEX. Its main contribution appears to be the combination of existing modules, without a sufficiently distinctive technical mechanism being demonstrated.
  2. The value of the core reconstruction module is unclear. In Table 3, the full method yields higher vehicle-region depth error than the localization-and-masking-only variant. This raises questions about how the results support the claim that geometry-guided reconstruction improves the core defense effectiveness.
  3. The semantic anomaly localization mechanism is insufficiently explained. SAL constructs a mask using similarity between DINO CLS tokens and patch tokens, but it is unclear how this similarity accurately reflects the semantic inconsistencies caused by camouflage.
  4. Practical deployability is limited: the full method runs at only 0.098 FPS, and the evaluation relies mainly on simulated data with relatively limited real-world testing. This is insufficient to support real-time autonomous driving applications.
Questions for authors
  1. What is the key technical novelty of this work compared with APEX?
  2. Why does the full reconstruction method yield higher vehicle-region depth errors than the masking-only variant in some experiments?
  3. How does the DINO similarity used in SAL represent semantic anomalies?
5 comments
AuthorsOpenReview ↗

Author Response to Reviewer YTYb (Part 1/3)

1. Response to Q1

What is the key technical novelty of this work compared with APEX?

We thank the reviewer for raising this important question. Although CamouPurifier and APEX share the general locate-and-restore paradigm, our work addresses fundamentally different challenges and introduces task-specific designs for full-surface adversarial camouflage defense in monocular depth estimation (MDE). The key distinctions are summarized below.

  • Different threat models and technical challenges. APEX primarily addresses localized adversarial patches targeting object detection, where removing the malicious patch can be sufficient to restore detection performance. In contrast, CamouPurifier tackles full-surface adversarial vehicle camouflage targeting MDE, which introduces two coupled challenges. First, camouflage patterns are spatially distributed, irregular, and can cover most visible vehicle surfaces, making accurate localization substantially more difficult than identifying compact patches. Second, MDE is sensitive to depth-relevant structural and appearance cues. Consequently, simply removing adversarial textures may disrupt vehicle appearance and depth predictions, making structure-preserving reconstruction particularly important.

  • Camouflage-aware localization. APEX relies on mutual-information-based analysis and semantic segmentation to identify suspicious regions. However, these mechanisms are not specifically designed for extensive, irregular vehicle camouflage and may either miss substantial camouflage regions (Appendix Fig. 8(a)) or identify overly large portions of the vehicle as anomalous (Appendix Fig. 8(c)), resulting in incomplete purification or unnecessary removal of vehicle information . In contrast, CamouPurifier combines Semantic-Aware Localization (SAL), which exploits global–local feature alignment in DINO representations, with Multi-domain Texture Anomaly Localization (MTAL), which captures complementary appearance irregularities. This design enables more precise localization of spatially distributed camouflage patterns while reducing reliance on a single anomaly cue.

  • Geometry-guided reconstruction. APEX employs LaMa for context-aware image inpainting, which is effective when sufficient clean surrounding information is available. However, full-surface camouflage presents a more under-constrained reconstruction problem: once large portions of the vehicle are masked, limited clean vehicle texture remains to guide restoration, increasing the risk of introducing background appearance into vehicle regions (Appendix Fig. 8(b)). To address this limitation, CamouPurifier adopts Canny-conditioned diffusion inpainting with semantic prompts, leveraging additional structural and generative priors to reconstruct plausible vehicle appearances while suppressing adversarial textures.

To further distinguish the contributions of localization and reconstruction, we conducted additional experiments on Monodepth2 under 3D-APA, using both APEX without LaMa and CamouPurifier with LaMa reconstruction.

DefenseEI↓\mathcal{E}_\mathbf{I} \downarrowEv↓\mathcal{E}_\mathbf{v} \downarrowEb↓\mathcal{E}_\mathbf{b} \downarrow
No Defense2.05713.8881.184
APEX w/o LaMa1.6854.4881.454
CamouPurifier$^\dagger$1.3111.5031.323
APEX w/ LaMa1.9408.1241.387
CamouPurifier$^\dagger$ + LaMa1.0791.5921.059
CamouPurifier$^\ddagger$0.9921.1361.001

The results provide two complementary observations. First, without generative reconstruction, CamouPurifier$^\dagger$ reduces vehicle-region discrepancy from 4.488 to 1.503 compared with APEX without LaMa, demonstrating the effectiveness of our localization and masking strategy. Second, when using CamouPurifier's localization, replacing LaMa with our geometry-guided diffusion reconstruction further reduces vehicle-region discrepancy from 1.592 to 1.136, while also improving global and background consistency. These controlled comparisons demonstrate that our improvements arise from both the localization design and the reconstruction strategy, rather than merely combining existing modules.

AuthorsOpenReview ↗

Author Response to Reviewer YTYb (Part 2/3)

2. Response to Q2

Why does the full reconstruction method yield higher vehicle-region depth errors than the masking-only variant in some experiments?

We thank the reviewer for this insightful observation. This phenomenon can be explained by two factors: the depth-prediction bias introduced by black masking and the influence of attack-strength-based sample selection.

  • Masking-induced depth bias. Our observations indicate that MDEs (such as monodepth2) tend to predict relatively small depth values (approximately 1.54 on our evaluation scale) for vehicle regions covered by black masks, with strong and weak attacks yielding nearly identical results. Since our initial experiments screened 500 attacks with stronger effects, adversarial camouflage shifts the vehicle’s depth predictions toward larger values. Replacing the affected areas with black pixels may inadvertently bring some predictions closer to the clean reference values, thereby reducing the measured absolute deviation. However, this improvement does not necessarily imply the ability to faithfully restore the vehicle’s original appearance or depth structure.

  • Influence of evaluation sample selection. Our original evaluation selects the 500 samples with the largest undefended vehicle-region discrepancies. To investigate the effect of this selection, we additionally evaluate the remaining 543 samples and the complete set of 1,043 scenes on Monodepth2, averaged over four attacks.

Evaluation subsetDefenseEI↓\mathcal{E}_\mathbf{I} \downarrowEv↓\mathcal{E}_\mathbf{v} \downarrowEb↓\mathcal{E}_\mathbf{b} \downarrow
Top-500CamouPurifier$^\dagger$1.3191.5371.326
CamouPurifier$^\ddagger$1.0361.7140.989
Remaining 543CamouPurifier$^\dagger$1.3641.5371.375
CamouPurifier$^\ddagger$1.0041.0751.007
All 1,043CamouPurifier$^\dagger$1.3421.5371.351
CamouPurifier$^\ddagger$1.0191.3810.998

The results reveal an important distribution-dependent effect. While CamouPurifier$^\dagger$ achieves a slightly lower average vehicle-region discrepancy on the Top-500 subset, CamouPurifier$^\ddagger$ substantially outperforms masking on the remaining 543 samples, with a 30.1% reduction in Ev\mathcal{E}_\mathbf{v}. Consequently, over all 1,043 scenes, CamouPurifier$^\ddagger$ achieves a 10.1% reduction in average Ev\mathcal{E}_\mathbf{v} (1.381 vs. 1.537).

These findings demonstrate that a lower masking-based depth discrepancy does not necessarily imply more faithful restoration, and that the relative advantage of CamouPurifier$^\dagger$ and CamouPurifier$^\ddagger$ depends on the evaluated attack distribution.

AuthorsOpenReview ↗

Author Response to Reviewer YTYb (Part 2/3)

3. Response to Q3

How does the DINO similarity used in SAL represent semantic anomalies?

We thank the reviewer for this insightful question. The key intuition behind SAL is that full-surface adversarial camouflage can perturb the global–local semantic relationships encoded by DINOv3, introducing representation-level changes beyond low-level texture irregularities.

Specifically, the CLS token summarizes global visual information, while patch tokens encode localized features. Their cosine similarity measures global–local feature alignment. Since adversarial camouflage introduces artificial patterns across vehicle surfaces, it can influence both global and local representations, resulting in altered spatial similarity responses.

Importantly, high CLS–patch similarity does not directly indicate a semantic anomaly. Instead, SAL uses relatively high-response regions as potential camouflage candidates, exploiting the spatial distribution of DINO features rather than treating similarity as an explicit anomaly score. These candidates are further filtered by MTAL, which provides complementary evidence of abnormal texture characteristics. This combination achieves a Precision of 0.630, Recall of 0.708, F1-score of 0.652, and IoU of 0.495 in our localization evaluation.

We appreciate your thoughtful comments. We hope our response addresses your concerns. Please let us know if there are any additional questions, and we will be happy to discuss further.

Reviewer_bucbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to defend monocular depth estimation against full-surface adversarial vehicle camouflage. CamouPurifier combines DINO-based and texture-based masks within vehicle proposals, then blacks out the selected pixels or reconstructs them using Canny-conditioned diffusion and boundary blending. Experiments on selected CARLA scenes with six depth models and four attacks show large vehicle-discrepancy reductions from masking and better global/background fidelity from reconstruction, with additional ablations, a localization-aware attack, and a small physical experiment.

Critical strengths
  1. CamouPurifier's masking variant reduces average vehicle-region discrepancy by 84.5% compared with no defense across six depth models and four attacks.

  2. Comparing masking with reconstruction makes the trade-off between attack removal and image fidelity clear.

Critical weaknesses
  1. The adaptive attack is not described clearly enough. Appendices A.1/B.3 do not give the loss, optimization budget, or explain how the attack targets the final masking/reconstruction output. Since Eqs. (6) through (10) use percentile thresholds, simply lowering localization scores may leave the mask unchanged. Appendix A.4 also changes the DINO percentile from 30 to 20 without saying whether the attacker knew this setting. Selecting 500 samples by undefended error may miss larger errors after defense. I therefore cannot tell whether Figure 7 gives a strong test of an attacker that knows the complete defense.

  2. I am not sure what the depth errors measure. Appendix A.3 calls D true depth, but A.1 seems to compare against clean-model predictions; sum(I) in Eq. (17), depth conversions, and scale alignment are not clearly defined. These are different evaluation targets. The benefit of Canny guidance is also unclear: the headline 84.5% gain comes from masking, and Table 3 gives vehicle discrepancy 1.132 for SAL+TR versus 2.502 for the full method. The full method improves other metrics, but this does not show what Canny adds. Since APEX already uses structure-guided restoration, a comparison using the same inpainting setup without Canny would be helpful.

Questions for authors
  1. What are D, sum(I), and the valid-pixel mask in Eqs. (17) through (19)? Please clarify each model's depth/disparity conversion and scale alignment, how the reference fits A.1's clean-set explanation, and the reference and units for the physical result of 26.53 to 3.57.

  2. What loss, weights, optimization budget, and gradient or query method were used for the adaptive attack? Did the attacker know q_d=20 and target the final defended depth output, including reconstruction when used? Please also report results with fixed defense settings over all 1,043 adaptive scenes.

  3. Is Canny computed from the adversarial crop or the black-masked crop, and which evaluation split is used in Table 3? To test Canny's contribution, could you compare results with and without it while holding masks, crops, prompts, seed, and diffusion settings fixed, and report vehicle/global/background discrepancies?

7 comments
AuthorsOpenReview ↗

Author Response to Reviewer bucb (Part 1/3)

R1. Response to Q1 & W2 [Metric Details]

Q1: What are D, sum(I), and the valid-pixel mask in Eqs. (17) through (19)? Please clarify each model's depth/disparity conversion and scale alignment, how the reference fits A.1's clean-set explanation, and the reference and units for the physical result of 26.53 to 3.57.

We sincerely thank the reviewer for highlighting the ambiguity in our evaluation metrics and depth conversion procedures. We apologize for the imprecise notation in Appendix A.3 and clarify the evaluation protocol as follows.

  • For D\mathbf{D}:

    • In Eqs. (17)–(19), D\mathbf{D} denotes the depth prediction of the target MDE model on the clean scene, rather than the ground-truth depth. D~\widetilde{\mathbf{D}} denotes the prediction of the same model on the evaluated input, which can be either adversarial or purified. Following the clean-reference evaluation principle adopted in 3D$^2$Fool [1], we measure the absolute difference (EI\mathcal{E}_\mathbf{I}, Ev\mathcal{E}_\mathbf{v}, Eb\mathcal{E}_\mathbf{b}) between these predictions to quantify attack-induced depth deviations and their recovery after purification. Therefore, the reported reduction from 26.53 m to 3.57 m quantifies the mean absolute depth-prediction discrepancy relative to the corresponding clean scene.

    • To further address the reviewer's concern, we additionally evaluate depth estimation errors against the ground-truth depth maps over all 1,043 scenes for each attack. Unlike our original clean-reference evaluation, this experiment directly assesses whether purification improves depth estimation accuracy relative to the actual scene depth. (All four attack types use the same set of 1,043 clean images, the “Clean” row is identical across all four columns.)

      Table 1: Ground-truth depth error of vehicle-region Ev(m)\mathcal{E}_\mathbf{v} (m) on Monodepth2

      Method3D2Fool3D-APA3D-SSAPRobCamou
      Clean1.5101.5101.5101.510
      No Defense4.00415.08314.90421.106
      Jedi3.1333.4383.8435.111
      PAD2.0383.3074.4647.545
      APEX3.7099.33310.64111.508
      CamouPurifier$\dagger$2.0122.1512.3844.193
      CamouPurifier$\ddagger$1.8582.0252.1183.517

      As shown in Table 1, CamouPurifier consistently reduces the vehicle-region ground-truth depth error across all four attacks. Specifically, our method achieves error reductions of 53.6% for 3D²Fool, 86.6% for 3D-APA, 85.8% for 3D-SSAP, and 83.3% for RobCamou. Notably, the purified results approach the clean-image baseline of 1.510. These results provide complementary evidence that CamouPurifier not only restores clean-reference depth predictions but also improves depth accuracy against ground truth.

  • For sum(I)\mathbb{sum}(\mathbf{I}): The sum(I)sum(\mathbf{I}) in Eq. (17) denotes the total number of image pixels H×WH\times W. Therefore, EI\mathcal{E}_\mathbf{I}, Ev\mathcal{E}_\mathbf{v}, and $ \mathcal{E}_\mathbf{b}$ all represent mean absolute depth-prediction discrepancies, computed over the entire image, vehicle region, and background region, respectively.

  • For Depth/disparity conversion & scale alignment: As described in Appendix A.1, the clean and adversarial images share identical viewpoints, vehicle geometry, and environmental conditions. The clean depth prediction serves as a fixed reference for evaluating both adversarial and purified outputs. The vehicle and background masks are defined from the corresponding clean scene and remain unchanged across the compared predictions. All evaluations are performed on spatially aligned depth maps. Since the evaluated MDE models may produce different depth representations, we apply model-specific depth/disparity conversion before computing discrepancies. For Monodepth2, Depth-Hints, ManyDepth, RobustDepth, and MIMDepth, we convert the predicted disparities into depth using the disp_to_depth() function provided by their respective official implementations, with the corresponding model-specific depth ranges and scaling configurations. For DepthAnything V2, we use the Outdoor Metric Depth checkpoint which directly predicts metric depth in meters with a maximum depth range of 80 m. Within each model, the same conversion and scaling procedure is used for clean, adversarial, and purified predictions, ensuring consistent within-model comparisons.

Ref: [1]. Junhao Zheng, Chenhao Lin, et. al. Physical 3d adversarial attacks against monocular depth estimation in autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pp. 24452–24461. IEEE, 2024.

AuthorsOpenReview ↗

Author Response to Reviewer bucb (Part 2/3)

2. Response to Q2 & W1 [Adaptive Attack Details]

Q2: What loss, weights, optimization budget, and gradient or query method were used for the adaptive attack? Did the attacker know q_d=20 and target the final defended depth output, including reconstruction when used? Please also report results with fixed defense settings over all 1,043 adaptive scenes.

We thank the reviewer for requesting further clarification. Since identifying abnormal areas is critical, similar to APEX, our adaptive evaluation considers a localization-aware white-box attack, where the attacker has full knowledge of SAL and MTAL, including their feature extraction, texture anomaly scoring, percentile thresholds, and mask fusion procedure.

Adaptive Attack Objective. Similar to APEX, we consider a localization-aware white-box attack in which the attacker has full knowledge of both SAL and MTAL, including their response functions, percentile thresholds, and mask fusion procedure. The adaptive attack jointly maximizes the vehicle-region depth discrepancy and minimizes the localization overlap between SAL and MTAL:

Ladaptive=−Ldepth+λlocLloc,L_{adaptive} = -L_{depth} + \lambda_{loc} L_{loc}, where λloc\lambda_{loc} controls the trade-off between attack effectiveness and localization evasion.

Depth Attack Loss. Let IadvI^{adv} and IclnI^{cln} denote geometrically aligned adversarial and clean images, respectively. Let MθM_\theta denote the victim MDE model and MvM_v denote the vehicle mask. The depth attack loss is defined as

Ldepth=∑Mv⊙∣Mθ(Iadv)−Mθ(Icln)∣∑Mv.L_{depth} = \frac{ \sum M_v \odot | M_\theta(I^{adv}) - M_\theta(I^{cln}) |}{\sum M_v}.

Maximizing this term encourages the adversarial camouflage to induce substantial depth prediction deviations within the vehicle region.

Localization Evasion Loss. Let Sd(p)S_d(p) and Sa(p)S_a(p) denote the DINO-based semantic response and the MTAL texture anomaly response, respectively. The attacker know the thresholds which are defined as td=Q20(Sd),ta=Q70(Sa)t_d=Q_{20}(S_d), t_a=Q_{70}(S_a). To approximate the non-differentiable binary thresholding operations, we construct differentiable soft masks:

M_a(p) = \sigma( \frac{S_a(p)-Q_{70}(S_a)}{\tau_a}),$$ where $\sigma(\cdot)$ is the sigmoid function and $\tau_d,\tau_a$ control the softness of the threshold approximations. Percentile-based masks depend on relative rankings, so uniformly reducing scores need not change their support. The surrogate therefore penalizes pixels jointly selected by the two branches. Attack success should be assessed using the actual binary masks after morphological and connected-component post-processing, not the surrogate alone. The localization evasion loss is formulated as $$\mathcal{L}_{\mathrm{loc}} = \frac{ \sum M_v(p) M_d(p) M_a(p) }{ \sum M_v(p)}.$$ Minimizing this loss encourages the two localization branches to exhibit less overlapping activation within the adversarial vehicle region. In the locate-adaptive attack setting, we adopt the Adam optimizer with a learning rate of 0.01. The model is trained for 100 epochs with a batch size of 8. $\tau_a=\tau_d=0.1$, $\lambda_{\mathrm{loc}}=5$. **Additional evaluation on all scenes.** To address the concern about the Top-500 sample selection, we evaluate both defense variants on all **1,043 scenes** of the localization-aware adaptive attack under the Monodepth2 setting. **Table 2: Evaluation of Adaptive Attack on monodepth2\$** | Defense | $\mathcal{E}_\mathbf{I}$ $\downarrow$ | $\mathcal{E}_\mathbf{v}$ $\downarrow$ | $\mathcal{E}_\mathbf{b}$ $\downarrow$ | |---|---:|---:|---:| | *No Defense* | *1.274* | *4.931* | *0.981* | | CamouPurifier\$^\dagger\$ (Masking) | 1.298 | 1.778 | 1.283 | | CamouPurifier\$^\ddagger\$ (Reconstruction) | 1.022 | 1.569 | 0.934 | The results show that CamouPurifier\$^\dagger\$ and CamouPurifier\$^\ddagger\$ reduce the vehicle-region depth discrepancy ($\mathcal{E}_\mathbf{v}$) by 63.9% and 68.2%, respectively, compared with the undefended baseline. Notably, the CamouPurifier\$^\ddagger\$ variant achieves a lower vehicle-region discrepancy than CamouPurifier\$^\dagger\$, while also providing better global and background consistency.
AuthorsOpenReview ↗

Author Response to Reviewer bucb (Part 3/3)

3. Response to Q3 & W2 [Contribution of Canny]

Q3: Is Canny computed from the adversarial crop or the black-masked crop, and which evaluation split is used in Table 3? To test Canny's contribution, could you compare results with and without it while holding masks, crops, prompts, seed, and diffusion settings fixed, and report vehicle/global/background discrepancies?

We sincerely thank the reviewer for this valuable suggestion. We agree that the original ablation study does not explicitly isolate the contribution of Canny-based geometry guidance. We clarify the implementation details and provide an additional controlled experiment below.

The Canny edge map is extracted from the original adversarial image crop before black masking, rather than from the black-masked crop. This allows the reconstruction module to utilize structural cues from the input image. Table 3 shows the average of 2,000 samples across four attack methods on monodepth2, whereas the additional Canny ablation is conducted on all 1,043 scenes against 4 attacks under monodepth2. To isolate the contribution of Canny guidance, we compare diffusion-based inpainting with and without Canny conditioning while keeping the localization masks, image crops, text prompts, random seeds, and diffusion settings identical.

Table 3: Ablation study of canny on the Monodepth2$

ConfigurationEI\mathcal{E}_\mathbf{I} ↓\downarrowEv\mathcal{E}_\mathbf{v} ↓\downarrowEb\mathcal{E}_\mathbf{b} ↓\downarrow
w/o Canny1.1121.3971.105
w/ Canny1.0191.3810.998

The Table 3 shows that Canny guidance reduces the background discrepancy by 9.59%, while also improving vehicle-region and full-image discrepancies by 1.15% and 0.27%, respectively. These results suggest that geometry guidance primarily benefits scene preservation in this evaluation, with a smaller but positive contribution to vehicle-region depth consistency.

We appreciate your thoughtful comments. We hope our response addresses your concerns. Please let us know if there are any additional questions, and we will be happy to discuss further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.