acceptodds
Under review as a conference paper at ICLR 2027

3DOSA: Unified Radiance Fields of 3D Objects and Scenes for Physically Realizable Adversarial Attacks in Autonomous Driving

Abstract

Achieving multi-view consistency in physical adversarial attacks requires an accurate 3D representation of the target object that captures intricate surface geometry. Recently, 3D Gaussian Splatting (3DGS) provides a promising representation for adversarial optimization due to its explicit and differentiable modeling of 3D objects and scenes. However, existing 3DGS-based attacks typically optimize physical adversarial camouflages without any physical environmental context or rely on 2D masks to identify optimizable Gaussians in the 3D scene, which can cause gradients to leak beyond the intended perturbation boundaries. In addition, volumetric Gaussians are difficult to align precisely with object surfaces, which can produce floating artifacts and geometric distortions that make the resulting camouflage physically undeployable. To address these limitations, we propose 3DOSA, a general physical attack framework that unifies the 3D object and the scene within a single adversarial radiance field. 3DOSA introduces an explicit spatial decoupling mechanism to isolate target Gaussians from the surrounding scene and prevent gradient leakage into background and non-printable regions. We further incorporate 2D Gaussian Splatting (2DGS) to represent adversarial Gaussians as planar surfels that closely adhere to the object surface. This representation suppresses floating artifacts and preserves the geometric fidelity of the object to maintain the physical realizability of the adversarial camouflage. We primarily evaluate 3DOSA on stereo matching, which imposes cross-view consistency beyond the multi-view requirements of general physical attacks, and further evaluate its versatility to object detection. Extensive experiments in simulated and physical environments demonstrate strong attack performance and robustness across diverse viewpoints and models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
physical adversarial attack, stereo matching, autonomous driving perception, 3D radiance field

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EQDMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes 3DOSA, a physical adversarial attack method to generate physically realizable adversarial camouflage against autonomous driving. The authors find two main limitations of existing solutions, namely gradient leakage and geometric misalignment. To address these, 3DOSA combines: 1) 3D Gaussian Grouping for explicitly separating optimizable object surfaces from the surrounding scene; 2) 2D Gaussian Splatting (2DGS) for surface-aligned adversarial texture optimization; and 3) Expectation over Transformation (EoT) for improving robustness across camera viewpoints and illumination conditions. The authors then apply the proposed the solution to both stereo matching models by manipulating the predicted disparity of a vehicle and object detection.

Critical strengths
  1. The knowledge gap is clear, e.g. the mismatch between digital optimization and real-world 3D geometry is indeed a concrete limitation of physical adversarial attack.

  2. The combination of explicit 3D target isolation and surface-constrained radiance optimization also make sense to me.

  3. The stereo matching evaluation is also a useful choice.

Critical weaknesses
  1. The methodological novelty seems limited. All the related components are from existing literature, making the contribution more like a integration of existing techniques.

  2. There is still gap between the experiments and the realistic traffic. It’s expected to perform experiments on more real world settings.

  3. More explanations are needed on why the same rendering framework can support both stereo matching and detection even when the attack is optimised in a task specific way.

  4. The whole system seems very large. e.g. the method uses many multi-view observations, segmentation supervision, and a differentiable 3D scene reconstruction. The practicality of obtaining these inputs for a real-world target deserves further investigation.

Questions for authors
  1. Can you explain intuitively on the chosen of those existing components?

  2. What will be the main challenges when apply the proposed solution in real world traffic?

  3. How to guarantee the practicality of having access to all the required inputs of the system in real world settings?

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

3DOSA puts a vehicle and the surrounding scene in one radiance field and builds the camouflage on the vehicle part of that field. Gaussian Grouping keeps the vehicle Gaussians separate from the scene. 2D Gaussian surfels are used so the pattern sits on the surface. Stereo matching is the main target. Object detection is a second target, run with a different objective that suppresses the car class.

Table 1, in CARLA, reports PSMNet, UniMatch, RAFT-Stereo, CREStereo, and FoundationStereo. 3DOSA records higher bad-pixel ratios and lower blending error than random texture, Cheng et al., Liu et al., and StereoAlign. Table 4, which does not name the model or viewpoint, lists bad-pixel ratios of 0.166 (2D texture), 0.272 (mask-based 3DGS), 0.520 (explicit decoupling), 0.359 (vanilla 2DGS), and 0.702 (full method). A 1:30 print is photographed as 100 stereo pairs with an iPhone 14 Pro Max at 50 cm and 100 cm, from the front, side, and back, on a rescaled KITTI baseline. A separate detection run drops average [email protected] from 88.73 to 21.73 on Faster R-CNN, YOLOv5, and DETR.

Critical strengths
  • Stereo matching is a harder physical target than single-image detection, and Table 1 reports five stereo models. The gaps are large. On PSMNet the bad-pixel ratio is 0.835 for 3DOSA, 0.553 for StereoAlign, 0.176 for Cheng et al., and 0.154 for Liu et al. FoundationStereo is the hardest column: 3DOSA reaches 0.676 there, against 0.253 for StereoAlign.
  • Table 4 is the right ablation. Decoupling is the large step, from 0.272 to 0.520. Vanilla 2DGS alone is 0.359, so the surfel representation is a smaller step on its own. The full method reaches 0.702. The full-method row matches the 5 m cell of Table 3 exactly (blending error 0.074, bad-pixel ratio 0.702). The paper never names the model behind either table.
  • Table 3 reports distance, heading, and weather. Among heading angles, 90° is the weakest (bad-pixel ratio 0.632). The paper attributes that to the flatter side of the vehicle.
  • Table 2 shows that the printed pattern still moves disparity on the phone rig. Adversarial bad-pixel ratios run from 0.467 to 0.574 across the three views and two distances.
Critical weaknesses
  • The novelty claim is ahead of the method. Gaussian Grouping and 2D Gaussian splatting are prior work. The related work already covers 3DGS camouflage, including PGA, 3DGAA, Meng et al., and Zhang et al. (2024), a bi-task detection-and-depth attack. The stereo objective is adapted from StereoAlign (Zhao et al., 2026). Explicit decoupling plus surface surfels in one scene field is the actual addition. Calling this the first fully cross-dimensional, task-agnostic physical attack framework for autonomous driving overreaches that addition.
  • "Task-agnostic" is two white-box objectives. Stereo is the main study. Detection changes the objective and is scored on Faster R-CNN, YOLOv5, and DETR. Table 5 drops average [email protected] from 88.73 to 21.73, ahead of PGA (35.68), 3DGAA (57.36), and Meng et al. (59.58). There is no BEV detector, no tracker, and no planner. The Town02 and Town03 appendix figures are described as the vehicle going undetected or being misclassified, and they report no stereo numbers.
  • The physical study does not support physical realizability on a vehicle. Section 4.1 uses a 1:30 print, an iPhone 14 Pro Max, 50 cm and 100 cm, a rescaled KITTI baseline, 100 pairs, and three viewpoints. The paper reports no motion, no automotive stereo pair, and no traffic scene. Table 2 is a real effect on that rig, and it is a much smaller result than the abstract. The back-view benign cell is printed as 0.24, well above the other benign cells (0.020–0.032).
  • The baseline comparison may credit the renderer. The paper does not say that Cheng et al., Liu et al., and StereoAlign were re-optimized in the same reconstructed scene, on the same printable region, with the same viewpoint randomization. In every Table 1 column those two older attacks still beat random texture. They are weak beside StereoAlign. On FoundationStereo they barely leave the floor: bad-pixel ratio 0.014 and 0.015, against 0.011 for both random texture and the benign render. Table 4 ablates only the authors' own variants.
  • Random texture already lowers blending error below the benign render on every model in Table 1. On PSMNet the numbers are 0.677 versus 0.752, while the bad-pixel ratio only moves from 0.013 to 0.052. That metric needs a definition before a lower blending error is read as attack success. In Table 3 the rainy benign bad-pixel ratio is already 0.250, so the rainy attack cell (0.550) is a different comparison from the clear-weather cells.
  • Section 4.1 names PSMNet, RAFT-Stereo, IGEV-Stereo, UniMatch, and FoundationStereo. Table 1 omits IGEV-Stereo and adds CREStereo, which the text never names. The UniMatch bibliography entry is Yang et al. (2023), a semi-supervised segmentation paper. The test-set text says three distances; Table 3 shows 5 m and 15 m and names no stereo model. The only back-matter disclosure is an AI-use statement limited to grammar and style. There is no ethics statement and no dual-use discussion.
Questions for authors
  1. Were Cheng et al., Liu et al., and StereoAlign re-optimized in the same Gaussian scene, with the same printable surface and the same viewpoint randomization? If each baseline kept its original renderer, which part of Table 1 is still an attack comparison?
  2. On the same 1:30 rig, what is the bad-pixel ratio of a non-adversarial printed pattern with the same coverage? Does any gap remain on a real automotive stereo pair, even for a parked vehicle?
  3. Is the Table 5 pattern the stereo camouflage, or a second camouflage trained with a detection objective? The section title says transfer. The method text says the objective is changed.
  4. Please define the ground-truth disparity in the bad-pixel ratio and the background reference in the blending error. Why does random texture beat the clean render on blending error for every model in Table 1? Which model and viewpoint produce Table 4, given that the full-method row matches the 5 m cell of Table 3?
1 comment
Reviewer_kZREOpenReview ↗

Official comments

After reviewing all reviewers' comment and authors' reply. I decided to keep my score.

I have read the paper and used AI to generate partial reviews. Frankly, I do not have sufficient expertise to assess this submission.

Reviewer_eV39OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes 3DOSA, a framework for generating physical adversarial camouflage through a joint representation of the target object and its surrounding scene. It combines Gaussian Grouping to identify the optimizable object primitives with 2D Gaussian Splatting to obtain a surface-aligned representation. During adversarial optimization, the background, geometry, and opacity remain fixed, while the target’s appearance parameters are optimized through jointly rendered stereo views.

The primary application is a stereo merging attack that pushes the target’s predicted disparity toward the surrounding background disparity. The framework is also applied to object detection using a task-specific objective. The evaluation includes CARLA experiments, component ablations, variations in viewpoints and environmental conditions, and a physical demonstration using a 1:30-scale vehicle model.

Critical strengths
  1. A useful formulation that connects the optimization domain with the physical target.

Explicitly separating the optimizable target primitives while retaining the surrounding scene is well motivated for physical stereo attacks. It offers a way to restrict appearance updates to the deployable object while preserving scene context during stereo image formation. The potential contribution lies in combining existing grouping and surface-representation techniques within this attack setting. This is a meaningful application of these techniques and could help make radiance-field-based attacks more practically usable.

  1. Component ablations and physical experiments provide initial supporting evidence.

Table 4 suggests that explicit grouping and the surface representation both contribute to the reported performance. For example, the reported E_bad increases from 0.272 for vanilla 3DGS to 0.520 with decoupling, while the full method reaches 0.702. The paper also evaluates a physical vehicle model using 100 stereo image pairs across different viewpoints and distances. These experiments provide useful evidence about the component combination and its initial physical feasibility, strengthening the case beyond simulation results alone.

Critical weaknesses
  1. The comparability of the reported baseline results requires clarification.

Table 1 explicitly attributes the baseline results to prior methods. Using published baseline results is appropriate when the evaluation protocols are comparable. However, the manuscript does not clearly distinguish quoted results from those reevaluated under the current setup.

The stated metric definitions also differ from the original StereoAlign paper: Zhao et al. report E_cover based on changes between benign and adversarial predictions, while this submission defines E_bad using disparity errors relative to ground truth. The stated E_blend thresholds are also different, at 20 and 3, respectively. It is unclear whether these differences concern the written descriptions, units or scaling, or the actual computations. Consequently, the manuscript does not yet establish that the reported gains in Table 1 represent directly comparable measurements. This clarification is important because the table provides central evidence for the claimed improvement over prior attacks.

Numerical agreement alone does not establish an error or misconduct. Nevertheless, the unexplained correspondence under different stated definitions prevents me from interpreting the reported improvements as a controlled comparison. This directly affects the principal evidence for superiority over prior stereo attacks.

  1. The physical texture generation and evaluation procedures are insufficiently specified.

The method optimizes spherical-harmonic appearance coefficients while keeping geometry and opacity fixed, but the conversion from this representation to the deployed texture is under-described. In particular, the paper should specify which SH components are optimized, how view-dependent appearance is handled, and how the resulting colors are extracted, baked, and mapped onto the physical vehicle.

The miniature-vehicle experiment provides valuable evidence of feasibility. Reproducing it, however, requires a clearer connection between the optimized radiance field and the fixed printed texture. The quantitative physical evaluation also needs a more explicit description of the reference disparity, stereo calibration, and evaluated model. These details directly affect the reproducibility and interpretation of the paper’s physical-realizability claim.

Questions for authors
  1. Baseline comparability

Which results in Table 1 are quoted from prior work, and which are reevaluated? Please explain how the baseline and 3DOSA results are made comparable given the different stated definitions of E_cover and E_bad and the different E_blend thresholds, including any differences in units or scaling. If the discrepancy is confined to the written descriptions, please identify the metric implementation actually used. If additional verification is necessary, a representative comparison under a shared evaluation implementation would be sufficient.

  1. Radiance field to printable texture.

Which SH components are optimized, and how are the optimized coefficients converted into a fixed printable surface texture? Please describe how view-dependent appearance is handled during extraction and how the texture is baked and mapped onto the vehicle. If available, evidence that the exported texture retains the attack effect would help establish the connection between the optimization and the deployed artifact.

  1. Physical measurements.

For Table 2, which stereo model is evaluated, how are the stereo images calibrated and rectified, and how are the reference and background disparities obtained? These details would help determine how the physical measurements correspond to the reported E_bad and E_blend.

These questions concern the existing claims and experiments. A new attack task or a full-scale vehicle experiment is not required to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.