acceptodds
Under review as a conference paper at ICLR 2027

ATFeaturePure: Learning A Class-Conditional Vector Field in Decision-Relevant Feature Space for Adversarial Defense

55 views

Abstract

Adversarial training (AT) provides a reliable foundation for adversarial robustness, yet adversarial perturbations can still substantially corrupt the internal representations that determine model predictions. Rather than recovering the entire clean input with a high-dimensional generative purifier, we investigate whether AT can be strengthened by directly correcting these decision-relevant representations. We introduce , 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. The key challenge is how to learn correction dynamics that actually improve classification robustness. To address this, we develop a local theoretical analysis that identifies the behavior required for effective feature correction and translate these insights into a learning mechanism that jointly trains the classifier and vector field while encouraging controlled local sensitivity, reliable class guidance, correction of adversarial features toward clean representations, and decision-consistent predictions. 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.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial defense, Adversarial attack, Flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tiqQOpenReview ↗
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
  1. The paper consider a interesting research problem.
  2. The experimental study is broad, which is conducted on multiple scenarios.
Critical weaknesses
  1. 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.
  2. The loss function and overview of the proposed method (Figure 1) are too complex and difficult to follow.
  3. The paper lacks sufficient novelty. The proposed method combines multiple existing techniques.
  4. The paper lacks the ablation study to show the effectiveness of each component in the loss function.
  5. The evaluation against adaptive attacks remains insufficient.
  6. 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
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

Reviewer_RyMmOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper adds a small class-conditional vector field to the pre-logit feature of an adversarially trained classifier, moving the feature a few steps toward the predicted class before classification. It reports 11.7- to 32.0-point robust-accuracy gains over AT baselines on CIFAR-10/100 and Tiny-ImageNet under PGD, AutoAttack, and DiffHammer, at a lower training cost than purification baselines.

Critical strengths

Working on the compact pre-logit feature instead of pixels is a cheap idea worth trying, but none of the evidence for it can be trusted yet.

Critical weaknesses

The robustness numbers look like gradient masking. In Tables 1 and 2, PGD robust accuracy is at or above clean accuracy (64.19 vs 60.93 on CIFAR-100, 51.90 vs 50.34 on Tiny-ImageNet), which a sound evaluation should not produce. Eq. 6 picks the guidance label with a hard argmax, and App. F.2 says gradients flow through the whole purification, but the paper never says how the argmax was differentiated. No standard checks are run (large-epsilon attacks, BPDA, attacking the raw K=0 classifier), so the adaptive-robustness claim is not supported.

The theory is not tied to the trained models. The constants in Assumption 3.2 and Thm 3.4 are never measured, so I can't tell whether the conditions hold, and a Lipschitz bound for a deep ResNet under l-infinity perturbations is likely too loose to matter. I read it as motivation, not as an explanation of the gains.

Questions for authors

How is the argmax label differentiated in the PGD, AA and DiffHammer attacks?

Why does PGD robust accuracy exceed clean accuracy on CIFAR-100 and Tiny-ImageNet?

What is the robust accuracy of the raw jointly trained classifier (K=0) under the same attacks?

What happens at epsilon 16/255 and above? Does robustness go to zero?

Can you report BPDA or a straight-through soft-label attack?

3 comments
AuthorsOpenReview ↗

Authors' response to Q1–Q2

We thank the reviewer for the detailed feedback. We would first like to emphasize that our original evaluation already addresses gradient-masking concerns through complementary gradient-based and gradient-free attacks. In particular, Appendix F.2 explicitly includes Square Attack as part of our AutoAttack evaluation. As demonstrated by Andriushchenko et al. (Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search, ECCV 2020) and Croce and Hein (Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse Parameter-Free Attacks, ICML 2020), Square Attack does not rely on local gradients and provides an important complementary evaluation when gradient masking is a concern. Our reported results therefore do not rely solely on potentially misleading PGD gradients; they also include gradient-free adversarial evaluation of the complete deployed defense. Below, we clarify the differentiation procedure, PGD evaluation results, theoretical foundations, and scope of our robustness assessment.

Q1. “How is the argmax label differentiated in the PGD, AA and DiffHammer attacks?”

The discrete guidance label should be distinguished from the differentiable feature-purification pathway. In Eq. (6), the raw prediction selects the class-conditioned vector field, whereas the continuous pre-logit representation (hθ(x)h_\theta(x)) remains the principal input to purification. Crucially, the guidance selection neither detaches these features nor interrupts their gradient flow.

For a locally guidance class, attack gradients propagate directly through the feature encoder, all unrolled Euler purification steps, and the final classification head. Thus, although the discrete argmax itself is not differentiated, the active computation path yields the exact local gradient of the deployed classifier, rather than a detached or approximate purification gradient, as described in Appendix F.2.

Furthermore, our evaluations attack the complete deployed ATFeaturePure model using multi-restart PGD, AutoAttack (including the gradient-free Square Attack), and DiffHammer. Tables 1–2 show substantial accuracy reductions under AutoAttack and DiffHammer relative to PGD, demonstrating that these attacks effectively challenge the complete defense. Therefore, nondifferentiability of the guidance selector should not be conflated with blocked gradient propagation through the feature-correction mechanism.

Q2. “Why does PGD robust accuracy exceed clean accuracy on CIFAR-100 and Tiny-ImageNet?”

The distinction between theoretical worst-case robust accuracy and empirical accuracy under a finite-step PGD evaluation is essential here. Theoretical robust accuracy considers all admissible perturbations within the specified (ℓ∞\ell_\infty) neighborhood, including the unperturbed input, and therefore cannot exceed clean accuracy. In contrast, PGD approximates adversarial optimization through a finite number of iterations and restarts. Our original PGD implementation measures classification accuracy at the returned adversarial points, where initially incorrect predictions may become correct. Consequently, this empirical statistic is not subject to the same inequality as theoretical worst-case robust accuracy.

To investigate the observation, we additionally conducted controlled diagnostics reproducing the original PGD-20×3 evaluation using the same trained checkpoints, full test sets, preprocessing, purification settings, and attack parameters (ε\varepsilon=8/255, α\alpha=2/255, 20 iterations, and 3 restarts). We additionally tracked prediction correctness at the clean inputs and throughout the attack trajectories.

On CIFAR-100 (seed 27837429), clean accuracy is 61.41%, while returned-point PGD accuracy is 65.78%. Specifically, 469 initially incorrect predictions become correct, versus 32 changes in the opposite direction, exactly accounting for the +4.37 percentage-point difference. On Tiny ImageNet (seed 19205881), the corresponding accuracies are 50.13% and 52.21%, with 277 incorrect-to-correct and 69 correct-to-incorrect transitions, accounting for the +2.08-point difference.

These controlled diagnostics quantitatively establish that the apparent inversion results from the finite PGD evaluation and returned-point accounting.

AuthorsOpenReview ↗

Authors' response to Q3–Q5

Q3. “What is the robust accuracy of the raw jointly trained classifier (K=0) under the same attacks?”

ATFeaturePure is not a simple combination of independently optimized adversarial training and post-processing modules. Rather, it is an integrated, jointly learned defense whose feature-correction mechanism is motivated by the theoretical analysis in Section 3.2 and implemented through the coupled learning objectives in Eqs. (21)–(28).

Importantly, setting (K=0) removes the class-conditional feature-correction dynamics that constitute the central mechanism of ATFeaturePure. Under Theorem 3.4, the theoretically motivated feature-region contraction is characterized by (rK=ρKr0r_K=\rho^K r_0); setting (K=0) eliminates this contraction effect (rK=r0r_K=r_0). The resulting raw classifier is therefore an altered inference configuration, not the complete defense defined in Eq. (6).

Accordingly, our primary empirical evaluation assesses the jointly optimized ATFeaturePure model as a complete deployed system, under established end-to-end PGD, AutoAttack, and DiffHammer protocols. The consistent improvements over independently trained AT or purification-based baselines in Tables 1–3 provide direct evidence of the effectiveness of the proposed integrated framework.

Q4. “What happens at epsilon 16/255 and above? Does robustness go to zero?”

Our experimental protocol follows the established (ℓ∞\ell_\infty) adversarial robustness benchmark at ε\varepsilon=8/255, rather than an arbitrarily selected perturbation budget. This is the standardized CIFAR-10/100 setting adopted by RobustBench (NeurIPS 2021) and the principal evaluation setting used by recent adversarial-training and purification methods, including DAT (NeurIPS 2024), DHAT (ICCV 2025), and ADBM (ICLR 2025).

Importantly, ATFeaturePure is trained and evaluated under the same specified perturbation budget (ε\varepsilon=8/255), consistent with the established benchmark protocol. As detailed in Section 4.1 and Appendix F.2, we evaluate the complete defense using multi-restart PGD, AutoAttack, and DiffHammer, with the evaluation perturbation budget matched across competing methods. This ensures a consistent and directly comparable assessment of adversarial robustness.

Evaluating at ε\varepsilon=16/255 introduces a different, larger threat model beyond the scope of this benchmark. Our reported empirical robustness comparisons concern the standardized ε=8/255\varepsilon=8/255 setting and do not claim robustness at arbitrary perturbation magnitudes. Consequently, performance at an unexamined larger budget cannot be inferred from the reported results and does not determine the validity of the established (8/255) comparison.

Q5. “Can you report BPDA or a straight-through soft-label attack?”

Our evaluation follows the direct-gradient methodology supported by existing adversarial-purification research, rather than relying on approximate gradients (like BPDA) where direct differentiation is available. In particular, Lee and Kim (Robust Evaluation of Diffusion-Based Adversarial Purification, ICCV 2023), explicitly cited in our manuscript, investigate this precise issue in their RQ3. Their Table 3 demonstrates that full/surrogate, backpropagation-based, attacks substantially outperform BPDA: robust accuracy decreases from 66.91% to 33.48% for ADP and from 75.59% to 24.06% for GDMP. The authors consequently recommend direct backpropagation through the purification process whenever feasible.

Our evaluation deliberately follows this established methodology. As detailed in Appendix F.2, attack gradients are computed by directly backpropagating through the complete purification trajectory, including the feature encoder, all unrolled Euler steps, and the final classification head. We additionally employ multi-restart PGD, all AutoAttack components (including gradient-free Square Attack), and DiffHammer, with EOT-based evaluation also reported in Table 3.

As explained in Q1, class guidance selects the active vector field without interrupting differentiation through the feature-correction pathway; direct backpropagation therefore yields the exact local gradient of the deployed model within each fixed-guidance region.

AuthorsOpenReview ↗

Authors' response to the concern “The theory is not tied to the trained models……. is likely too loose to matter”

Theorem 3.4 provides a sufficient-condition characterization of the proposed robustness mechanism, rather than a numerical certificate for an individual trained checkpoint. Under explicitly stated local conditions, it establishes that class-conditional feature correction contracts the perturbation-induced feature region and preserves classification margins. Corollary 3.5 further connects this analysis to the pseudo-label guidance used at inference. These results identify the theoretical properties underlying our defense design and are explicitly connected to the practical learning objectives in Section 3.3 and Appendix C.

Crucially, the theory is explicitly connected to the implemented learning procedure. Section 3.3 and Appendix C establish these connections: adversarial cross-entropy promotes reliable raw-label guidance through Lemma 3.6; class-conditional flow matching encourages inward corrective directions through Eq. (25); feature alignment and purified classification promote decision-consistent representations through Lemma C.1; and parameter regularization encourages controlled sensitivity.

Furthermore, Assumption 3.1 concerns local, rather than global, Lipschitz behavior. The potential looseness of global ResNet Lipschitz bounds therefore does not invalidate the stated local theoretical results. The theory explains the design principles, while the realized robustness of the trained models is evaluated empirically through the end-to-end attacks reported in Tables 1–3.

Reviewer_hG4XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ATFeaturePure, a defense framework that combines adversarial training with feature-space purification. The central motivation is that classification may not require reconstructing an entire clean image in high-dimensional pixel space. Experiments cover CIFAR-10, CIFAR-100, and Tiny ImageNet with ResNet-18 and WRN28-10. Across comparisons with AT baselines, the paper reports an average robust-accuracy improvement.

Critical strengths
  1. Feature-space purification is a promising design direction. The resulting approach can be jointly trained with an adversarially robust classifier and has potentially substantial advantages over diffusion-based input purification in parameters and training cost.
  2. The experimental scope and basic reporting practices are strong.
Critical weaknesses
  1. The deployed purifier depends on the raw prediction. An attacker can first force the raw classifier to select an incorrect guidance class and then exploit the corresponding incorrect vector field. Standard gradients generally treat the integer pseudo-label as fixed during a forward-backward pass and do not capture the non-smooth branch change. Thus, adaptive evaluation is not yet convincing for the discrete pseudo-label routing mechanism.
  2. The clean-feature preservation assumption concerns purification starting from a clean feature. The alignment loss instead compares a purified adversarial feature with its paired clean representation. The paper explicitly acknowledges that this does not directly enforce the assumption. Parameter-norm regularization is also only a loose surrogate for the local Lipschitz and vector-field quantities in the theorem.
Questions for authors
  1. Does an incorrect class-conditional vector field reinforce an incorrect raw prediction? Please compare ground-truth, raw-prediction, soft-label, random-label, and worst-case-label guidance.
  2. Why is PGD accuracy approximately 15–23 points higher than AutoAttack accuracy?
  3. What causes the 4.55-point and 4.98-point standard deviations for WRN28-10 under PGD?
  4. Can the authors empirically measure whether purification contracts within-class features and preserves clean features?
2 comments
AuthorsOpenReview ↗

Authors' response to Q1–Q2

We thank the reviewer for recognizing the potential of feature-space purification and the breadth of our experimental evaluation. We appreciate the opportunity to clarify the class-guidance mechanism, the scope of our theoretical results, and the evaluation protocol.

Q1: Class-conditional guidance and the evaluation protocol

We would first clarify that class guidance in ATFeaturePure does not dictate the final classification. As defined in Eq. (6), the raw prediction selects the class-conditional vector field, while the final prediction is determined by the classification head after feature correction. Thus, an incorrect raw prediction is not mechanically propagated into the final prediction. Moreover, the raw classifier is adversarially trained, and the correction mechanism is jointly optimized through the objectives in Eqs. (21)–(28).

Raw-prediction guidance is the prescribed inference mechanism of ATFeaturePure, rather than an interchangeable option. Ground-truth guidance would introduce oracle information unavailable at inference. Soft-label guidance would require a different conditioning mechanism, whereas random-label and externally imposed worst-case-label guidance would replace the inference rule explicitly defined in Eq. (6). Consequently, these alternatives constitute different conditioning schemes rather than controlled evaluations of the proposed inference mechanism.

Our robustness evaluation follows the established input-perturbation protocols and evaluates the complete, unchanged ATFeaturePure classifier end-to-end using PGD, AutoAttack, and DiffHammer, as detailed in Section 4.1 and Appendix F.2. This ensures that our comparisons with the established baselines assess the method as proposed and deployed.

Q2: Difference between PGD and AutoAttack

The observation that AutoAttack (AA) yields lower robust accuracy than PGD is well established in adversarial robustness evaluation. For example, the original DAT (NeurIPS 2024) and DHAT (ICCV 2025) papers report the same pattern across CIFAR-10 and CIFAR-100, with AA consistently producing lower robust accuracy than PGD. This difference arises because AA combines complementary attacks, such as APGD-CE, APGD-DLR, FAB, and Square Attack, etc, which employ diverse optimization objectives and mechanisms, allowing them to identify adversarial examples that conventional fixed-step PGD may miss. Consequently, the magnitude of the accuracy difference depends on the evaluated model and the effectiveness of the respective attacks.

AuthorsOpenReview ↗

Authors' response to Q3–Q4

Q3: Standard deviations under PGD

The reported standard deviations reflect variations across four independently trained random seeds (See Appendix F.3), rather than repeated attack evaluations of a fixed model. Specifically, the PGD-20×3 accuracies are 83.77%, 78.45%, 77.57%, and 87.20%, yielding 81.75 ± 4.55%, while PGD-100×10 produces 81.51%, 76.39%, 75.75%, and 86.40%, yielding 80.01 ± 4.98%. Both statistics are computed using the sample standard deviation.

Importantly, even the lowest individual PGD results substantially exceed the strongest competing baseline means reported in Table 2. Furthermore, the variability is considerably smaller under AutoAttack (64.91 ± 0.97%) and DiffHammer (standard deviations of 1.39–1.42 points), demonstrating that the larger dispersion is concentrated in the PGD evaluations rather than observed uniformly across attack methods. Our comparative robustness advantage therefore does not depend solely on the PGD results.

Q4. Empirical feature contraction and clean-feature preservation

We would clarify the direction of our theoretical analysis: Theorem 3.4 establishes sufficient local conditions for robust feature correction and margin preservation, rather than necessary conditions characterizing every robust representation. These properties guide the learning design in Section 3.3 and Appendix C, where class-conditional flow matching encourages corrective dynamics toward clean representations, feature alignment promotes recovery of clean features, and purified classification encourages positive post-purification margins.

Importantly, the theorem characterizes contraction toward a class reference, which is not equivalent to requiring every pairwise within-class feature distance to decrease. Our primary objective is adversarial robustness, not feature-distance minimization itself. Accordingly, the proposed mechanism is evaluated through the complete deployed classifier under established PGD, AutoAttack, and DiffHammer protocols, consistently applied across competing baselines. These end-to-end robustness results provide direct empirical evidence of the method's effectiveness, while geometric measurements would offer complementary characterization rather than replace robustness evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.