acceptodds
Under review as a conference paper at ICLR 2027

MAAT: Massive-Activation-Aware Transferable Adversarial Attacks

Abstract

Transfer-based adversarial attacks optimise a perturbation on a surrogate model and apply it to target models without access to them. Such attacks can include objectives defined on intermediate representations, whose values depend on how representation magnitude is distributed across coordinates. We ask whether this distribution can be exploited to improve transfer. In DINOv2-L/14 and CLIP-L/14, the representation block used by the attack contains a token whose norm exceeds that of a typical token by and times, respectively, with the squared magnitude of that token concentrated in an effective dimensionality of only and channels out of . A clean-data calibration identifies the corresponding token–channel coordinates in -% of image–channel pairs. We do not assume that these coordinates are intrinsically non-transferable. Instead, MAAT uses this structure to control which channels enter the feature objective, while leaving the task loss and update rule unchanged. This makes the intervention compatible with attacks that modify the gradient, forward pass, backward pass, or input. Across two surrogates, eight targets, and four base attacks, MAAT increases mean transfer attack success rate by about four percentage points. Matched random-channel controls with six seeds support the channel selection effect for CLIP-L/14 on all seven transfer targets; the same evidence is absent for DINOv2-L/14, where an ablation identifies the source of the difference. These results show that intermediate-activation structure can be used to shape transferable adversarial objectives, with the effect depending on the surrogate.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial transferability, visual foundation models, massive activations, black-box attacks, vision transformers

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_H9p2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to leverage the activation magnitude distribution of intermediate representations in Vision Transformers to improve the cross-model transferability of adversarial attacks. The authors empirically observe that certain tokens exhibit unusually large feature norms, with their activations largely concentrated in a small subset of channels. Motivated by this observation, they propose MAAT to enhance adversarial transferability by calibrating highly activated channels, suppressing feature variations within highly activated channels and maximizing feature discrepancies across the remaining channels. Experiments show that integrating MAAT with existing transfer-based attacks can further improve their attack performance.

Critical strengths
  1. Different from prior studies on transferable adversarial attacks, this work explores how to exploit the concentration phenomenon of activation magnitudes in intermediate representations to improve attack transferability, and designs the attack objective from the perspective of the feature activation magnitude distribution.
  2. Superior attack performance and method compatibility. MAAT can be readily integrated with existing transfer attacks, including MI-FGSM, MTE, PNA, and PatchOut, consistently improving their average attack success rate across different target models and datasets.
Critical weaknesses
  1. Although the authors investigate the concentration of activations in intermediate representations of Vision Transformers, they do not sufficiently justify why this phenomenon warrants further investigation in the context of transferable adversarial attacks, nor do they clearly articulate how it relates to the limitations of existing attack methods.
  2. Separate ablations on suppressing variations in highly activated channels and maximizing discrepancies across the remaining channels are missing, making their individual contributions to attack performance unclear.
Questions for authors
  1. Could the authors further clarify why the concentration of activation magnitudes in intermediate representations deserves particular attention in the context of transferable adversarial attacks? Specifically, could exploiting this phenomenon help overcome certain limitations or technical challenges of existing methods, or provide new insights into improving adversarial transferability?
  2. Since the experiments only use DINOv2-L/14 and CLIP-L/14 as surrogate models, additional evaluations across different model scales (e.g. DINOv2-B/14, CLIP-B/14) would help verify the effectiveness and generality of MAAT.
  3. Existing work [1] shows that introducing register tokens mitigate the phenomenon of abnormally high-norm tokens in Vision Transformers. Could the authors evaluate MAAT on DINOv2 with registers, and compare the degree of activation concentration as well as attack performance before and after register incorporation, to examine whether MAAT remains effective when the feature structure of high-norm tokens changes?

[1] Darcet, Timothée, et al. "Vision transformers need registers." International conference on learning representations. Vol. 2024. 2024.

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

This paper investigates massive activations in the intermediate representations of Vision Transformers and proposes MAAT to improve adversarial transferability. The method identifies high-activation channels through clean-data calibration, then encourages feature changes outside these channels while penalizing changes at their peak-token locations. Experiments on DINOv2 and CLIP show improvements over several baseline attacks, with random-channel controls providing stronger evidence for CLIP.

Critical strengths

S1. The paper provides a quantitative analysis of high-norm activations in intermediate ViT representations and introduces a simple channel-selection strategy;

S2. The experiments consider different base attacks and include matched random-channel controls, which provide useful evidence for the effectiveness of the proposed channel selection, particularly on CLIP.

Critical weaknesses

W1. The motivation and novelty are not sufficiently convincing. Sections 3 and 4 mainly characterize high-norm tokens and channel concentration, but their relationship with existing studies on register-related artifacts remains unclear. Vision Transformers Don't Need Trained Registers (Jiang et al.) has already investigated the sparse MLP neurons responsible for high-norm tokens in CLIP and DINOv2. The manuscript neither establishes a clear distinction from these findings nor explains why massive-activation channels should be excluded and penalized to improve adversarial transferability. Moreover, related feature importance-aware attacks, such as FIA (ICCV 2021), NAA (CVPR 2022), and APFD (ECAI 2025), are insufficiently discussed. Without a clear mechanism or convincing comparisons, MAAT appears more like an empirically designed feature-loss modification than a well-motivated novel approach.

W2. The experimental validation is insufficient to support the claimed contribution and generality. The main experiments involve only two surrogate models, both exhibiting massive activations. There is no evaluation on models with register tokens or architectures where such activations are reduced. Moreover, the contributions of channel exclusion and the additional penalty are not separately evaluated, and the sensitivity to key hyperparameters is unclear. The controlled improvements are also strongly dependent on the surrogate and perturbation budget: at ε = 4/255, no target shows a statistically significant gain over random-channel controls. These limitations make it difficult to determine the actual contribution and applicability of MAAT.

Questions for authors

Q1. The authors cited Vision Transformers Need Registers, which already investigated high-norm artifact tokens in ViTs. More importantly, Jiang et al. further identified sparse MLP neurons responsible for these outliers in CLIP and DINOv2. How do the massive-activation channels analyzed in Sections 3 and 4 differ from these known phenomena? What additional insights are provided by the present analysis?

Q2. Why should massive-activation channels be excluded from the feature objective and their changes further penalized? Are these channels less transferable or more surrogate-specific? The paper provides no direct evidence supporting this motivation. Besides, Eq. (5) normalizes the feature gradient, so the influence of the feature-scale denominator in Eq. (2) needs clarification.

Q3. The ablation experiments are incomplete. What are the respective contributions of channel exclusion and the additional penalty in Eq. (4)? Comparisons with the full-channel feature objective, exclusion only, and penalty only are necessary. How were the hyperparameters (k=5), (q=0.05), (\lambda=0.5), and the selected feature layers determined? Their influence on attack performance should be investigated.

Q4. The experiments only use DINOv2-L/14 and CLIP-L/14 as surrogates. Why not consider DINOv2 with registers and DINOv3, which already incorporates register tokens? It is important to determine whether MAAT remains effective when high-norm artifacts are mitigated, and whether stronger models without such artifacts could already achieve better transferability. In addition, comparisons with feature importance-aware attacks such as FIA, NAA, and APFD under consistent settings are missing.

Q5. The E-commerce experiments described in Sec. 6.1 are missing from the results. The authors should report the corresponding transfer attack performance rather than only the activation statistics in Appendix A.

Q6. There is an inconsistency in the calibration protocol. Sec. 5 explicitly states that the first 500 attacked images are used for calibration, whereas Sec. 6.1 claims that calibration and evaluation images are disjoint. Which setting was actually adopted? Furthermore, under MTE, the clean and adversarial representations are computed through different forward procedures. The authors acknowledge this issue in Sec. 5 but do not provide the matched-forward control, making it difficult to isolate the contribution of the proposed feature objective.

Q7. The statistical testing protocol in Appendix E is also questionable. The authors take the median of the McNemar-test p-values across six random-channel seeds and then apply Benjamini–Hochberg correction. How is this aggregation justified as a valid combined significance test? The reported statistical conclusions require a more rigorous explanation.

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

This paper investigates whether massive-activation structure in vision transformers can be used to improve the transferability of adversarial attacks. The authors characterize activation concentration in intermediate representations of DINOv2-L/14 and CLIP-L/14, and use a low-quantile statistic computed on clean images to identify a small set of channels with consistently large activations. The proposed method, MAAT, encourages feature changes outside the selected channels while penalizing deviations from the clean reference at their peak token–channel coordinates. The feature and task gradients are normalized separately and then combined. The paper reports improvements over several base attacks and uses matched random-channel controls to assess the contribution of the calibrated channel selection. The reported benefits depend on the surrogate model, perturbation budget, and base attack.

Critical strengths

1、The paper formulates a concrete, testable intervention and reports meaningful improvements in transfer performance. MAAT translates a measurable property of intermediate representations into an explicit feature objective. In the main ImageNet experiments with the MTE base attack at (epsilon=16/255 ), mean transfer ASR increases from 57.0% to 61.2% for the DINOv2 surrogate and from 47.5% to 51.6% for the CLIP surrogate. These results support the potential practical value of the proposed intervention. 2、The evaluation attempts to distinguish the benefit of adding a feature objective from the benefit of selecting particular channels. Matched random-channel controls with six seeds provide a more informative comparison than evaluation against the base attack alone. The paper also reports limitations, including weaker evidence at smaller perturbation budgets and negative results when MAAT is combined with PNA–PatchOut. These observations help delineate the method’s scope.

Critical weaknesses

1、Inconsistencies in the experimental protocol and incomplete result reporting limit the verifiability of the central claims. Section 5, lines 223–226, explicitly states that calibration uses the first 500 images subsequently attacked and is not held out. In contrast, Section 6.1, line 317, states that the calibration and attack evaluation sets are disjoint. Section 6.5 refers to Appendix I for the controlled MI-FGSM comparison and controls for other bases, but Table 6 contains no MI-FGSM results and explicitly states “Single runs, no control seeds.” In addition, I could not locate E-commerce attack results supporting the conclusion’s claim of evaluation across three datasets. Appendix E also leaves the construction of “median-seed p-values” ambiguous. Together, these issues make it difficult to determine which protocols and statistical evidence support the reported conclusions. The calibration inconsistency does not, by itself, establish label leakage, but it must be resolved. 2、The specific contribution relative to prior work and the necessity of the individual design choices are not yet sufficiently established. Section 4.8 of Robustness Tokens(ECCV 2024) already examines the relationship between massive activations and adversarial attacks, making the broad statement in Section 2 that this relationship remains unstudied too strong. FDA, FIA and NAA also provide relevant precedents for selectively manipulating intermediate features. MAAT’s particular calibration rule and objective may still constitute a contribution, but more focused comparisons and component ablations are needed to establish it. Furthermore, the main MTE configuration computes the clean reference and adversarial representation using different forward procedures, and the paper acknowledges that the matched-reference control has not been run. The current evidence therefore does not sufficiently isolate the contributions of channel selection, the peak-coordinate preservation term, and interaction with the modified forward pass.

Questions for authors

1、Are the calibration and attack evaluation sets overlapping or disjoint? Which protocol was used for each reported table? Please specify the data split. If the current results use overlapping sets, do the main improvements persist with a disjoint calibration set? 2、How exactly are “median-seed p-values” constructed? Please provide the per-target, per-seed results, the discordant-pair counts underlying the McNemar tests, and the adjusted p-values for the main controlled comparison. Where are the controlled MI-FGSM results, the comparison with forward modification applied to all blocks, and the E-commerce attack results referenced in the manuscript? 3、Control needed to assess the main result: When the clean reference and adversarial representation are computed using the same forward procedure, do the improvements over both the base attack and the random-channel controls persist? This result would directly affect the interpretation of the main MTE experiments. 4、Component analysis: How do the complement-only feature term, the peak-coordinate preservation term alone, and the complete objective compare? Given the separate gradient normalization in Equation (5), how do the authors distinguish the effect of channel selection from changes in the relative weighting of the two feature terms? 5、Positioning and comparative evidence: What is the central contribution of MAAT relative to Robustness Tokens, FDA, FIA/NAA, and activation-norm-based gradient modulation such as TESSER? Can the authors provide a representative feature-level attack comparison under matched surrogate models and perturbation budgets, with computational costs accounted for?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.