acceptodds
Under review as a conference paper at ICLR 2027

SAR-DiSCO: Differentiable Scattering Coefficient Optimization for 3D Adversarial Attacks on SAR Target Recognition

Abstract

Synthetic aperture radar (SAR) automatic target recognition (ATR) systems are highly vulnerable to adversarial attacks, which can cause classification errors and threaten system reliability. Many existing attacks on SAR ATR inject perturbations directly into the 2D image domain without explicitly modeling target-surface scattering or the SAR image formation process. Such perturbations do not necessarily correspond to consistent target-level modifications across 3D viewing conditions. To address this gap, we propose SAR-DiSCO, a simulation-based adversarial attack framework that optimizes target-surface scattering coefficients through differentiable SAR rendering. The proposed approach employs a custom differentiable SAR renderer to perform end-to-end gradient optimization of vertex-wise scattering coefficients, producing adversarial perturbations while preserving target geometry. Specifically, we (1) build an end-to-end differentiable SAR imaging and recognition framework that links 3D surface coefficients to classification predictions, and (2) design a loss function that balances misclassification and scattering-coefficient deviation. We then (3) iteratively update the coefficients via backpropagation for view-specific attacks or shared multi-view optimization, without requiring a separately fitted surrogate mapping. Experiments on a simulated five-class vehicle dataset demonstrate an average view-specific fooling rate of 88.96% across five CNN- and Transformer-based classifiers. Further experiments demonstrate strong cross-architecture transferability and partial cross-view effectiveness when a single coefficient field is shared across viewing conditions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial example, Synthetic Aperture Radar (SAR), adversarial attack, deep neural network (DNN), SAR imaging simulation

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_ZefwOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes SAR-DiSCO, which optimizes target-surface scattering coefficients through a differentiable SAR renderer while keeping geometry fixed. Experiments on simulated vehicle data report an average view-specific fooling rate of 88.96% across five classifiers, alongside cross-model transfer and shared multi-view evaluations.

Critical strengths

Treating surface scattering coefficients as learnable parameters connects adversarial optimization to target-level changes rather than image pixels alone. This allows classification gradients to directly update the simulated surface response while preserving target geometry.

Critical weaknesses

The paper acknowledges that physical realizability remains unverified, but does not establish how the optimized coefficients could be implemented on real targets. Their sensitivity to imperfect material control is also unexplored. This limits the practical significance of the results beyond the adopted simulator.

Questions for authors
  1. The experiments remain simulation-based, without a demonstrated way to realize the optimized surface coefficients physically. Precisely controlling these coefficients on a real target would be difficult, yet attack effectiveness under coefficient fluctuations or implementation errors is not evaluated.

  2. The evaluation covers only five classification architectures. the limited model coverage leaves the generality of the attack insufficiently established.

  3. The External setting does not clearly specify whether the component’s coefficients are optimized independently for each view or shared across multiple views.

  4. In Table 2, Multi-view achieves a higher fooling rate than View-specific on Bus (54.17% versus 42.78%), unlike the other categories. why?

  5. The paper states that a common ground background is included, but Figures 2, 4, and 7 do not clearly show how background returns contribute to the classifier inputs. The evaluation does not establish attack effectiveness in more realistic scenes containing varied backgrounds and clutter.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Zefw

We thank the reviewer for recognizing the value of optimizing target-surface scattering coefficients and for the constructive feedback on physical implementation and evaluation. We have uploaded a revised manuscript with changes highlighted in red. We address the five questions below.

Q1. Physical realization and sensitivity to coefficient errors

Physical realization. SAR-DiSCO optimizes effective surface-scattering coefficients. Physically realizing them requires mapping the optimized values to material or component parameters and calibrating their scattering responses. This mapping and a physical demonstration have not yet been established. Our External experiment offers a potential direction for localized implementation, achieving 47.11% mean FR by optimizing only an attached component’s coefficients within the simulator.

Sensitivity to coefficient errors. We added an experiment that perturbs saved adversarial coefficients without re-optimization,

t~=clip⁡(tadv+ϵ,0,1),ϵv∼U(−δ,δ).\widetilde{\mathbf t}=\operatorname{clip}(\mathbf t_{\mathrm{adv}}+\boldsymbol\epsilon,0,1),\qquad \epsilon_v\sim\mathcal U(-\delta,\delta).

Errors are applied independently to the effective third coefficient channel. We re-render all 1,800 test samples per classifier while keeping the geometry, background, observation configurations, and classifier fixed. The resulting FR (%) is reported below and in Section 4.6 and Appendix I (Table 10).

Classifierδ=0\delta=0δ=0.01\delta=0.01δ=0.05\delta=0.05δ=0.08\delta=0.08δ=0.1\delta=0.1
ResNet-5077.2270.0641.7224.7217.61
VGG-1689.9481.5054.9438.7831.11
ViT-B/1694.2292.5076.8961.6149.83
Swin-V2-S93.2887.7863.3946.2237.28
EfficientNetV2-S90.1187.4458.6738.3327.78
Average88.9683.8659.1241.9332.72

Mean FR remains 83.86% at δ=0.01\delta=0.01, compared with 88.96% without errors, and decreases to 32.72% at δ=0.1\delta=0.1. Thus, the attacks retain substantial effectiveness under the smallest tested nonzero error bound, while larger deviations reduce their effectiveness. These results measure coefficient sensitivity within the simulator; interpreting δ\delta as a physical manufacturing tolerance requires material-specific calibration.

Q2. Architecture coverage and cross-model transfer

Our evaluation covers three CNNs (ResNet-50, VGG-16, and EfficientNetV2-S) and two Transformers (ViT-B/16 and Swin-V2-S). SAR-DiSCO achieves FRs of 77.22%–94.22%, averaging 88.96% (Table 1). Figure 3 further reports all 20 cross-model transfer combinations, with FRs of 36.44%–67.39% without accessing target-model parameters or gradients or re-optimizing the attacks. These results demonstrate effectiveness across the tested architectures and transferability beyond the model used for optimization.

Q3. Shared or view-specific coefficients in External

In the External setting, one component coefficient field is shared and jointly optimized across multiple views of each target. Only these shared coefficients are updated; the target body’s coefficients and the assembled geometry remain fixed. We have clarified this protocol in Section 4.4 and Appendix G.

Q4. Why Multi-view performs better on Bus

In Multi-view, an update from one view also affects the attack on other views. Some updates may benefit several views, while others may improve one view at the expense of another. For Bus, where view-specific optimization is less effective, joint updates may help find a more effective coefficient configuration. For the other four categories, sharing coefficients may instead limit adjustments tailored to individual views. This offers a possible explanation for the differences across categories, although the reported results do not establish the specific cause of the improvement on Bus.

Q5. Ground returns, varied backgrounds, and clutter

The SAR images used for classifier evaluation include ground returns. Both the ground and target participate in ray intersections, and their returns are accumulated on the same imaging grid to form the classifier inputs. Ground parameters remain fixed during attacks; only the target or attached component’s scattering coefficients are optimized. We have clarified this in Section 4.1. Using a common background provides consistent scene conditions for comparing methods. The current evaluation does not cover varied backgrounds and complex clutter, so effectiveness in those settings remains to be assessed separately.

Reviewer_km74OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAR-DiSCO studies adversarial vulnerability in synthetic aperture radar (SAR) target recognition by optimizing surface-scattering coefficients on fixed three-dimensional meshes. A differentiable renderer connects vertex-wise coefficients to simulated images and classifier outputs, avoiding a separate surface-fitting stage. The method uses a guided untargeted objective, bounded coefficients and a deviation penalty, and evaluates both independently optimized views and a coefficient field shared across views. On a synthetic dataset from five vehicle meshes, it reports a mean view-specific fooling rate of 88.96% across five recognition architectures (Table 1). On ResNet-50, the shared-field result is 56.94%, compared with 77.22% for view-specific optimization (Table 2). Cross-architecture transfer and an external-component setting provide additional evidence within the simulator.

Critical strengths
  1. A coherent formulation with an explicit, piecewise-differentiable path. Sections 3.1-3.2 and Appendix A explain how coefficient-dependent image formation connects surface parameters to recognition loss. Keeping ray intersections and grid assignments fixed is appropriate for coefficient-only optimization. As described in Appendix D, the direct formulation removes SAR-PATT's separate surface-mapping stage. Integrating a mesh-based coefficient field with differentiable SAR simulation provides a useful framework for studying recognition sensitivity under a controlled forward model.

  2. The evaluation distinguishes several experimental settings. Table 1 covers five architectures. Its fooling rate measures error over all evaluated samples, including those already misclassified when clean; the low clean error rates nevertheless make the degradation clear. Figure 3 reports cross-architecture transfer, while Table 2 separates view-specific, shared-field and external-component results. These comparisons distinguish effectiveness across the reported views from success on separately optimized images. The sensitivity studies and rendering appendix provide useful detail, and Section 5 explicitly limits the physical interpretation of the simulation results.

Critical weaknesses
  1. The comparisons do not isolate the benefit of direct coefficient optimization. Section 4.2 acknowledges that equal iteration counts do not match perturbation budgets across BIM, Patch, SMGAA and SAR-DiSCO. Without quantitative measures of the resulting perturbations under comparable constraints, the fooling rates cannot establish superiority at a matched perturbation budget. Appendix D and Table 7 appropriately identify the SAR-PATT experiment as a comparison of complete pipelines, with separately trained classifiers and different imaging procedures. This is informative, but those differences prevent attributing the performance gap specifically to direct optimization. Tables 3-4 examine iteration count and regularization strength, rather than isolating the contribution of the guided objective. The evidence therefore supports the effectiveness of the complete system more strongly than its individual methodological choices.

  2. Evaluation breadth and the effective parameterization need clarification. Section 4.1 and Appendix C describe five meshes, one per category, limiting the range of target instances tested. Shared-field optimization traverses all five sensor configurations and the azimuth grid, but a disjoint set of evaluation views is not clearly identified. Table 2 therefore supports effectiveness across the reported views without establishing generalization to views excluded from optimization. Separately, Section 3.1 defines three coefficients per vertex, while Appendix A's scalar response uses the third component. The roles of the other channels, the effective degrees of freedom and the achieved coefficient changes need clarification. The piecewise rendering description would also benefit from numerical validation of its implemented derivatives. These concerns address the simulation study itself; physical realization remains a separate question, as the authors acknowledge.

Questions for authors
  1. What coefficient-space or image-space perturbation magnitudes accompany the results, and which constraints permit a meaningful comparison across methods? Please clarify which conclusions are supported by the complete-pipeline comparison in Table 7 and which would require additional controls.

  2. Does Table 2 evaluate the same azimuths and sensor configurations used to optimize the shared field? Please distinguish the classifier's training/test split from any split between views used to optimize that field and views used to evaluate it.

  3. What does each of the three coefficient channels represent, and how does it affect the rendered scalar response? Please reconcile the parameterization in Section 3.1 with Appendix A and report the magnitude and spatial extent of the resulting coefficient changes.

  4. Have the implemented coefficient derivatives been numerically checked against the forward calculation away from changes in thresholded PSF support? Appendix A explains the piecewise differentiation, but a reported consistency check would strengthen confidence in its implementation.

4 comments
AuthorsOpenReview ↗

Response to Reviewer km74 (Part 1/2)

We thank the reviewer for the detailed comments on comparison conditions, coefficient parameterization, and derivative verification. We have uploaded a revised manuscript with changes highlighted in blue. Below we answer Q1–Q4, then address the guided-objective contribution and target-instance coverage raised in the critical weaknesses.

Q1. Perturbation magnitudes, comparison constraints, and Table 7 conclusions

Perturbation settings and comparison conditions. Section 4.2 now specifies each method’s constraints. Random independently samples vertex-wise coefficients uniformly from [0,1)[0,1), whereas SAR-DiSCO optimizes them within (0,1)(0,1) with a squared-deviation penalty of weight c=0.25c=0.25. Both bound each coefficient’s absolute deviation from its clean value of 0.50.5 by 0.50.5. For images normalized to [0,1][0,1], BIM uses an L∞L_\infty budget of 8/2558/255, and Patch uses a circular patch of 30-pixel diameter. SMGAA uses three scattering centers with amplitude parameter A∈[0,10]A\in[0,10], followed by clipping to [0,1][0,1]. This amplitude parameter controls the scattering-model response rather than directly bounding pixel-wise changes.

Within each classifier block, all methods use the same 1,800 test samples, observation configurations, trained classifier, and untargeted FR metric. BIM, Patch, SMGAA, and SAR-DiSCO each use 50 iterations. These conditions provide a common task-level evaluation under explicit method-specific constraints. Random additionally provides a reference within the same surface-coefficient bounds: its mean FR is 9.54%, compared with 88.96% for SAR-DiSCO. Appendix I reports the achieved coefficient changes, detailed under Q3. These comparisons establish attack effectiveness under the stated constraints, rather than superiority at a matched cross-method perturbation budget.

Conclusions from original Table 7 (Table 8 in the revised manuscript). Appendix D evaluates both methods after rendering from their final surface coefficients, including the fitting stage for SAR-PATT. The methods share target meshes, training/test views, and classifier training settings, with a separate ResNet-50 trained for each imaging pipeline. Both gradient-optimization stages use 50 Adam updates at learning rate 0.05; SAR-PATT subsequently performs surface fitting. Its constraints include sigmoid-bounded intermediate intensities, a squared-deviation penalty of weight 1, a sparsity threshold of 10−410^{-4}, and fitting over [0,1][0,1].

SAR-DiSCO improves FR by 13.06, 28.89, and 10.56 percentage points on Car, Truck, and Van, respectively, performs similarly on Jeep, and achieves lower FR on Bus. These results support its effectiveness as a complete surface-attack pipeline. Producing surface coefficients directly without a separate fitting stage is a property of the formulation.

Isolating the performance contribution of direct optimization would require holding the renderer, classifier, final surface parameterization, and perturbation constraints fixed while varying the optimization path. Table 8 therefore supports conclusions at the complete-pipeline level.

Q2. Classifier split versus shared-field optimization and evaluation views

Yes. The evaluation views in Table 2 are included among the views used to optimize the shared field. Sections 4.1 and 4.4 now distinguish the two protocols explicitly.

Classifier split. The 9,000 clean images are split 80:20 for classifier training and testing. Each classifier is trained on the training subset and frozen during attack optimization. Evaluation uses the 1,800 classifier-test samples, with 360 per category.

Shared-field views. For each target, one coefficient field is optimized across all five sensor configurations and all 360 azimuths at 1∘1^\circ intervals. This includes the classifier-test views; there is no separate held-out-view split for field optimization.

Accordingly, the 56.94% mean FR measures the effectiveness of one shared field across optimized views, compared with 77.22% for independently optimized view-specific fields. It does not measure generalization to unseen views.

AuthorsOpenReview ↗

Response to Reviewer km74 (Part 2/2)

Q3. Coefficient channels, scalar response, and perturbation magnitude and extent

Effective parameterization. Section 3.1 clarifies that the three channels are storage components, not distinct scattering mechanisms or polarizations. Only the third channel contributes to the scalar SAR response, giving t=T:,3\mathbf t=T_{:,3}. Its values are barycentrically interpolated at intersections and modulate return amplitude (Appendix A). The first two channels remain at their clean initialization during SAR-DiSCO optimization. There is one effective scattering coefficient per represented vertex.

Magnitude. Appendix I defines the mean absolute change and the mean per-sample maximum absolute change relative to the clean value 0.50.5. Statistics cover all 1,800 test samples per classifier, averaging within each category and then equally across the five categories. Figure 10 reports:

ClassifierSAR-DiSCO mean absolute changeSAR-DiSCO mean per-sample maximumRandom mean absolute change
ResNet-500.15%15.73%24.97%
VGG-160.17%17.38%25.02%
ViT-B/160.14%16.25%24.98%
Swin-V2-S0.18%18.53%25.01%
EfficientNetV2-S0.15%16.87%24.99%

Percentages represent absolute coefficient changes multiplied by 100, without division by the clean value of 0.50.5.

Spatial extent. Figure 6 marks nonzero third-channel changes in red, spanning broad surface regions with target-dependent distributions. Red indicates locations, not magnitudes; geometry remains fixed.

Q4. Numerical verification away from PSF-support changes

Yes. Appendix H, Table 9 compares implemented coefficient-to-image derivatives with central finite differences. Each step size uses 30 attempted directional checks per category, totaling 150 across five categories.

For a fixed scalar projection f(t)=⟨w,vec⁡I(t)⟩f(\mathbf t)=\langle\mathbf w,\operatorname{vec}I(\mathbf t)\rangle, we compare

\qquad g_{\mathrm{FD}}(h)=\frac{f(\mathbf t+h\mathbf d)-f(\mathbf t-h\mathbf d)}{2h}.$$ Geometry, imaging configurations, projection weights, perturbation directions, and sampling randomness are fixed within each comparison. Both perturbed forward evaluations recompute the PSF support. Valid checks require feasible coefficients, unchanged support relative to the unperturbed evaluation, and stable relevant piecewise branches. The check uses float32 coefficients and complex64 echo buffers. Relative error is $$e_{\mathrm{rel}}(h)= \frac{|g_{\mathrm{AD}}-g_{\mathrm{FD}}(h)|} {\max(|g_{\mathrm{AD}}|,|g_{\mathrm{FD}}(h)|,10^{-8})}.$$ | Step size $h$ | Valid / attempted | Median relative error | P95 relative error | Maximum relative error | |:---:|:---:|:---:|:---:|:---:| | $10^{-2}$ | 109 / 150 | 1.84% | 6.27% | 14.36% | | $10^{-3}$ | 137 / 150 | 0.32% | 1.46% | 4.82% | | $10^{-4}$ | 146 / 150 | 0.21% | 0.78% | 1.15% | At $h=10^{-4}$, **146 of 150 checks are valid, with a median relative error of 0.21%**. These results support numerical consistency at the tested configurations. ### Additional points raised in the critical weaknesses **W1. Contribution of the guided objective.** We compare standard untargeted cross-entropy maximization against our guided untargeted objective across all five classifiers. The baseline maximizes CE for the ground-truth label; ours minimizes CE for the highest-scoring incorrect class, reselected each iteration. Both use the same untargeted success criterion, 1,800 test samples per classifier, renderer, observation configurations, frozen classifier, coefficient initialization and parameterization, regularization weight ($c=0.25$), and Adam settings with 50 updates. **Only the classification objective changes.** Section 4.5, Table 5 gives category-level results; means follow. | Classifier | Standard untargeted CE FR (%) | Ours FR (%) | |---|:---:|:---:| | ResNet-50 | 67.67 | 77.22 | | VGG-16 | 84.61 | 89.94 | | ViT-B/16 | 87.11 | 94.22 | | Swin-V2-S | 84.39 | 93.28 | | EfficientNetV2-S | 83.83 | 90.11 | | **Average** | **81.52** | **88.96** | Our objective improves mean FR from **81.52% to 88.96%**, with gains in **all 25 classifier–category combinations**. The mean improvement is 7.43 percentage points, calculated from unrounded rates. This controlled comparison supports the contribution of the guided objective, complementing the iteration-count and regularization analyses in Tables 3–4. **W2. Target-instance coverage.** Section 4.1 and Appendix C specify **one mesh per category**. The evaluation covers five target instances across the stated viewing conditions. The classifier-test split separates rendered observations of these instances; generalization to unseen vehicle instances has not been evaluated.
Reviewer_km74OpenReview ↗

Official Comment by Reviewer km74

Thank you for the detailed responses and additional experiments. The revision addresses several of my concerns and strengthens the evidence for the proposed framework. However, the specific benefit of direct coefficient optimization remains insufficiently isolated, and generalization to held-out views and unseen targets remains unverified. I therefore maintain my original rating of 2/4.

AuthorsOpenReview ↗

Response to Reviewer km74

Thank you for acknowledging the improvements and clarifying your remaining concerns. We have added an optimization-path comparison, transfer to a held-out 20° elevation, and an evaluation on Pickup/SUV to examine direct optimization, unseen-view transfer, and target coverage. These experiments are included in the revised manuscript, with changes highlighted in blue.

1. Isolating the benefit of direct coefficient optimization

The original SAR-PATT comparison involves different renderers and classifiers. We therefore added a two-stage SAR-DiSCO variant within the same imaging and classification pipeline (Appendix D, Table 9).

Direct SAR-DiSCO optimizes vertex coefficients from their clean initialization. The two-stage variant starts from the corresponding clean intermediate scattering responses, optimizes them, and then fits vertex coefficients. Both variants share the renderer, frozen ResNet-50, meshes, observation configurations, classification objective, and final coefficient parameterization and bounds. Intermediate-response optimization uses 50 Adam updates, learning rate 0.05, and regularization weight 0.25, followed by surface fitting. We evaluate the two-stage variant after re-rendering its fitted coefficients, not at the intermediate-response stage. Both variants use the same 360 test images per category, totaling 1,800.

Table 9 retains the original pipeline comparison and adds the controlled variant. Values are fooling rates (FR, %).

MethodMetricBusCarJeepTruckVanMean
SAR-PATTClean FR1.943.063.892.222.782.78
SAR-PATTAttacked FR51.3970.2859.7271.1189.4468.39
SAR-DiSCOClean FR1.111.391.671.111.391.33
SAR-DiSCOAttacked FR42.7883.3360.00100.00100.0077.22
SAR-DiSCO (two-stage)Attacked FR40.5681.6757.78100.0099.1775.83

Direct optimization improves mean FR from 75.83% to 77.22%, a gain of 1.39 percentage points, with higher FR on four categories and a tie on Truck. This comparison supports a modest performance benefit under the tested settings, while direct optimization also avoids a separate fitting stage. SAR-PATT remains a complete-pipeline reference; the two SAR-DiSCO variants provide the within-pipeline comparison.

2. Transfer to an elevation excluded from attack optimization

We re-rendered saved view-specific attacks from the five source elevations (15°, 17°, 25°, 35°, and 45°) at 20°, retaining each sample's azimuth, mesh, coefficients, and the original ResNet-50. No attack optimization or classifier retraining was performed at 20°. Each attack has a clean control at the same test view. We pool 360 paired evaluations per category across the five source elevations, retaining repeated clean views when source samples share an azimuth. Section 4.3, Table 2, reports:

TargetPaired evaluations20° clean accuracy (%)20° transfer FR (%)
Bus36091.6735.00
Car36090.2866.39
Jeep36094.4443.61
Truck36091.1161.94
Van36093.3364.17
Overall1,80092.1754.22

The saved attacks achieve 54.22% transfer FR at the held-out elevation, compared with a clean error rate of 7.83%. All five categories show higher error rates under attack. This tests transfer of independently optimized fields to an unseen elevation, complementing the shared-field evaluation on optimized views in Section 4.4, Table 3.

3. Evaluation on additional targets

We added Pickup and SUV in a separate two-class task, training a separate ResNet-50 with the same procedure and 80:20 observation split as in Section 4.1. We optimized attack coefficients for these targets using the original attack settings. Appendix J, Table 13, reports results on 360 test images per category:

TargetTest imagesClean accuracy (%)FR (%)
Pickup36098.0686.67
SUV36092.2282.78
Overall72095.1484.72

SAR-DiSCO achieves 84.72% overall FR, extending the evaluation beyond the original five targets. These results demonstrate that SAR-DiSCO can be applied to target classification tasks beyond the original five categories.

Thank you for identifying these points. The optimization-path comparison and held-out-elevation experiment provide direct new evidence, while the additional-target experiment broadens the evaluation. We hope these results provide a stronger basis for reassessing our work.

Reviewer_9BJoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces SAR-DiSCO, a simulation-based physical-domain attack framework for SAR target recognition. The framework directly optimizes 3D surface scattering properties through a differentiable SAR renderer to generate adversarial SAR images instead of introducing perturbations in the image or signal domain. The generated SAR images are then passed through a pretrained classifier where the classification loss is used to optimize the surface scattering coefficients and to promote misclassification. The authors evaluate the method using 5 synthetic vehicle classes and 5 CNN/Transformer classifiers. They report an average fooling rate of 88.96%, demonstrate cross-architecture transferability, and investigate whether the same surface modifications can fool classifiers across different viewing angles.

Critical strengths
  1. The paper presents an end-to-end differentiable framework that directly optimizes 3D surface scattering properties to generate adversarial SAR images, providing an interesting alternative to conventional image-domain or signal-domain attacks.

  2. The experimental results demonstrate that the proposed framework achieves high fooling rates across five architectures (CNN and Transformer based classifiers) and generally outperforms the evaluated baselines. It demonstrates the effectiveness within the simulated SAR environment.

Critical weaknesses
  1. As discussed by the authors, all experiments remain simulation-based. Therefore, the real-world effectiveness of the perturbations remains unverified. Furthermore, there is limited evaluation to support generalization to unseen targets as the training and testing datasets use different rendered views of the same five 3D CAD vehicle models.

  2. The scattering model appears to be oversimplified. The renderer does not fully account for complex electromagnetic scattering effects or realistic noise. The experiments do not evaluate the sensitivity of the proposed adversarial attack to the modeling assumptions.

Questions for authors
  1. Since all experiments are simulation-based, how can the optimized scattering coefficients be implemented on real targets?

  2. The dataset uses only five 3D CAD vehicle models. How well does the proposed approach generalize to unseen vehicle models?

  3. How sensitive is the proposed adversarial attack to changes in the SAR simulation settings, such as noise or different scattering assumptions?

  4. The multi-view attack achieves a much lower fooling rate (56.94%) compared to the view-specific attack (77.22%). Could you explain the main reasons for this performance drop? Are there any limitations in the proposed optimization approach when handling multiple viewing angles?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9BJo

We thank the reviewer for recognizing the value of direct surface-coefficient optimization and its effectiveness in the simulated SAR environment. We have uploaded a revised manuscript with changes highlighted in red. Below we address the four questions, including the concerns about physical realization, target-instance coverage, and simulation fidelity raised in the critical weaknesses.

Q1. Physical realization of the optimized coefficients

SAR-DiSCO optimizes effective surface-scattering coefficients. Physically realizing them requires mapping the optimized values to material or component parameters and calibrating their responses under the relevant radar frequencies and viewing conditions. This mapping and a physical demonstration have not yet been established. Our External experiment offers a potential direction for localized implementation, achieving 47.11% mean FR by optimizing only an attached component’s coefficients within the simulator. This shows that modifying the entire target surface is unnecessary in the tested setting, while realizing the component physically still requires electromagnetic modeling and measurement.

Q2. Generalization to unseen vehicle instances

The current evaluation uses one CAD model per category, as specified in Section 4.1 and Appendix C. The 80:20 classifier training/test split separates rendered observations of these five models, not vehicle instances. It therefore does not measure generalization to unseen vehicles. SAR-DiSCO optimizes a coefficient field for a given target mesh; applying the procedure to a new mesh would involve optimizing a field on that mesh, which is distinct from transferring an existing field to an unseen vehicle. The reported results establish effectiveness on the five tested instances across the stated viewing conditions.

Q3. Sensitivity to noise and scattering assumptions

Section 4.6 and Appendix I (Table 10) report a new sensitivity experiment that perturbs the saved adversarial coefficients while keeping the imaging model fixed. We add independent uniform errors to the effective third channel,

\qquad \epsilon_v\sim\mathcal U(-\delta,\delta),$$ and re-render all 1,800 test samples per classifier without re-optimizing the attack. The classifier, geometry, background, and observation configurations remain fixed. The resulting FR (%) is shown below. | Classifier | $\delta=0$ | $\delta=0.01$ | $\delta=0.05$ | $\delta=0.08$ | $\delta=0.1$ | |---|---:|---:|---:|---:|---:| | ResNet-50 | 77.22 | 70.06 | 41.72 | 24.72 | 17.61 | | VGG-16 | 89.94 | 81.50 | 54.94 | 38.78 | 31.11 | | ViT-B/16 | 94.22 | 92.50 | 76.89 | 61.61 | 49.83 | | Swin-V2-S | 93.28 | 87.78 | 63.39 | 46.22 | 37.28 | | EfficientNetV2-S | 90.11 | 87.44 | 58.67 | 38.33 | 27.78 | | **Average** | **88.96** | **83.86** | **59.12** | **41.93** | **32.72** | Mean FR remains 83.86% at $\delta=0.01$, compared with 88.96% without coefficient errors, and decreases to 32.72% at $\delta=0.1$. Thus, the optimized attacks retain substantial effectiveness under the smallest tested nonzero coefficient-error bound, while larger deviations substantially reduce their effectiveness. This experiment directly evaluates sensitivity to inaccuracies in the optimized surface response under the adopted renderer; robustness to image or echo noise and alternative scattering models remains an open question. ### Q4. Lower mean FR under multi-view optimization **View-specific optimization can tailor a separate coefficient field to each view, whereas Multi-view must use one field across all views.** Different views observe different surface regions and scattering responses, so a change that helps one view may be less effective for another. Sharing coefficients therefore requires compromises across views. The optimization procedure adds another limitation. As described in Section 3.3, views update the shared field sequentially, so later updates may weaken the attack on earlier views. The objective aims to minimize the average loss across views and does not guarantee misclassification at every view. These features help explain the lower mean FR, although the experiments do not isolate their individual contributions. The 56.94% result nevertheless shows that one shared field can remain effective across multiple optimized views. As clarified in Section 4.4, this is evaluated on views included in field optimization, rather than on held-out views.

I went through the rebuttal, and the authors have provided useful clarifications and an additional sensitivity experiment. However, some of the main limitations regarding physical realizability, generalization to unseen targets, and the simplified scattering model still remain.

I think I will maintain my original recommendation, and I don't have any further questions for the authors at this point.

AuthorsOpenReview ↗

Response to Reviewer 9BJo

Thank you for reviewing our response and acknowledging the clarifications and sensitivity experiment. Regarding your remaining concerns about physical realizability, generalization to unseen targets, and the simplified scattering model, we clarify the material implementation pathway and report additional evaluations on new targets, an unseen elevation, and scenes composited with measured clutter. The new experiments are included in the revised manuscript, with changes highlighted in blue.

1. Physical realizability: from optimized responses to surface materials

Our proposed implementation pathway is to apply absorbing materials, reflection-enhancing materials, or surface structures to different regions according to the optimized spatial response distribution. Absorbing treatments reduce local echo contributions, while reflection-enhancing treatments increase them under specified observation conditions.

Specifically, candidate materials or structures would be characterized at the corresponding radar frequency and incidence conditions. Their scattering responses would then be fitted to the renderer's effective coefficients to select surface configurations approaching the optimized responses. SAR-DiSCO specifies the desired spatial response distribution; material-response calibration guides the selection of surface treatments. This pathway also applies to External attacks, where treatments can be confined to the attached component. The connection is established through scattering responses, rather than treating coefficient values as material parameters.

2. Target coverage and view transfer: two additional targets and an unseen elevation

To extend beyond the original five CAD models, we added Pickup and SUV, trained ResNet-50 on a separate two-class task, and optimized attack coefficients for these targets. We retained the original training procedure, 80:20 observation split, and attack settings. Appendix J, Table 13, reports results on 360 test images per category, totaling 720.

TargetTest imagesClean accuracy (%)FR (%)
Pickup36098.0686.67
SUV36092.2282.78
Overall72095.1484.72

These results provide direct evidence of applicability beyond the original five targets. They evaluate reapplying attack optimization to new targets with a separately trained classifier, which differs from unseen-instance transfer under a fixed classifier.

We also tested cross-view transfer without re-optimization. Saved view-specific attacks from five source elevations (15°, 17°, 25°, 35°, and 45°) were re-rendered at 20°, which was excluded from attack optimization. Azimuths, meshes, coefficients, and the original ResNet-50 remained unchanged. Across 360 paired evaluations per category and 1,800 overall, mean clean accuracy was 92.17% and transfer FR was 54.22% (Section 4.3, Table 2). Thus, the saved attacks retain effectiveness at an elevation not used during optimization.

3. Scene simplification: evaluation with measured MSTAR clutter

To assess performance beyond the original simulated backgrounds, we used the MSTAR dataset's independent Public Clutter collection to evaluate saved attacks in scenes composited with measured backgrounds. These backgrounds introduce measured SAR textures, grayscale variations, and non-target scattering responses.

For each clean–adversarial pair, we randomly cropped a 224×224224\times224 background patch. Both images used the same patch and processing parameters, with the original classifier checkpoints and no attack re-optimization. We retained foreground pixels above grayscale 80, scaled background values by 0.20 and rounded them, added the images pixelwise, and clipped to [0,255][0,255]. Each classifier was evaluated on all 1,800 test images.

Appendix K, Table 14, reports the results; Figure 11 shows paired examples before and after compositing.

ClassifierComposited clean accuracy (%)Composited FR (%)
ResNet-5074.3355.89
VGG-1677.6163.22
ViT-B/1680.2867.94
Swin-V2-S78.9465.39
EfficientNetV2-S76.0660.72
Mean77.4462.63

The saved attacks achieve 62.63% mean FR in these measured-clutter composites. Under identical processing conditions, the mean clean error rate is 22.56%, 40.07 percentage points below the adversarial error rate. All five classifiers show higher error rates on adversarial images, so the errors do not arise solely from background changes. Mean FR is 26.33 percentage points below the original-scene result, reflecting the effect of foreground extraction and background compositing.

Thank you for identifying these aspects requiring further evidence. We hope the implementation pathway and additional experiments provide a more concrete basis for reassessing our work.

Comments on the paper

7 comments
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We thank the reviewers and Area Chair for their time and constructive feedback. We have uploaded a revised manuscript with changes highlighted in red. The revision adds three experimental analyses and clarifies the parameterization and evaluation protocols, as summarized below.

1. Contribution of the classification objective (km74, Critical Weakness 1). We compare our guided untargeted objective with standard untargeted cross-entropy maximization while keeping the renderer, classifier, initialization, regularization, and optimization budget fixed. Our objective increases mean FR from 81.52% to 88.96%, with improvements across all 25 classifier–category combinations. This comparison provides evidence for the objective’s contribution under the tested settings (Section 4.5, Table 5).

2. Numerical consistency of coefficient derivatives (km74, Q4). We check backpropagated directional derivatives against central finite differences, retaining comparisons with feasible coefficients, unchanged thresholded PSF support, and stable piecewise branches. At h=10−4h=10^{-4}, 146 of 150 checks are valid, with median, P95, and maximum relative errors of 0.21%, 0.78%, and 1.15%, respectively (Appendix H, Table 9).

3. Sensitivity to coefficient inaccuracies (Zefw, Q1; 9BJo, Q3). We perturb saved adversarial coefficients with bounded uniform errors and re-render all 1,800 test samples per classifier without re-optimization. Mean FR remains 83.86% at an absolute error bound of 0.01, compared with 88.96% without errors, and decreases to 32.72% at a bound of 0.1. These results quantify sensitivity to coefficient inaccuracies within the adopted renderer. Our responses distinguish this experiment from physical implementation and robustness to image/echo noise or alternative scattering models (Section 4.6; Appendix I, Table 10).

4. Coefficient parameterization and comparison conditions (km74, Q1 and Q3). We clarify how the effective third coefficient channel determines the scalar SAR response, report the magnitude and spatial locations of coefficient changes, and specify each baseline’s perturbation constraints. We also clarify that the SAR-PATT comparison evaluates complete attack pipelines after rendering from their final surface coefficients (Section 4.2; Appendices A, D, E, and I; revised Table 8).

5. Evaluation protocols and interpretation (9BJo, km74, and Zefw). We distinguish the classifier training/test split from the views used to optimize the shared field, clarify that External uses coefficients shared across multiple views, and explain how ground returns contribute to classifier inputs. The reviewer-specific responses also address the multi-view performance differences, including the Bus exception, and distinguish the evidence for cross-architecture transfer from generalization to unseen vehicle instances.

These additions provide further evidence for SAR-DiSCO’s effectiveness within the simulated SAR setting and make the experimental conditions and supported conclusions clearer. Detailed results and explanations are provided in our responses under each review. We appreciate the reviewers’ consideration of the revised manuscript and welcome further discussion.

Dear Reviewers,

The authors have provided their rebuttals to the reviewers' comments. Please take a moment to review the responses and continue the discussion on any remaining concerns, unresolved issues, or points that may require further clarification.

Your further input and discussion are highly appreciated.

Regards, AC for this paper

Thank you for the response. The new experiment addresses my concern about sensitivity to coefficient errors, and I appreciate the clarification of the External setting. However, the response does not explain how the optimized coefficients could be realized on a physical target, and it cites existing results rather than testing additional classifiers. Even one or two additional classifiers would have helped assess how well the method generalizes.

For the Bus result, the authors offer a possible explanation but note that “the reported results do not establish the specific cause of the improvement on Bus.” Similarly, the response confirms that the inputs contain ground returns, but concludes that “effectiveness in those settings remains to be assessed separately.” These are fair qualifications, but they leave my original concerns about the multi-view result and performance under realistic backgrounds and clutter unresolved.

I would have appreciated a more direct response to the intent of these questions. If AI tools are used to help draft responses, I would encourage using them sparingly and keeping the focus on concrete explanations and evidence. Since most of my concerns remain open, I am revising my score to 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.