acceptodds
Under review as a conference paper at ICLR 2027

Bending Early, Curbing Deep: Test-Time Adversarial Defense for Large Vision-Language Models Based on Residual Curvature

Abstract

Although Large Visual-Language Models (LVLMs) have achieved significant advances, their vulnerability to adversarial perturbations poses a persistent challenge to their reliable deployment in the real world. Prevalent test-time defenses largely focus on mitigating adversarial effects while overlooking how adversarial deviations propagate through the ViT-based visual encoder. In particular, their reliance on multi-view transformations and closed-set class space limits their applicability to generative tasks in LVLMs. In this paper, we find that adversarial perturbations first induce a geometric abnormality of early residual trajectories by shifting them from differential toward common component, with the resulting deviations subsequently amplified predominantly through deep MLP branches. Therefore, we propose Curvature-based Universal Residual Bending (CURB), a test-time defense that selectively bends abnormally flattened residual trajectories through input updates to curb adversarial growth. CURB does not depend on downstream outputs space, allowing universal and direct operation across open-ended generation and zero-shot classification. Experiments on LLaVA-1.5-7B and InternVL3.5-8B show consistent robustness gains in various tasks, together with substantial reductions in adversarial representation deviations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Attacks, Test-Time Defense, Robustness

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

This paper proposes a test-time adversarial defense method based on early residual curvature. The authors find experimentally that adversarial perturbations reduce the curvature of early residual trajectories in ViT and amplify representation discrepancies through deep MLP branches. CURB uses a gating mechanism and gradient-based input pixel updates to increase early residual curvature for abnormal samples, thereby curbing adversarial propagation. Experiments show that the method effectively improves adversarial robustness on image captioning, visual question answering, and zero-shot classification tasks while largely preserving clean performance.

Critical strengths
  1. By comparing the clean–adversarial representation discrepancies across different layers of ViT-based visual encoders, the authors observe that the representation discrepancies caused by adversarial perturbations gradually increase toward deeper layers. The authors also find that the early residual curvature of adversarial samples tends to be lower than that of clean samples, and attempt to explain the connection between this phenomenon and deep-layer representation discrepancy growth through an analysis of the common component and differential component. The overall research motivation is clear and has some value for mechanistic analysis.
  2. The authors propose CURB, a test-time defense method that does not require retraining the model. It increases early residual curvature by optimizing the input image to alleviate representation discrepancies caused by adversarial perturbations. The method is independent of the downstream output space and introduces selective activation to avoid unnecessary modifications to all inputs.
Critical weaknesses
  1. The current experiments are mainly based on statistical analysis of curvature distributions and do not explicitly report the proportions of successful adversarial examples whose curvature decreases or increases. Therefore, I am uncertain whether the phenomenon proposed in the paper is general.
  2. The gating threshold proposed in the paper should be a very important hyperparameter. However, the paper only states that the 5th percentile of the clean curvature distribution is used as the gating threshold, without providing a detailed explanation or discussing robustness to the choice of this gating threshold.
Questions for authors
  1. Among successful adversarial examples, what proportions exhibit a decrease and an increase in early residual curvature, respectively? Furthermore, how effective is CURB on successful adversarial examples whose curvature does not decrease or that do not trigger the CURB gating mechanism?
  2. What is the specific rationale for choosing the 5th percentile rather than other percentiles as the gating threshold? When the threshold is set to the 1st, 5th, or 10th percentile, how do clean performance, adversarial robustness, and the gating activation rate change?
4 comments
AuthorsOpenReview ↗

W1: Prevalence of decreased curvature among successful attacks

Thanks for your comments. We conducted a paired analysis on the same 500 APGD examples per dataset used in Table 1. For every successfully attacked pair, we compute ΔC=C(xadv)−C(xclean)\Delta C=C(x_{\mathrm{adv}})-C(x_{\mathrm{clean}}). An attack is considered successful when its per-sample task score is lower than that of the corresponding clean input.

ModelDatasetSuccessful attacksΔC<0\Delta C<0ΔC>0\Delta C>0
LLaVACOCO494494 (100.0%)0 (0.0%)
LLaVAFlickr30k495495 (100.0%)0 (0.0%)
LLaVATextVQA202202 (100.0%)0 (0.0%)
LLaVAVQAv2403403 (100.0%)0 (0.0%)
InternVLCOCO492484 (98.4%)8 (1.6%)
InternVLFlickr30k491488 (99.4%)3 (0.6%)
InternVLTextVQA401401 (100.0%)0 (0.0%)
InternVLVQAv2404402 (99.5%)2 (0.5%)

For Q1.1: Overall, curvature decreases for 1,594/1,594 (100%) successful attacks on LLaVA and 1,775/1,788 (99.3%) on InternVL. Thus, the observation is not merely distributional, it holds for 3,369/3,382 (99.6%) successfully attacked clean–adversarial pairs.

For Q1.2: Importantly, ΔC>0\Delta C>0 does not necessarily prevent CURB from working, because the gate depends on the adversarial input’s absolute curvature, rather than its change relative to the unavailable clean counterpart. Among the 13 InternVL exceptions with ΔC>0\Delta C>0, 12 still activate the gate and all 12 obtain improved task scores:

  • COCO: 7/8 recover; CIDEr increases from 10.03 to 147.02.
  • Flickr30k: 3/3 recover; CIDEr increases from 1.93 to 41.05.
  • VQAv2: 2/2 recover; accuracy increases from 0.0 to 80.0.

This analysis clarifies that the paired curvature observation is highly prevalent, while the remaining robustness gap is primarily attributable to the conservative gate rather than violations of the curvature observation. We will add these paired statistics and the exception analysis to the revised appendix.

AuthorsOpenReview ↗

W2: Sensitivity to the gating threshold

Thanks for your comments. We evaluated LLaVA on the same APGD-500 pairs, using one cross-dataset threshold defined by the 1st, 5th, or 10th percentile of the clean-curvature distribution.

PercentileClean gateAdv. gateCOCO C/AFlickr30k C/ATextVQA C/AVQAv2 C/A
1st1.0%15.1%115.5 / 23.377.8 / 18.137.2 / 6.776.1 / 12.0
5th5.0%88.9%115.4 / 116.077.2 / 73.936.9 / 28.074.5 / 65.3
10th10.0%91.0%115.2 / 116.376.9 / 74.536.7 / 29.474.2 / 65.7

Here, C/A denotes clean/adversarial CIDEr for captioning and accuracy for VQA; gate rates are pooled across the four datasets. The 1st percentile is overly conservative and misses most adversarial inputs. Increasing the threshold from the 5th to the 10th percentile doubles clean activation, while adversarial robustness improves only marginally.

Notably, this sensitivity analysis shows that activating CURB for the small subset of low-curvature clean inputs has only a marginal effect on clean performance, consistent with our finding in Sec. 4.4, Ablation of Gating Threshold and Optimization Direction, that gate activation on clean inputs does not necessarily lead to clean performance degradation. This is why a lower percentile of the clean distribution can serve as the gating threshold. And we will add this sensitivity analysis to the Sec. 4.4

Dear Reviewer JaXM,

We would be grateful if you could review our rebuttal at your convenience. We have provided additional experiments and revised the manuscript accordingly, with the changes highlighted in blue. we would be grateful if you could let us know whether our response and revisions adequately address your concerns, and we are happy to provide additional clarification for further questions.

Thank you for your time and consideration.

AuthorsOpenReview ↗

Summary of Concerns Addressed and Request for Reassessment

Dear reviewer,

Thank you again for your thoughtful review. As the discussion period closes soon, we would appreciate your feedback on whether our additional response resolve your concerns:

  • W1: Our paired analysis shows decreased curvature in 3,369/3,382 (99.6%) successfully attacked samples. Among the 13 exceptions, 12 still activate the gate and all 12 obtain improved task performance, directly quantifying both prevalence and exceptional cases.

  • W2: Our sensitivity analysis over the 1st, 5th, and 10th percentiles shows that the 1st percentile misses most adversarial inputs, whereas increasing from the 5th to the 10th percentile doubles clean activation with only marginal robustness activation. This supports the 5th percentile as the best trade-off.

Both analyses have been incorporated into the revised manuscript. Could you please let us know whether these additions resolve your concerns? If so, we would greatly appreciate your reconsideration of the assessment and score.

Reviewer_reb4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to promote robustness to adversarial attacks for LVLMs, which is important for reliable uses of large models. The authors study the influence of adversarial attacks across different layers and accordingly propose to increase directional variation in the early residual trajectory. Experiments show that the proposed method can outperform compared methods.

Critical strengths
  1. The research problem is important and interesting.
  2. The study reveals important characteristics of different layers in LVLMs to adversarial attack.
  3. Experiments show that the proposed method can outperform compared methods.
Critical weaknesses
  1. The authors propose to target important layers to implement adversarial attack, which has been previously studied by related works, such as [1-2].
  2. The authors theoretical analyses the important of early layers, which is important. However, in Observation 3, the analysis depends on γm>γr\gamma_m>\gamma_r, which might not be practical in real world applications.
  3. In the experiments, the method is evaluated under only APGD with an (\ell_\infty) perturbation budget of 4/255, which is not enough to verify the effectiveness and practical ability of the proposed method.
  4. The proposed method focuses on increasing directional variation in early residual trajectories. However, it remains unclear whether targeting early layers is consistently more effective than targeting other individual layers or combinations of layers.
  5. What is the rationale of setting the value of per-step rectification magnitude \alpha?
Questions for authors

Please refer to the Critical Weaknesses.

9 comments
AuthorsOpenReview ↗

Clarifications and Additional Evidence on CURB

W1: Distinction from Layer-Targeted Adversarial Attacks

Thanks for your comments. We would like to clarify that we do not propose targeting important layers to construct adversarial attacks. Instead, CURB is a training-free defense.

Could you please provide the bibliographic information for the related works[1]–[2]? To the best of our knowledge, no prior adversarial defense has identified the link between early abnormal residual update and deep discrepancy amplification. CURB first reveals this connection and uses it to correct residual curvature in the early frozen visual layers.

W2: Validity and Scope of the Local Condition γm>γr\gamma_m>\gamma_r

Thank you for raising this concern. We would like to clarify that γm>γr\gamma_m>\gamma_r is a local sufficient condition along the considered finite-step rectification trajectory, as explicitly stated in Proposition 3. Empirically, we verified γm>γr\gamma_m>\gamma_r for both LLaVA and InternVL across all four generative benchmarks, as well as eight classification benchmarks, without observing a violation in the evaluated settings.

Importantly, these models and datasets are widely common used, community-established benchmarks for evaluating multimodal vision tasks, rather than settings selected specifically to favor our hypothesis. Therefore, the condition is a consistently supported local property across representative architectures and tasks.

W3: Robustness across Different Adversarial Perturbation Budgets

Thanks for your comments. We would like to clarify that the current evaluation is not limited to standard APGD: the generative experiments use white-box ensemble APGD, the eight classification benchmarks use PGD, and Table 4 further evaluates BPDA combined with AutoAttack against the complete test-time defense. To assess budget sensitivity, we additionally evaluate CURB at ϵ∈{2/255,4/255,8/255}\epsilon\in\{2/255,4/255,8/255\} on generative tasks using the same APGD-500 protocol. We will add this budget sweep in the final version.

BudgetCOCO +BaseCOCO +CURBFlickr30k +BaseFlickr30k +CURBTextVQA +BaseTextVQA +CURBVQAv2 +BaseVQAv2 +CURB
2/2553.6122.11.184.50.232.61.873.2
4/2552.7116.00.873.90.228.00.165.3
8/2552.3115.60.472.40.026.70.063.1

W4: Layer-Wise Ablation of the Curvature Measurement

Thanks for your comments. We would like to clarify that this comparison has already been performed in Table 6. Moving the curvature objective from C2C_2 to C5C_5 and C8C_8 significantly reduces the adversarial robustness, while the clean performance remains stable.

Moreover, CURB defines discrete residual curvature using consecutive residual updates. Therefore, arbitrary combinations of nonconsecutive layers are not applicable to our curvature objective. These results support targeting the earliest measurable directional abnormality before downstream nonlinear transformations amplify the discrepancy.

W5: Rationale for the Rectification Step Size α\alpha

Thanks for your comments. The parameter α\alpha is the learning rate, or more precisely, the pixel-space step size of CURB’s gradient optimization. We set α=0.015\alpha=0.015 because it consistently enables CURB to reach stable (as shown in Figure. 3), near-optimal performance within only 3–4 optimization steps across different models and tasks.

Dear Reviewer reb4,

We would be grateful if you could review our rebuttal at your convenience. We have provided additional clarifications, experiments and revised the manuscript accordingly, with the changes highlighted in yellow. we would be grateful if you could let us know whether our response and revisions adequately address your concerns, and we are happy to provide additional clarification and experiments for further questions.

Thank you for your time and consideration.

Reviewer_reb4OpenReview ↗

Response

Thanks for your response. Sorry for the missing bibliographic information, which is [1] SMP-Attack: Boosting the Transferability of Feature Importance-based Adversarial Attack with Semantics-aware Multi-granularity Patchout [2] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, Ser-Nam Lim

Reviewer_zedaOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a test-time defense based on their analysis of how adversarial perturbations affect the internal representation within the ViT architecture. In particular, they find that residual trajectories in the earlier layers differ for adversarial samples when compared to that of real samples, with deviations being more amplified by the MLP branches inside the later transformer blocks. Based on this, they propose a method which will be Curvature-based Universal Residual Bending (CURB) that, at inference, optimizes the input when the curvature is below a threshold, such that residual trajectories are more curved at the earlier blocks, a property which they find is differing real samples from adversarial samples. They have demonstrated the method using various Vision language models, such as CLIP, LLaVA-1.5-7B and InternVL3.5-8B and show consistent gains.

Critical strengths
  1. The paper is well written and motivated.

  2. Insightful mechanistic finding: The findings regarding how adversarial samples impact the model from within are interesting. They find that in the deep layers, MLP branches mostly propagate the error growth compared to the attention branches. Based on their findings, they calculate residual curvature in the early layers and use it as a metric to optimize and reduce propagated errors.

  3. The theoretical part is adequate and shows that optimizing early residual curvature helps curb downstream error propagation when using existing attacks.

  4. The method is training-free, light, and has no dependence on closed-set labels or augmentations, and can be used across various tasks from classification to generation.

Critical weaknesses

Inadequate adaptive attacks:

  1. While the theory does show that increasing the curvature in early layers suppresses the error growth in the later layers, it does not prove that adversarial attacks must have low curvature. An adversary could create an adversarial example that has higher curvature. This gives rise to two scenarios:

    a) The default setting in paper uses a hard threshold to selectively activate CURB (shown in eq (14)). An adversary could exploit this detection mechanism and could use it to generate adversarial samples while staying above the threshold.

    b) The authors leave the vulnerability of the middle layers unexplored. For example, would it be possible to have attacks with higher curvature in the beginning layers, but then the middle layers may flatten the curvature and propagate the error.

  2. The results with BPDA (Table 4) only use 200 samples, and results vary significantly from what has been shown in Table 1(uses 500 samples), even for non-adaptive attacks. While I understand adaptive attacks can be expensive, in this case, the defense is really lightweight, so with this, it's not clear why the authors did not use 500 samples as they did in Table 1.

Questions for authors
  1. Could an adversary generate an adversarial example that can bypass the gating threshold shown in eq (14).
  2. Would it be possible to have attacks with higher curvature in the beginning layers, but then the middle layers may flatten the curvature and propagate the error?
4 comments

W1: Adaptive Attacks against Curvature Gating

We greatly appreciate you raising this exciting question. As reported in Sec. 4.4, “Ablation of Gating Threshold and Optimization Direction,” Table 5, we include a variant denoted as CURB ↑\uparrow, which removes the gate and optimizes all samples by increasing early residual curvature. The results show that the gate avoids unnecessary rectification of clean samples, but avoiding it remains marginal effect on clean performance compared with advanced methods such as ET3. This confirms that upward curvature optimization remains effective independently of selective gating. Based on this existing variant, we conduct the following two adaptive adversarial evaluations.

W1.1: Adaptive Attack That Bypasses the C2C_2 Gate

The first-new attack jointly maximizes the original attack objective and forces the adversarial curvature above the gating threshold:

\max_{\left\\|x-x_c\right\\|_\infty\leq\epsilon} \mathcal{J}_{\mathrm{task}}(x) - \lambda_g \left[\tau-C_2(x)\right]_+^2

where [z]+=max⁡(z,0)[z]_+=\max(z,0), λg=105\lambda_g=10^5, and τ\tau is the gating threshold. Thus, the attack is optimized to remain adversarial while satisfying C2(xadv)≥τC_2(x_{\mathrm{adv}})\geq\tau.

DatasetC2(xadv)≥τC_2(x_{\mathrm{adv}})\geq\taubaselineCURB$^{\uparrow}$
COCO498/500 (99.6%)3.05109.55
Flickr30k496/500 (99.2%)1.0667.73
TextVQA477/500 (95.4%)0.0029.48
VQAv2497/500 (99.4%)0.7264.86

W1.2: Adaptive Attack With High Early but Low Middle-Layer Curvature

We next requires the second-new attack to remain above the early C2C_2 threshold while reducing its curvature at C8C_8 relative to the paired clean input:

\max_{\left\\|x-x_c\right\\|_\infty\leq\epsilon} \mathcal{J}_{\mathrm{task}}(x) - \lambda_g \left[\tau-C_2(x)\right]_+^2 - \lambda_d \left[C_8(x)-C_8(x_c)\right]_+^2

where λg=λd=105\lambda_g=\lambda_d=10^5. This objective enforces both C2(xadv)≥τC_2(x_{\mathrm{adv}})\geq\tau and C8(xadv)≤C8(xclean)C_8(x_{\mathrm{adv}})\leq C_8(x_{clean}).

DatasetC8adv≤C8cleanC_8^{\mathrm{adv}}\leq C_8^{\mathrm{clean}}C2(xadv)≥τC_2(x_{\mathrm{adv}})\geq\tau & C8adv≤C8cleanC_8^{\mathrm{adv}}\leq C_8^{\mathrm{clean}}baselineCURB$^{\uparrow}$
COCO499/500 (99.8%)496/500 (99.2%)3.22107.41
Flickr30k500/500 (100%)498/500 (99.6%)1.0567.65
TextVQA500/500 (100%)476/500 (95.2%)0.1228.62
VQAv2500/500 (100%)498/500 (99.6%)0.7063.42

This indicates that deliberately flattening the middle-layer curvature does not invalidate the original early-curvature rectification mechanism. More importantly, the rectification point DOES NOT need to be moved to C8C_8.

These two new results agrees with proposed Proposition 3: increasing early residual curvature strengthens the differential component, suppresses downstream discrepancy amplification, and reduces the final feature deviation. We will add both adaptive experiments and more detailed analysis to the final paper.

W2: 200-Sample in Table 4

Table 4 uses 200 samples to follow the same evaluation protocol as the compared baselines, including ET3, TeCoA, and FARE, which also use 200 samples. Because this adaptive attack is expensive, reproducing all baselines on 500 samples is difficult within the rebuttal period. We therefore provide an additional 500-sample evaluation of CURB using the standard CLIP encoder.

DatasetAdv.Adv.*
COCO33.630.9
Flickr30k31.730.2
TextVQA18.517.0
VQAv226.223.1

Thank you for the additional experiment and the rebuttal. However, I believe W1.1 exposes a fundamental weakness in the gating mechanism rather than resolving my concern.

The adaptive attack is explicitly designed to raise the adversarial sample’s curvature above the threshold. If it succeeds, the gate in Eq. (14) should skip the rectification and pass the adversarial input directly to the model. The response instead appears to rely on CURB↑, which removes the gate and optimizes every sample. This may demonstrate that the ungated optimization remains effective, but it does not establish the robustness of the original gated method.

More importantly, if removing the gate is necessary to defend against samples that deliberately cross the threshold, then the gate itself creates an avoidable attack surface. An adaptive attacker can target the curvature threshold specifically to prevent the defense from being activated. I therefore believe the gating mechanism is a significant flaw in the current method and should be removed, with CURB↑ used as the primary method rather than merely as an ablation. Otherwise, the authors need to demonstrate that the original gated CURB remains effective specifically on samples that successfully cross the threshold.

AuthorsOpenReview ↗

Core Mechanism and the Auxiliary Role of Gating

Thank you for identifying this limitation. We agree that a gate-aware adversary can force C2(xadv)≥τC_2(x_{\mathrm{adv}})\geq\tau, causing the original gated configuration to skip rectification. Therefore, the deterministic gate itself should not be interpreted as secure against an adversary that explicitly targets its threshold.

We would, however, like to clarify that our core contribution is the mechanistic exploration in Section 2, Rethinking Test-Time Defense for LVLMs: adversarial perturbations suppress directional differentiation between early residual updates, and thereby induce discrepancy that is preferentially amplified by downstream branches. We further show theoretically and experimentally that increasing early residual curvature suppresses the subsequent propagation of adversarial discrepancy.

The second-new adaptive attack provides stronger evidence.It deliberately postpones the curvature abnormality to the later layer. Nevertheless, increasing early residual differentiation still improves adversarial robustness from 1.271.27 to 66.7866.78. Thus, even when residual flattening is deliberately delayed, increasing directional differentiation early remains sufficient to suppress its downstream propagation. This result provides additional mechanistic evidence for our central claim: early residual geometry is not merely a signal for detecting adversarial inputs, but an effective intervention point for curbing the subsequent adversarial discrepancy. It is also consistent with Proposition 3, which predicts that increasing the early differential component reduces the discrepancy subsequently accumulated through deep MLP branches.

CURB is introduced as a concrete operational tool for validating this mechanism. The gating threshold is not part of this mechanistic claim; it is auxiliary to be added only to avoid unnecessary rectification of clean inputs (even just small impact on them).

We know that the current presentation may not distinguish these two levels clearly enough. We will therefore revise the Abstract, Introduction, Proposed Method, and other related sections to state explicitly that (i) CURB is the mechanism-instantiating operation used to validate our central early-curvature rectification, whereas (ii) hard gating is only an external clean-preservation tool and is limited to a curvature-aware adversary.

Thank you for your time and consideration. We have revised the presentation and corresponding content in the updated manuscript, with all changes highlighted in red.

AuthorsOpenReview ↗

Strongly Constraining Early Curvature Does Not Alter the Mechanism of Adversarial Propagation

We further examine whether forcing early curvature higher (or stated as above τ\tau) restores residual dynamics (or stated as invalidating our central mechanism). For each actively perturbed clean–adversarial pair, we compute $ \Delta C_k = C_k(x_{\mathrm{adv}})-C_k(x_{\mathrm{clean}}) $ ( Refer to the Figure 5 in revised PDF).

DatasetAttackSuccessful AttackedMean ΔC2\Delta C_2ΔC2<0\Delta C_2<0Mean ΔC5\Delta C_5ΔC5<0\Delta C_5<0
COCOStandard APGD494-0.137100.0%-0.07799.2%
Gate-Bypass493-0.06994.3%-0.06097.4%
Flickr30kStandard APGD495-0.132100.0%-0.07099.6%
Gate-Bypass491-0.06193.5%-0.05298.0%
TextVQAStandard APGD202-0.128100.0%-0.06798.0%
Gate-Bypass202-0.06691.1%-0.05395.0%
VQAv2Standard APGD403-0.141100.0%-0.07899.5%
Gate-Bypass399-0.07396.0%-0.06198.5%

These paired results reveal an important distinction between crossing the gate threshold and restoring normal residual geometry. The Gate-Bypass attack (the first-new adaptive attack) successfully raises C2(xadv)C_2(x_{\mathrm{adv}}) above τ\tau, but its curvature remains lower than that of the paired clean input for 94.1% of samples. More importantly, this abnormal flattening persists at C5C_5: the mean ΔC5\Delta C_5 remains negative on every dataset, and 97.5% of gate-bypassing samples still satisfy C5(xadv)<C5(xclean)C_5(x_{\mathrm{adv}})<C_5(x_{\mathrm{clean}}). Although the attack reduces the pooled C2C_2 deficit from -0.135 to -0.067, the C5C_5 deficit remains substantial at -0.057. Therefore, forcing an input across the scalar C2C_2 threshold does not reconstruct its clean residual trajectory. It primarily conceals the abnormality from the hard gate, while the suppressed directional differentiation remains observable downstream.

This experiment does not imply that the hard gate itself cannot be targeted. Rather, it shows that such targeting does not invalidate our central finding: adversarial perturbations continue to suppress residual directional differentiation even after their early curvature has been explicitly raised. Thus, your concern identifies a limitation of the selective hard gate, but not a failure of the rectification mechanism that constitutes our central contribution.

We are happy to provide additional clarification and experiments for further questions. We are very appercriate you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.