- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes GAUSS, a lightweight image perturbation framework for selective soft-biometric privacy protection. The method constructs low-dimensional Spectral Attribute Subspaces (SAS) through spectral decomposition of the input gradients of frozen attribute classifiers and generates attribute-specific protective perturbations within these subspaces. Building on this, GAUSS introduces an Instance-Adaptive Perturbation (IAP) module that predicts image-dependent subspace coefficients and spatial masks to selectively suppress different sensitive attributes. In addition, the authors employ joint supervision from multiple attribute classifiers and semantic directions derived from CLIP text embeddings to improve the transferability of the perturbations to unseen model architectures.
Critical strengths
The paper investigates the role of low-dimensional attribute-sensitive subspaces in selective perturbation generation by examining the spectral structure of attribute classifiers' input gradients. The experiments demonstrate that low-dimensional learnable subspaces can achieve high attribute prediction flip rates and that the initial spectral subspaces estimated from different data subsets exhibit a certain degree of stability. These findings provide a valuable representation-geometric perspective on selective adversarial perturbations.
Critical weaknesses
-
The paper primarily uses Flip Rate to evaluate attribute privacy protection, measuring the proportion of protected images for which the classifier's prediction differs from the ground-truth attribute label. However, for binary sensitive attributes, a high prediction flip rate does not necessarily imply that the original attribute information has been effectively concealed. For example, if the protection mechanism consistently causes a binary attribute classifier to predict the opposite of the true label for the vast majority of samples, an attacker aware of this behavior could potentially recover the original attributes simply by inverting the classifier's predictions. This requires no classifier retraining. The paper reports a white-box prediction flip rate of 99.56% for the Young attribute, but this result alone does not establish a corresponding level of privacy protection. On the contrary, for the same binary predictions, a simple label-inversion strategy could potentially achieve a comparable attribute recovery rate. Therefore, the current experiments more directly demonstrate that attribute classifier predictions are disrupted, rather than that sensitive attribute information has been effectively removed from the protected images.
-
GAUSS is fundamentally an adversarial perturbation generation method with a bounded perturbation magnitude, yet the paper primarily compares it against generative facial attribute editing and privacy-preserving methods. Although these approaches share similar application objectives, they differ substantially in how they modify images, their optimization objectives, generative priors, and computational complexity. Consequently, some of GAUSS's advantages in terms of PSNR, identity preservation, model size, and inference latency may arise from the inherent differences between localized pixel-level perturbations and generative image editing, rather than solely from the proposed spectral subspace design. Moreover, although the Related Work section discusses attribute-based adversarial perturbations, mutual information-based attacks, class-wise universal perturbations, and other structured perturbation methods, the main experiments lack systematic comparisons with these more closely related approaches.
Questions for authors
-
The paper primarily uses the Flip Rate of binary attribute classifiers as a privacy protection metric. However, when classifier predictions are consistently flipped, an attacker aware of the protection mechanism may recover the original attributes simply by reversing the predictions. Could the authors explain how GAUSS ensures that sensitive attribute information is effectively concealed, rather than merely causing existing classifiers to make systematic prediction errors?
-
Could the authors compare the proposed gradient-SVD initialization with a randomly initialized orthonormal (k)-dimensional subspace, while keeping the IAP module, training losses, training budget, and orthogonal refinement procedure identical? Additionally, I suggest reporting the principal-angle overlap between the initial spectral subspace and the final learned subspace to quantify the geometric changes during training.
-
Could the authors compare GAUSS against strong adversarial attack methods, such as PGD, MI-FGSM, and DI-FGSM, as well as recent perturbation-based approaches for soft-biometric privacy, under the same perturbation budget and perceptual quality constraints?
-
The experimental setup explicitly identifies DINOv2-L, SigLIP-B, and CLIP-H as held-out foundation-model evaluators. However, Table 3 reports protected-image results only for SigLIP-B and CLIP-H, without providing the corresponding results for DINOv2-L.
5 comments
Gradient-SVD Initialization and Principal-Angle Analysis and DINOv2-L Results
2. Gradient-SVD Initialization and Principal-Angle Analysis
To directly assess the contribution of gradient-SVD initialization, we conducted additional experiments comparing it with randomly initialized orthonormal subspaces of identical rank k=8, while maintaining the same IAP architecture, training objectives, perturbation budget, optimization schedule, and orthogonal refinement procedure. We further analyzed the geometric evolution of the learned subspaces using principal-angle measurements throughout the 40-epoch training.
Table R3C2: Comparison of random orthonormal and gradient-SVD initialization. Attribute suppression (%) and PSNR (dB); higher is better.
| Initialization | Male ↑ | Young ↑ | Smiling ↑ | Mean ↑ | PSNR (dB) ↑ |
|---|---|---|---|---|---|
| Random orthonormal | 92.92 | 95.48 | 85.74 | 91.38 | 30.99 |
| Gradient-SVD (Ours) | 96.82 | 99.56 | 92.88 | 96.42 | 35.79 |
Effectiveness of spectral initialization: Table R3C2 demonstrates that gradient-SVD initialization consistently outperforms random orthonormal initialization, improving suppression by 3.90%, 4.08%, and 7.14% for Male, Young, and Smiling, respectively. This corresponds to a 5.04% improvement in mean suppression, accompanied by a 4.80 dB improvement in PSNR. These findings demonstrate that gradient-SVD provides a more informative attribute-sensitive initialization than arbitrary orthonormal directions, enabling better suppression while preserving higher visual fidelity.
Geometric analysis of learned subspaces: To quantify the influence of subsequent refinement, we measure the principal angles between the initial Gradient-SVD subspace and the final learned subspace. Following Section 3.2 and Appendix A.1, each attribute-specific subspace has rank k = 8 within an ambient image space of dimension d = 49,152. Therefore, comparing the initial and learned subspaces produces eight principal angles, θ₁, θ₂, ..., θ₈, calculated from the singular values of the product of their orthonormal basis matrices. Each principal angle measures alignment along a corresponding pair of principal directions. An angle of 0° indicates perfect alignment, whereas 90° indicates orthogonality. We summarize these eight angles using two complementary metrics:
-
Mean principal angle: The arithmetic mean of the eight individual angles:
θ̄ = (1/8) × Σᵢ₌₁⁸ θᵢ
This indicates their average angular separation.
-
Subspace overlap: The average squared cosine of the eight principal angles:
Overlap = (1/8) × Σᵢ₌₁⁸ cos²(θᵢ)
Table R3C2_1: Principal-angle analysis between initial and final learned subspaces.
| Attribute | Overlap ↑ | Mean Angle (°) | Individual Principal Angles (°) |
|---|---|---|---|
| Male | 0.416 | 51.7 | 16.6, 20.8, 33.7, 49.5, 62.1, 65.7, 80.2, 85.1 |
| Young | 0.427 | 51.8 | 21.5, 27.8, 35.0, 42.1, 58.7, 59.3, 82.4, 87.4 |
| Smiling | 0.431 | 52.0 | 19.9, 22.7, 36.7, 37.3, 52.5, 77.2, 82.7, 86.9 |
Interpretation of subspace refinement: Table R3C2_1 reveals substantial geometric adaptation, with mean principal angles of approximately 52° across all three attributes. The individual angles exhibit a nonuniform distribution: the two smallest angles remain approximately 17° to 28°, indicating relatively strong alignment along certain principal directions, whereas the two largest reach approximately 80° to 87°, indicating considerable reorientation along others. Thus, the learned subspaces neither simply preserve nor completely discard their initial spectral structure; instead, refinement adapts the gradient-informed representation to the selective privacy objective. This observation is consistent with the original ablation in Table 4(a), where freezing the initial basis reduces male suppression from 96.82% to 32.92%, confirming the importance of learnable refinement.
Comparison with Strong Adversarial Attack Methods
3. Comparison with Strong Adversarial Attack Methods
We conducted an additional evaluations using PGD, MI-FGSM, and DI-FGSM, examining white-box attribute suppression, transferability to held-out CNN classifiers, and per-image inference efficiency.
Table R3C3: Comparison with conventional adversarial perturbation methods.
| Method | WB Male ↑ | WB Young ↑ | WB Smiling ↑ | WB Mean ↑ | Held-out CNN Mean ↑ | Time/Image ↓ |
|---|---|---|---|---|---|---|
| PGD | 81.00 | 98.40 | 87.00 | 88.80 | 11.51 | 10,441.21 ms |
| MI-FGSM | 99.80 | 99.80 | 100.00 | 99.87 | 32.00 | 9,853.14 ms |
| DI-FGSM | 96.80 | 99.80 | 97.80 | 98.13 | 32.71 | 9,744.25 ms |
| GAUSS (Ours) | 96.82 | 99.56 | 92.88 | 96.42 | 63.83 | 9.33 ms |
White-box suppression values and held-out CNN flip rates are reported in percentages (%). The held-out CNN mean is averaged over ResNet-50, EfficientNet-B0, and MobileNetV3. Inference time is reported in milliseconds per image for the single-image setting.
Cross-model transferability: The results highlight an important distinction between white-box optimization and cross-model generalization. Conventional adversarial attacks obtain strong mean white-box suppression (88.80% to 99.87%), but their effectiveness decreases considerably on unseen CNN architectures, yielding only 11.51% to 32.71% mean flip rates. In contrast, GAUSS achieves 63.83% mean suppression on held-out CNNs, exceeding the best conventional attack baseline by 31.12%. These findings support the value of evaluating transferability beyond the classifiers directly targeted during perturbation optimization, rather than relying exclusively on white-box suppression performance.
Inference efficiency: GAUSS also demonstrates a substantial computational advantage over conventional iterative attacks. PGD, MI-FGSM, and DI-FGSM require approximately 9.74 to 10.44 seconds per image, whereas our proposed approach GAUSS requires only 9.33 milliseconds in the single-image setting. This corresponds to approximately 1,119×, 1,056×, and 1,044× faster inference than PGD, MI-FGSM, and DI-FGSM, respectively.
We will refine the paper with all these concerns wherever best fit and applicable in the paper.
Attribute Recovery via Prediction Inversion and DINOv2-L Evaluation
1. Attribute Prediction Inversion and Privacy Protection
We thank the reviewer for highlighting the possibility of recovering protected attributes through prediction inversion. While we acknowledge that reversing predictions may counteract protection under certain conditions, its effectiveness mostly depends on the adversary's knowledge and the consistency of prediction reversal across different models. In a white-box setting, where perturbations are specifically optimized against a known classifier, GAUSS achieves approximately 96% to 98% suppression. Under such conditions, an adversary who knows the protection mechanism and its systematic prediction-reversal behavior could potentially recover the original binary attributes through inversion. However, this assumption does not directly generalize to practical black-box scenarios, where the underlying classifiers, decision boundaries, and responses to protected images may differ substantially from those used during perturbation optimization. Importantly, our cross-model evaluations demonstrate that the prediction flip rates vary considerably across unseen architectures, typically ranging from approximately 50% to 75%. For binary attributes, indiscriminately reversing these predictions would yield only 50% to 75% recovery accuracy, substantially lower than the near-perfect recovery achievable in the white-box setting. Also, when the flip rate approaches 50%, prediction inversion provides little advantage over random guessing. Thus, successful inversion against a known classifier does not establish a universally reliable recovery strategy across heterogeneous inference systems. Also, real-world automated inference pipelines generally lack explicit information about whether a particular image has been protected, which attributes are selected for suppression, and whether the corresponding classifier predictions should be reversed. Applying inversion indiscriminately to both protected and unprotected images may introduce additional prediction errors. For multi-class attributes, the challenge becomes more complex because incorrect predictions do not necessarily correspond to a unique, reversible mapping between the original and predicted classes.
4. Missing DINOv2-L Evaluation
We have added the complete protected-image evaluation across all three held-out foundation models below.
Table R3C4: Attribute prediction flip rates (%) on held-out foundation models and perceptual quality.
Although PrivAD achieves higher suppression on CLIP-H and DINOv2-L, GAUSS achieves a higher mean suppression rate across the three foundation-model evaluators (32.61% versus 30.82%). Importantly, GAUSS also maintains substantially higher visual fidelity (35.79 versus 19.40 dB PSNR). These findings demonstrate that GAUSS provides a favorable balance between average cross-model attribute suppression and preservation of the original facial appearance. We will include the missing DINOv2-L results in the revised manuscript.