acceptodds
Under review as a conference paper at ICLR 2027

Read the Trajectory, Not Just the Endpoint: Calibrated Utility Readout for Transfer Attacks

Abstract

Iterative transfer attacks are widely used to evaluate adversarial vulnerability on unseen models. Although their optimization trajectories contain multiple legal adversarial actions, existing methods almost universally deploy the final iterate. However, source optimization and cross-model transferability need not peak at the same iteration, and the preferred action can vary across images. Replacing the endpoint with one global checkpoint is therefore insufficient, while target-guided action selection would violate the zero-query transfer setting. To address these problems, we propose Calibrated Pointwise Relative Utility (CPRU), a target-output-free trajectory readout that estimates the signed utility of each saved action relative to the conventional final iterate. We show that the target-clean mean of this utility is exactly equal to the paired attack success rate (ASR) gain. CPRU uses four target-disjoint proxy models and an attack-specific utility regressor to score all non-final actions, independently returning the highest predicted positive-utility action while retaining final as an exact zero-utility fallback. Extensive experiments across different iterative attacks, data sets, source models, and target architectures demonstrate the effectiveness of image-specific trajectory readout.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Calibrated pointwise relative utility, target-output- free trajectory readout, cross-model transferability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_npNCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CPRU (calibration point relative utility), which is a trajectory reading method for transferable confrontational attacks. CPRU does not always deploy the final iteration, but uses target-to-target unrelated proxy models and attack specific utility readers to evaluate the saved intermediate confrontation state, and selects image-specific operations related to regular endpoints. This method retains the original attack generation process, and the official target is still inaccessible before deployment. Experiments on CIFAR-10 and ImageNet-1K in multiple attack and target architectures show that trajectory reading can improve the success rate of transfer attacks.

Critical strengths
  1. The paper identifies image-specific trajectory selection as an underexplored degree of freedom in transfer attacks.
  2. Source, proxy, development, and formal-target models are explicitly separated, reducing the risk of formal-target leakage.
  3. The paper evaluates multiple attacks, targets, source models, two datasets, uncertainty intervals, and several controlled ablations.
Critical weaknesses
  1. There is a mismatch between the theoretical utility formulation and the implemented reader. The theory defines μt(z)=E[yi,t∣Zi=z]\mu_t(z)=\mathbb{E}[y_{i,t}\mid Z_i=z], whereas the ExtraTrees regressor is trained with highly asymmetric 31/3/1 sample weights. Therefore, the learned score is generally not an unbiased estimate of the conditional utility used in the theoretical analysis, which weakens the interpretation of the zero-utility threshold and the fitted-reader bound. 2.The method relies on substantial auxiliary resources. CPRU requires four proxy models, five development models, a labeled development population, and 44 additional proxy forward passes per image, yet there is no strong resource-matched comparison against simpler proxy-based selection or ensemble strategies.
  2. The calibration claim is insufficiently validated. Although the method is described as a calibrated utility readout, the paper does not provide direct calibration, utility-prediction, or sign-accuracy analysis to verify that predicted scores correspond reliably to actual relative utility.
  3. Some experimental comparisons and protocol choices need stronger justification. The LPAA comparison contrasts a trajectory-readout method with a different attack-generation method, while the reported runtime excludes the offline reader-development cost. In addition, the ImageNet reader is trained on a subset of the validation set rather than on a fully separate development split.
  4. Several implementation details remain insufficiently justified. Important design choices such as the 31/3/1 weighting, 180 trees, minimum leaf size, and the exact BP14 definition lack adequate motivation or sensitivity analysis.
Questions for authors
  1. How do the asymmetric 31/3/1 sample weights affect the interpretation of the reader output as μt(z)=E[yi,t∣Zi=z]\mu_t(z)=\mathbb{E}[y_{i,t}\mid Z_i=z]? Does the theoretical analysis still apply to the weighted regressor?
  2. Were the reader hyperparameters, including the 31/3/1 weighting, 180 trees, and minimum leaf size, fixed using only the development set before any formal-target evaluation?
  3. In the ConvNeXt-Small source-transfer experiment, is the reader reused without retraining, or is it refitted using trajectories generated from the new source model?
  4. How does CPRU compare against simpler resource-matched proxy-based trajectory selection rules, such as selecting the step with the highest mean proxy margin or proxy success rate?
4 comments

We thank the reviewer for recognizing image-specific selection, model-role separation, and the breadth of the evaluation.

1. Weighting and theoretical alignment (Q1).

We agree that sign-dependent weighting changes the squared-loss estimand. We replaced 31/3/1 with 1:1:1: ordinary MSE now targets the conditional signed utility in Eq. 16 (Eq. 29), retaining the target-clean normalization in Eq. 14. Proposition 2 describes the ideal selector; the fitted-reader bound in Eqs. 21-22 explicitly depends on prediction error to that conditional utility. Gains remain +7.347 pp on CIFAR-10, while ImageNet improves from the original +2.280 to +2.693 pp (Tables 1, 4).

2. Parameter sensitivity and the origin of BP14 (Q2).

All reader hyperparameters were determined using only the development set. Varying trees over 90/180/360 and minimum leaf size over 8/16/32, one factor at a time, gives +7.137-7.519 pp on CIFAR-10, with only 0.382 pp ASR spread (Tables A.29-A.30). Gains therefore do not depend on the precise 180/16 setting. Tables A.2-A.3 document model roles, constructors and preprocessing.

BP14 arose from our early work on MDCS: we sought to make its scaling responsive to disagreement between the current normalized gradient qi,tq_{i,t} and its historical mean qˉi,t−1\bar q_{i,t-1}. We introduced the coordinate-wise ratio ri,tr_{i,t} (Eq. A.73) and adaptive exponent pi,t=1/4+ri,t/4p_{i,t}=1/4+r_{i,t}/4 (Eq. A.74) to couple perturbation and momentum scaling (Eq. A.76). Although this improved the baseline, fixed-endpoint deployment remained a key transfer bottleneck, motivating CPRU. We retained BP14 as one of five trajectory generators; Appendix A.4.3, Eqs. A.70-A.77, provides its complete definition.

3. The original source experiment refitted readers; frozen reuse also works (Q3).

The original ConvNeXt-Small experiment used source-specific development trajectories to test CPRU with a different source. We now also directly reuse frozen ViT-B/32 readers: +4.004 pp, retaining 94.0% of the dedicated readers' +4.260 pp (Table 3). This verifies direct reuse across these sources without separately retraining the reader.

4. Direct proxy selectors and CPRU utility analysis (Q4).

RelMargin and RelVote use the same trajectories, actions and four-proxy outputs (44 forwards per image and attack), without reader training or inference. RelMargin retains per-image intermediate-state selection, clean-logit normalization, final-relative margins and the zero-score Final fallback (Eqs. 24-26, 30). RelVote uses final-relative proxy-success votes; both retain Final unless the best score is positive. CIFAR-10 mean ASRs are 54.707% (Final), 60.049% (RelMargin), 57.822% (RelVote), and 62.054% (CPRU). The learned reader adds 2.005/4.232 pp over the two controls (Tables A.17-A.18). RelMargin is itself part of our contribution: a lightweight, reader-free variant of the CPRU trajectory-readout framework.

CPRU utility analysis. Recovery exceeds destruction: 9.330 vs 1.983 pp on CIFAR-10 and 4.726 vs 2.033 pp on ImageNet (Table A.31). Appendix A.9, Table A.33 compares mean predicted/observed signed utility on unseen formal targets: 6.054/7.347 pp on CIFAR-10 and 2.912/2.693 pp on ImageNet-1K. Predictions reflect positive aggregate gains and successfully capture the contrast between larger gains on CIFAR-10 and smaller gains on ImageNet.

5. LPAA: training-free deployment and comparison purpose.

  • Strong performance without reader training. RelMargin achieves 66.10% ASR versus LPAA's 63.28% (+2.82 pp), taking 106.010 vs 595.531 s for 1,000 images; LPAA takes 5.62 times as long (Table 2). Timing includes online generation/selection and one target's loading and clean/adversarial forwards.

  • No new reader training is required. Our reader-free CPRU variant retains strong ASR gains without fitting a reader. Together with the reuse evidence in Q3, this supports online-cost comparisons for deployments requiring no new reader training.

  • Comparison purpose. We agree LPAA changes attack generation, whereas CPRU performs post-hoc readout. Section 4.3 tests whether selecting intermediate states from an unchanged ten-step trajectory, with zero additional source gradients for readout, can match or outperform a strong gradient-enhancement attack. Implementation and complete budgets are in App. A.4.3 and Table A.5(b).

6. ImageNet development and evaluation are image-disjoint.

The classifiers were pretrained on ImageNet training images. We fit utility on held-out validation images to reflect attacks on images outside classifier training. We use 2,000 development images and a disjoint 8,000-image held-out pool, sampling 1,000 evaluation images from that pool (Sec. 4.1). Evaluation images are excluded from reader fitting and tuning; formal targets are not queried during reader development or action selection (App. A.3).

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer npNC ,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Deadline for Discussion is Approaching in Less than 2 Days

Dear reviewer npNC ,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in One Day]

Dear reviewer ,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only one day, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

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

The topic is interesting, and the results show that selecting an intermediate action can improve existing attacks. The separation of source, proxy, development, and formal-target models is also clearly considered. CPRU can be added to an attack without modifying its trajectory generation.

Critical strengths

This paper shows that selecting an intermediate action can improve existing attacks. The separation of source, proxy, development, and formal-target models is also clearly considered.

Critical weaknesses
  1. The core seems to be: evaluate the relative utility of different attack checkpoints, train an ExtraTrees regressor to predict this utility, and use it to select a checkpoint. But why do different target models prefer different checkpoints of the same attack? Why does the preferred checkpoint also change from one sample to another? These are the more interesting questions, and the paper does not give a convincing explanation. The propositions do not really fill this gap. They mainly formalize the utility objective and the benefit of an ideal selector, while the actual reason behind the trajectory differences remains unclear. In the end, the paper observes these differences and fits a predictor to exploit them. I do not think this alone provides enough insight.

  2. The gain also looks rather small for the extra machinery involved. The default setting evaluates saved attack states using four proxy models, requiring 44 additional proxy forwards per image and attack, and also needs a separately trained utility reader. After all this, the mean improvement on ImageNet-1K is only about 2.28 percentage points. This cost–benefit trade-off is not convincing to me, especially without a stronger comparison using the same total computation on ImageNet-1K.

  3. If the ExtraTrees reader makes a wrong prediction, it can directly make the attack worse. It may discard a final iterate that already succeeds and replace it with an intermediate one that fails. Keeping the final iterate as a zero-utility fallback does not solve this problem when the reader wrongly predicts a positive utility. However, the paper does not properly analyze these failure cases. How often does the selector destroy an existing success? On which samples and target models does this happen, and why? The whole selection strategy depends on the reader, but its failure behavior is left largely unexplained. Reporting a positive average ASR gain is not enough to establish that this extra selection step is reliable.

Questions for authors

See Weaknesses.

4 comments

We thank Reviewer SeYB for recognizing the value of intermediate-state selection and the separation of model roles. We added mechanism, runtime, and failure diagnostics.

W1. Margin geometry explains variation across images and targets.

Our contribution is image-specific checkpoint selection without target queries. Section 3.5 and Appendix A.7 explain trajectory variation. Let Mq,iM_{q,i} be target model qq's correct-class margin (true-class minus largest incorrect-class logit); M<0M<0 means attack success. For the actual source update vi,t=xi,t+1−xi,tv_{i,t}=x_{i,t+1}-x_{i,t},

aq,i,t=∇xMq,i(xi,t)⊤vi,t.a_{q,i,t}=\nabla_x M_{q,i}(x_{i,t})^\top v_{i,t}.

This inner product of the target-margin gradient and source update approximates the margin change (Eq. 33): a<0a<0 favors attack success, while a>0a>0 works against it.

  • Across images: Among clean-correct image-target pairs with any successful step, lower-initial-margin groups reach first success earlier for all five attacks (Table A.13; PGD-A15 low/high means: 2.603/5.159 steps). For 374-518 of 1,000 images per attack, an intermediate state succeeds on a target where Final fails (Table A.12). At last success-to-failure transitions, 93.26% of updates have a>0a>0, versus 45.89% in matched stable-success controls (Table A.14).
  • Across targets: The same update yields opposite inner-product signs across targets in 2,775/4,365 comparable updates (63.57%); disagreement remains 52.84% among targets already fooled at that step (Table A.15). An update can favor an attack on one target while hindering it on another.
  • Connection to CPRU: Post-hoc diagnostics around selected checkpoints show that the fraction of updates with a>0a>0, averaged across attacks, rises from 25.16% for the incoming update to 58.62% for the outgoing update (Fig. 2, Table A.16), linking CPRU's choices to checkpoints preceding locally counterproductive updates.

W2. Gains persist at approximately matched online runtime.

ImageNet-1K gains +2.693 pp (95% CI [2.246, 3.141]); all five attack-wise intervals exclude zero (Table 4).

  • Matched budget: Ten-step CPRU, including trajectory generation, 44 proxy forwards, and reader inference, costs 14.4-16.1 plain attack steps (mean 15.0). It achieves 52.877% versus Final15's 49.297% ASR, a paired gain of +3.579 pp, 95% CI [2.997, 4.164] (Sec. 4.5, Tables A.10-A.11).
  • Longer trajectories: At 18 steps, CPRU achieves 52.650% versus Final18's 48.342%. Its gain expands from +2.693 to +4.308 pp (Table A.9).
  • Lower overhead: Two proxies halve additional forwards to 22 and retain 87.1% of the ImageNet gain (+2.346 pp) (Table 4). One proxy gains +5.024 pp on CIFAR-10 (Table A.19). On the four-attack CIFAR-10 benchmark, CPRU/reader-free RelMargin/LPAA achieve 68.06%/66.10%/63.28% ASR, with mean runtimes of 109.056/106.010/595.531 s. LPAA takes 5.62 times as long as RelMargin. Timing covers online generation/selection and one target's loading and clean/adversarial forward evaluation (Table 2; full breakdown: Table A.5(b)).
  • Reader reuse: Frozen ViT-B/32 readers applied to ConvNeXt-Small trajectories gain +4.004 pp, retaining 94.0% of dedicated readers' +4.260 pp (Table 3), demonstrating reuse across these sources without retraining.

W3. Failures vary across targets and checkpoint positions.

The zero-score Final fallback does not guarantee preservation of an existing success. Recovery (R) means Final fails but selection succeeds; destruction (D) means the reverse. Table A.34 reports every target (R/D in pp, averaged over five attacks):

Target (CIFAR / ImageNet)CIFAR-10 R/DImageNet-1K R/D
DeiT-Tiny / DeiT-Small9.767 / 1.3744.916 / 1.742
CoAtNet-07.265 / 2.8284.623 / 2.299
CaiT-XXS249.990 / 1.3014.739 / 1.836
NesT-Tiny11.158 / 1.9585.171 / 1.960
PiT-Ti8.470 / 2.4534.179 / 2.328
Mean9.330 / 1.9834.726 / 2.033

Recovery outweighs destruction; mean gains are +7.347/+2.693 pp. Net gains are positive in all 25 attack-target combinations per dataset (Tables A.4, A.8).

Where losses occur: CoAtNet-0 has the highest mean attack-wise conditional destruction rate among targets: 7.665%/6.295% on CIFAR/ImageNet. Specific combinations include PI-FGSM/CoAtNet-0 on CIFAR (48/428 Final successes lost, 11.215%) and PI-FGSM/NesT-Tiny on ImageNet (26/281, 9.253%) (App. A.9, Table A.34 and following paragraph, p. 46).

Early selection: The selected checkpoint precedes the affected target's first success in 344/474 image-attack-target destruction events (72.57%) on CIFAR-10 and 258/401 (64.34%) on ImageNet (App. A.9, Table A.35, p. 47). For CoAtNet-0: 103/137 (75.18%) and 65/93 (69.89%), respectively (p. 46). This identifies premature checkpoint choice as a major failure mode. A reader-preferred state can precede the affected target's boundary crossing, even though later states succeed.

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer SeYB,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Deadline for Discussion is Approaching in Less than 2 Days

Dear reviewer SeYB,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in One Day]

Dear reviewer ,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only one day, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Reviewer_NG6kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Calibrated Pointwise Relative Utility (CPRU), which uses four target-disjoint proxy models to describe each saved trajectory state and an ExtraTrees regressor trained using separate development models to predict whether replacing the final iterate by a particular intermediate state will increase target-clean attack success. The predicted best positive-utility state is selected independently for each image, with the final iterate assigned exact utility zero and retained as a fallback. On the primary CIFAR-10 experiment, CPRU improves equal-target ASR by 7.617 percentage points on average over the final iterate and 7.262 points over a dev-selected global checkpoint; gains are smaller on ImageNet-1K, where the mean improvement is 2.280 points.

Critical strengths

Clean problem formulation with strong empirical evidence. The paper identifies per-image trajectory-state selection as a meaningful degree of freedom that is distinct from changing the attack optimizer or choosing one globally fixed early stopping point. The CIFAR-10 results are strong across all five attacks with gains of roughly 4.4-9.5 percentage points, and the improvements persist across all 25 attack-target pairs.

The target-access protocol is carefully designed. Source, proxy, development-utility, and formal-target checkpoints are assigned disjoint roles, and the formal target is used only after the action has been frozen.

Critical weaknesses

The implemented weighted regression objective does not appear to estimate the conditional utility used in the theoretical analysis. The theory analyzes the unweighted conditional utility [ \mu_t(z)=\mathbb{E}[Y_{i,t}\mid Z_i=z], ] for which a positive value means expected recovery exceeds expected destruction. However, the implemented ExtraTrees reader weights positive, negative, and zero targets by 31, 3, and 1. Under weighted squared-error regression, the model estimates a weighted conditional mean rather than (\mathbb{E}[Y\mid Z]), which can shift the zero threshold. For (Y\in{-1,0,1}), the weighted prediction can be positive when (31P(Y=1\mid Z)>3P(Y=-1\mid Z)), whereas the ASR-aligned utility is positive only when (P(Y=1\mid Z)>P(Y=-1\mid Z)). Thus, the connection between the theoretical utility and the implemented zero-threshold decision needs clarification.

The experiments do not isolate learned signed-utility prediction from substantially simpler pointwise uses of the same proxy models. CPRU uses four proxy models and 44 proxy forwards per image, while Final and Dev-fixed do not receive comparable per-image proxy information. It is therefore unclear how much of the gain comes from the learned signed utility reader versus simply selecting states using final relative proxy statistics. The ablations show that removing final-relative features sharply reduces the gain, while aggregate-only, mean-only, and binary-utility variants remain close to the full method.

Questions for authors

How should the 31/3/1 target-dependent weighting be reconciled with Eq. 16 and Proposition 2? The experiments establish that proxy-informed pointwise trajectory readout is useful. They currently provide substantially weaker evidence that CPRU's calibrated signed-utility regression is necessary or is the main technical source of the gains. This distinction affects both novelty and the interpretation of the mechanism.

Most CIFAR checkpoints are specified only by architecture family rather than exact checkpoint identifier/training recipe, and BP14 is described but does not include enough reproducibility details. How do simple resource-matched proxy selectors perform? For example, what happens if one chooses the trajectory state with the largest mean final-relative proxy margin, or uses final-relative proxy-success voting, while retaining final as the fallback?

In the ConvNeXt-Small source experiment, are the original ViT-B/32-source readers reused unchanged, or are readers retrained on ConvNeXt-Small development trajectories?

Why do the primary CIFAR experiment and the controlled-ablation record have different endpoint and full-CPRU ASRs?

Missing references BP14 does not include any the citation if the authors are referring to some existing work. Literature on model-ensemble / multi-surrogate transfer attacks is missing and requires to be discussed and differentiated with the proposed approach.

Examples: Liu et al., “Delving into Transferable Adversarial Examples and Black-Box Attacks,” ICLR 2017. Chen et al., ICLR 2024, Rethinking Model Ensemble in Transfer-based Adversarial Attacks. Yao et al., ICML 2025, Understanding Model Ensemble in Transferable Adversarial Attack.

4 comments

We thank the reviewer for the constructive comments. All locations below refer to the revised manuscript.

1. Uniform weighting aligns the objective and preserves strong gains.

We replaced the sign-dependent 31/3/1 weights with 1:1:1. The unweighted MSE (Eq. 29) now targets the conditional signed utility in Eq. 16, consistent with the ideal selector in Proposition 2. The target-clean normalization in Eq. 14 is retained. Performance remains strong: CIFAR-10 mean ASR improves by 7.347 pp (54.707% to 62.054%; Table 1). On ImageNet-1K, the revised 1:1:1 setting achieves +2.693 pp, compared with +2.280 pp in the original submission (Section 4.5, Table 4).

2. Model details and the origin and definition of BP14 are now explicit.

Appendix A.4.1, Tables A.2-A.3, lists all 17 CIFAR models' roles, constructors/model IDs, the fine-tuning recipes used to obtain their checkpoint weights, resize/interpolation settings, and clean accuracies. BP14 arose from our early work on MDCS: we sought to make its scaling responsive to disagreement between the current normalized gradient qi,tq_{i,t} and its historical mean qˉi,t−1\bar q_{i,t-1}. We introduced the coordinate-wise ratio ri,tr_{i,t} (Eq. A.73) and adaptive exponent pi,t=1/4+ri,t/4p_{i,t}=1/4+r_{i,t}/4 (Eq. A.74) to couple perturbation and momentum scaling (Eq. A.76). Although this improved the baseline, fixed-endpoint deployment remained a key transfer bottleneck, motivating CPRU. We retained BP14 as one of five trajectory generators; Appendix A.4.3, Eqs. A.70-A.77, now provides its complete definition.

3. RelMargin is a lightweight CPRU variant; the learned reader further improves ASR.

We added both requested selectors using the same trajectories, candidate actions, and four-proxy outputs (44 forwards per image and attack), without reader training or inference. RelMargin is a reader-free lightweight variant of the CPRU framework. It retains three core designs: (i) per-image intermediate-step selection; (ii) Final as an exact zero-utility fallback (Eq. 30); and (iii) clean-image RMS logit normalization sp,is_{p,i} and final-relative features dp,i,t=mp,i,t−mp,i,Td_{p,i,t}=m_{p,i,t}-m_{p,i,T} (Eqs. 24-26). RelMargin averages dp,i,td_{p,i,t}; RelVote averages proxy-success changes relative to Final. Both return Final unless the best score is positive.

Mean ASRs are Final 54.707%, RelMargin 60.049%, RelVote 57.822%, and CPRU 62.054% (Appendix A.8.2, Tables A.17-A.18). CPRU adds 2.005 pp over RelMargin and 4.232 pp over RelVote, improving all five attack-wise means. Removing final-relative features reduces the learned reader's gain from 7.347 to 2.753 pp (Table A.21).

4. The original source-transfer experiment refitted readers; we now also test frozen-reader reuse.

The original ConvNeXt-Small experiment trained readers on ConvNeXt-generated development trajectories to test CPRU's applicability with a different source. We have added direct reuse of the frozen, attack-specific ViT-B/32-source readers without ConvNeXt-specific refitting. On the same ConvNeXt evaluation record, Final achieves 43.775% MacroASR, the dedicated readers 48.035% (+4.260 pp), and the frozen readers 47.779% (+4.004 pp; 95% CI [3.430, 4.591]). The two settings are now explicitly distinguished in Section 4.4, Table 3, and Appendix A.5.4/Table A.7.

5. The CIFAR main results and ablations now use one locked evaluation archive.

For the ablations covering nine proxies, we independently reran the evaluation script to evaluate all configurations consistently. The original main and ablation tables differed in Final ASR by only about 0.13 pp; small floating-point differences in independent GPU reruns can be amplified by sign(∇xL)\mathrm{sign}(\nabla_x\mathcal L). After auditing the logs, we unified Table 1, Table A.4, and all controlled ablations in Appendix A.8 (Tables A.17-A.30) under one locked primary experiment archive, with shared default results Final = 54.707%, CPRU = 62.054%.

6. Ensemble generation and post-hoc trajectory readout play different roles.

We now cite MDCS (Tao et al., 2026) as the origin of our BP14 variant and expand the discussion of ensemble transfer attacks in Section 1 (Liu et al., 2017; Dong et al., 2018; Chen et al., 2024; Yao et al., 2025).

The distinction is between perturbation generation and predictive readout. These ensemble attacks backpropagate through multiple white-box surrogates during trajectory generation and conventionally deploy the final iterate xTx_{T}. CPRU instead uses proxy logits obtained solely through forward passes for post-hoc selection of an image-specific saved state xt^ix_{\hat{t}_{i}}. The proxies are excluded from source optimization and remain disjoint from the source (Ms∩P=∅\mathcal{M}_{s} \cap \mathcal{P} = \emptyset, Eq. 23), serving as external observers that provide signals of transfer saturation along the trajectory. This readout requires zero additional source-gradient evaluations (Eq. A.63).

Deadline for Discussion is Approaching in Less than 3 Days

Dear reviewer NG6k,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Deadline for Discussion is Approaching in Less than 2 Days

Dear reviewer NG6k,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in about 3 days only, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in One Day]

Dear reviewer ,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only one day, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.