- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes ATFeaturePure, which retains AT as the robust learning foundation and augments the classifier with a lightweight class-conditional vector field operating on the compact pre-logit feature. ATFeaturePure uses a theory-guided class-conditional vector field to correct adversarially corrupted pre-logit features and strengthen adversarial training. Extensive experiments across three datasets, two backbones, and multiple strong attacks show that ATFeaturePure consistently improves robustness over recent AT and purification baselines, including against adaptive attacks, while maintaining favorable training efficiency. These results show that theory-guided correction of decision-relevant features can effectively strengthen AT without requiring full input reconstruction.
Critical strengths
- The paper consider a interesting research problem.
- The experimental study is broad, which is conducted on multiple scenarios.
Critical weaknesses
- The paper suffers from poor readability and presentation quality. Much of the content appears to be AI-generated, containing numerous unnecessary details, and overly engineering-oriented descriptions. The manuscript reads more like an experimental report than an academic paper.
- The loss function and overview of the proposed method (Figure 1) are too complex and difficult to follow.
- The paper lacks sufficient novelty. The proposed method combines multiple existing techniques.
- The paper lacks the ablation study to show the effectiveness of each component in the loss function.
- The evaluation against adaptive attacks remains insufficient.
- The computational cost metrics reported in the paper are not sufficiently intuitive. I suggest directly comparing the inference time per image (in seconds) across different methods.
Questions for authors
See weakness
3 comments
Authors' response to W4–W6
W4. Contribution of individual learning components.
The concern appears to treat the proposed framework as a collection of independently functioning components, whereas ATFeaturePure is formulated and optimized as an integrated defense. Although Eq. (28) expresses the training objective as a weighted sum, its terms jointly optimize shared model parameters and complementary properties of the same feature correction mechanism. Their roles are explicitly connected to the theoretical analysis in Section 3.3 and Appendix C.
A decomposition of the training objective should not be confused with a decomposition of the proposed defense into independently operating methods. Our central contribution concerns the jointly learned correction mechanism and its realized performance as a complete system. Tables 1–3 directly evaluate this system against established adversarial training and purification baselines across multiple datasets, architectures, and attack families.
W5. Adaptive-attack evaluation.
We respectfully disagree that the adaptive-attack evaluation is insufficient, as the original manuscript already incorporates established advanced attacks and follows published evaluation recommendations for adversarial purification. Specifically, our protocol includes multi-restart PGD, the complete AutoAttack configuration, including gradient-free Square Attack, and DiffHammer, which explicitly targets purification defenses.
Importantly, Appendix F.2 follows the direct-backpropagation evaluation principle established by Lee and Kim (Robust Evaluation of Diffusion-Based Adversarial Purification, ICCV 2023), who demonstrated that direct-gradient attacks can be substantially more effective than BPDA and recommended differentiating through the purification process whenever feasible. Our attacks therefore propagate gradients through the complete feature-correction trajectory rather than bypassing or approximating the purifier. Table 3 additionally reports EOT-based evaluation. These experiments provide a methodologically grounded, complementary evaluation of the complete deployed defense.
W6. Computational-efficiency evaluation.
The computational-cost measurements in Table 4 are specifically designed to assess total training computation and model complexity, not hardware-specific inference latency. We report total PFLOPs and learnable parameters for all compared methods, including the validation-based hyperparameter selection cost of ATFeaturePure.
Unlike execution time measured in seconds, FLOP counts provide a hardware-independent measure of computational work, avoiding variations arising from GPU architecture, CPU–GPU communication, memory bandwidth, batch size, and implementation-specific optimizations. Inference latency is a distinct hardware-dependent performance metric and cannot replace a comparison of total training computation.
Our reported measurements therefore provide a quantitative and directly comparable assessment of the training-efficiency claims made in Section 4.3.
In summary, ATFeaturePure presents a theoretically motivated, jointly optimized feature correction mechanism, supported by extensive adversarial evaluation and quantitative training efficiency comparisons. We believe these substantive theoretical, methodological, and empirical contributions warrant assessment on their demonstrated scientific merits.
Thank you for the rebuttal. However, the authors have not provided substantial evidence, such as additional experiments or revisions to the manuscript, to address my concerns. Therefore, I will maintain my original score.
Authors' response to W1–W3
We thank the reviewer for acknowledging the interesting research problem, broad experimental evaluation, and generally sound methodology. However, we respectfully disagree with several characterizations of our work, which overlook important methodological, theoretical, and experimental evidence explicitly presented in the submission. We clarify these points below.
W1. Presentation quality and AI usage.
We respectfully disagree with the characterization of our manuscript as an engineering-oriented experimental report. The paper follows a rigorous progression from theoretical characterization (Section 3.2), to theory-motivated learning (Section 3.3), and systematic empirical evaluation (Section 4). Theorem 3.4, Corollary 3.5, Lemma 3.6, and the complete derivations in Appendices B–C constitute substantive mathematical contributions.
Regarding the impression of AI-generated content, our original submission already includes an explicit AI Use Statement on page 10. AI assistance was limited to language polishing, manuscript drafting and editing, and presentation and formatting. The research ideas, methodology, experimental design and execution, mathematical results, literature review and citation selection, and scientific conclusions were developed and finalized by the authors. All AI-assisted materials were reviewed and revised by the authors, who assume full responsibility for the manuscript.
W2. Complexity of Figure 1 and the learning objectives.
The learning objectives are not an arbitrary collection of loss functions. Each component has a specific role derived from the theoretical analysis: adversarial classification promotes reliable guidance (Lemma 3.6), class-conditional flow matching encourages inward corrective dynamics (Eq. 25), feature alignment and purified classification promote margin preservation (Lemma C.1), and regularization encourages controlled sensitivity. These connections are explicitly derived in Section 3.3 and Appendix C. Figure 1 depicts the corresponding integrated training and inference procedures.
W3. Methodological novelty.
The characterization of ATFeaturePure as merely combining existing techniques overlooks its central methodological contribution. ATFeaturePure introduces a jointly learned, class-conditional feature-correction mechanism operating on compact, decision-relevant pre-logit representations. Its corrective vector field performs iterative feature transformations rather than relying on conventional input-space reconstruction or fixed feature-denoising operators.
More importantly, this mechanism is developed from an explicit theoretical characterization, connecting contractive feature dynamics, classification-margin preservation, and practical pseudo-label guidance (Theorem 3.4 and Corollary 3.5). These properties motivate the coupled learning objective in Section 3.3 and its mathematical derivations in Appendix C. The contribution is therefore a distinct, theory-motivated robustness mechanism with a unified learning and inference formulation, not simply an aggregation of existing methods.