acceptodds
Under review as a conference paper at ICLR 2027

When Competitor Rankings Mislead: Dynamic Competitor Discovery for Untargeted Adversarial Attacks

Abstract

Untargeted adversarial attacks need only cross the decision boundary of a single competing class, yet determining which competitor to pursue remains challenging. Class logits provide readily available guidance, but their rankings can change during optimization. Early target commitment or score-dependent gradient weighting can therefore place excessive emphasis on transient rankings and steer optimization toward less promising competitors. We introduce Dynamic Competitor Discovery (DCD), which treats target selection as a coarse-to-fine process under uncertain competitor rankings. DCD begins with a broad candidate set, progressively narrows it as optimization proceeds, and selects a single target only after sufficient evidence has accumulated. DCD- extends this process to multiple targets by repeating DCD while excluding previously selected competitors. Across ten benchmark models and four datasets, DCD with PGD, APGD, GAMA, FAB, and FMN improves the corresponding base attacks in 49 of 50 attack–model comparisons under equal or lower gradient budgets, with consistent gains also observed under . Applied after AutoAttack, APGD-DCD9 uncovers 88 additional adversarial examples across all ten models, revealing vulnerabilities missed by AutoAttack.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Attacks, White-box Attacks, Untargeted Adversarial Attacks, Competitor Discovery, Adversarial Robustness

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.80 over 5 reviews, above 85% of the others.
Reviewer_Vs31OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Dynamic Competitor Discovery (DCD), which repeatedly re-ranks competing classes, progressively contracts the candidate set, and uses uniform mean-margin updates to select an attack target. A targeted attack then restarts from the clean input. Experiments demonstrate improvements across multiple attack families, models, and perturbation norms, as well as complementary coverage beyond AutoAttack.

Critical strengths
  1. Broad empirical effectiveness. DCD improves 49/50 standard L∞ attack–model comparisons under equal or lower gradient budgets, with consistent benefits under L2.
  2. Practical applicability. The procedure integrates with several existing attacks. Figure 3 also demonstrates strong performance with relatively few target attempts.
  3. Extensive experimental analysis. The paper includes mechanism ablations and sensitivity studies, and explicitly reports degradation under gradient masking.
Critical weaknesses

1. A central ablation contains an unresolved consistency issue

CIFAR-10 has nine incorrect classes. Under Eq. 91, setting K₀ = K_f = 9 yields K_t = 9 throughout discovery. According to Appendix C.8, this should be equivalent to fixed-width TKMM with K = 9. However, Table 12 reports:

ConfigurationCarm.Rebu.
K_f = 960.06%67.01%
TKMM K = 960.67%67.62%

Both differences are 0.61 percentage points. Since the stated ablation protocol changes only the named component, these discrepancies cannot be explained by the current description. Clarifying implementation differences or correcting the results is important for interpreting the contribution of support contraction.

2. The contribution of progressive contraction under uniform weighting is not fully isolated

On datasets with many classes, fixed-width controls only use K ∈ {2, 3, 5, 9}, so their degradation may partly reflect insufficient initial coverage. The FullSupportCE versus T-CE (τ = 1) comparison provides evidence for contraction under CE weighting. However, the default mean-margin formulation lacks a matched fixed-full-support control on these datasets. Consequently, attributing the principal gains to progressive contraction requires stronger evidence separating contraction from broad initial coverage.

Questions for authors
  1. Please explain why K_f = 9 and TKMM K = 9 differ on CIFAR-10. Are there undocumented differences in budgets, updates, target selection, or initialization?
  2. Please provide a fixed-full-support mean-margin control on datasets with many classes, keeping the 15-step discovery phase, 85-step downstream attack, and remaining settings identical.
  3. Optional, not central to my recommendation: Compare DCD against existing attacks given the same additional budget on AutoAttack survivors, to quantify its practical value as an additional evaluation component.
4 comments
AuthorsOpenReview ↗

Reply to Reviewer Vs31 Part 1

Reply to W1 and Q1.

We thank the reviewer for identifying this inconsistency. The reviewer is correct that, on CIFAR-10, Kf=9K_f=9 and TKMM K=9K=9 should be equivalent under the stated ablation setting. The discrepancy resulted from an implementation error in the original TKMM experiments, which we have identified and corrected.

1. Cause of the inconsistency.

The inconsistency arose because the original TKMM implementation used a different update rule. Unlike DCD, which updates the current input xtx_t at each iteration, TKMM computed each new input relative to the original input x0x_0. As a result, the original comparison changed both the support size and the update rule, rather than isolating the effect of support size. We have corrected TKMM to use the same update rule as DCD, with only the support size fixed at Kt=KK_t=K.

2. Corrected ablation results.

We reran all affected TKMM experiments and updated the corresponding results in the paper. The corrected results are reproduced below.

Corrected support ablations (robust accuracy, %; lower is better). All variants use 15 discovery steps and 85 targeted attack steps. TKMM uses the same DCD update with fixed Kt=KK_t=K. Avg. is the mean across eight models.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.
Standard59.8066.7837.0119.0830.3636.8214.8916.4835.15
TKMM K=2K=259.9867.0137.4519.3031.3237.4015.6716.6135.59
TKMM K=3K=359.8366.9937.2719.2530.8637.0215.4016.6435.41
TKMM K=5K=559.8967.0637.4019.3730.6636.8615.2916.5935.39
TKMM K=9K=960.0667.0137.4419.4730.9237.0215.2016.6835.48
Kf=9K_f=960.0667.0137.4919.4930.9437.1015.1416.6635.49
$K_f=\vert \mathcal U\vert $60.0667.0137.5219.6031.3037.3215.4116.7435.62

The corrected results fully resolve the reported equivalence issue. On CIFAR-10, TKMM K=9K=9 and Kf=9K_f=9 now produce identical robust accuracies of 60.06% on Carmon and 67.01% on Rebuffi. The Kf=∣U∣K_f=|\mathcal U| variant also yields the same results because CIFAR-10 has exactly nine competitors.

3. Progressive support contraction remains beneficial after the correction. With the corrected implementation, DCD outperforms all fixed-KK variants on every evaluated model. DCD progressively reduces the active support and achieves a mean robust accuracy of 35.15%, compared with 35.39--35.59% when the support size remains fixed and 35.62% when all competitors remain active. Although the improvements are smaller than originally reported, these results consistently demonstrate the benefit of progressive support contraction under the same update rule.

We have corrected the TKMM implementation and updated the affected ablation results and discussion. This discrepancy also prompted us to conduct a systematic review of all experiments. We verified that the reported results matched their actual implementations and corrected any inconsistencies. We have documented the implementation fix in the code and confirmed that the revised results are consistent with the experimental protocols described in the paper. We thank the reviewer for identifying this important issue.

Reply to W2 and Q2.

We agree that the fixed small-KK baselines do not fully isolate the benefit of progressive support contraction. Unlike DCD, these variants consider only a few competitors from the beginning, so their weaker performance may also result from limited initial candidate coverage. To address this concern, we have added a matched fixed-full-support control (Kf=∣U∣K_f=|\mathcal U|) to Table 15, reproduced below.

Effect of support contraction under uniform mean-margin weighting (robust accuracy, %; lower is better). Both variants use full initial candidate coverage and identical gradient budgets. Avg. is the mean across eight models.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.
Standard DCD59.8066.7837.0119.0830.3636.8214.8916.4835.15
$K_f=\vert \mathcal U\vert $60.0667.0137.5219.6031.3037.3215.4116.7435.62
AuthorsOpenReview ↗

Reply to Reviewer Vs31 Part 2

1. Isolating support contraction from initial coverage.

Both Standard DCD and the new control start with all eligible competitors and use identical mean-margin weighting, optimization updates, target selection, and gradient budgets (15 discovery steps and 85 targeted attack steps). The only difference is whether the active support progressively contracts. Standard DCD gradually reduces the support size, whereas Kf=∣U∣K_f=|\mathcal U| retains all competitors throughout discovery.

2. Experimental evidence for progressive contraction. Standard DCD outperforms the fixed-full-support control on all eight models, reducing mean robust accuracy from 35.62% to 35.15%, an improvement of 0.47 percentage points. The benefit is particularly clear on large-class models. On ImageNet, the improvements are 0.94 and 0.50 percentage points, while ImageNet-12K shows improvements of 0.52 and 0.26 points.

These results directly isolate the contribution of progressive contraction under uniform mean-margin weighting, independent of initial candidate coverage. Together with the existing FullSupportCE versus T-CE (τ=1\tau=1) comparison, they show that progressive support contraction improves attack performance under both mean-margin and CE-style weighting. We have incorporated the new control results and clarified these findings in the revised manuscript.

Reply to Q3.

We thank the reviewer for this valuable suggestion. We have extended Table 3 to compare APGD-DCD9 with MALT, MM, and MultiTargeted on the same examples that remain robust to AutoAttack. Each attack is evaluated independently, allowing a direct comparison of the additional adversarial examples found by APGD-DCD9 and the competing attacks.

1. Experimental setup and computational budget. Each supplementary attack is evaluated independently on the same AutoAttack survivors. We follow each method's evaluation protocol and gradient budget reported in Table 2. For APGD-DCD9, the stated budget includes all gradients used for target discovery. Table 8 (reproduced below) reports end-to-end runtime, and shows that APGD-DCD9 achieves comparable average runtime to the multi-target baselines.

Average attack runtime per image (seconds, lower is better). Times include target discovery or ranking where applicable, including the additional gradient computations used by MALT for target ranking. Avg. is the arithmetic mean of the displayed runtimes across the eight models, rounded to three decimal places.

CIFAR-10CIFAR-100ImageNetImageNet12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.↓\downarrow
APGD-T0.5172.8833.4880.4731.3071.0513.0172.7911.941
MM+0.4913.1813.3230.4621.3281.0142.9732.8081.948
MALT0.5102.9133.4230.4971.3591.1803.1882.9812.006
APGD-DCD90.4952.9043.3980.4661.2971.1213.0962.7861.945

Additional adversarial examples found after AutoAttack. Each attack independently evaluates the same examples that remain robust to AutoAttack. Higher is better. Total is the sum across ten models.

CIFAR-10CIFAR-100ImageNetImageNet-12K
AttackCarm.Rebu.GowalRiceEngs.Salm.MIM.CXNCXCViTCTotal
APGD-DCD955835619202688
MALT01814616142162
MM+46412205132158
MultiTargetedPGD01100304172551
+APGD-T-R542112023424

2. Additional adversarial examples beyond AutoAttack.

Table 3 (reproduced above) shows that APGD-DCD9 uncovers 88 additional adversarial examples, compared with 62 for MALT, 58 for MM, 51 for MultiTargetedPGD, and 24 for APGD-T-Restart. It achieves the highest or tied-highest success count on nine of the ten models. Notably, on ImageNet-12K CXN, CXC, and ViTC, DCD9 finds 9, 20, and 26 additional examples, respectively, exceeding all competing attacks.

Overall, APGD-DCD9 uncovers the most additional adversarial examples missed by AutoAttack, leading or tying on nine of the ten models. It outperforms MultiTargetedPGD, MM+, and MALT in total additional successes, at comparable gradient budgets per image (Table 2) and achieving comparable average runtime (Table 8). These results demonstrate the effectiveness of dynamic competitor discovery in revealing vulnerabilities missed by existing attacks and enabling more thorough robustness evaluation.

Thank you for the detailed response and the additional experiments. Most of my concerns have been addressed. After reading the other reviews, I decide to maintain my current rating

Thank you again for your careful review. Your feedback was essential to improving the quality of our work.

Reviewer_fF6sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose a method to improve untargeted evasion attacks by dynamically selecting the optimal class to push the adversarial examples towards. It is based on a 15-step discovery phase that starts from the top-k most confident classes and iteratively reduces them to 2. The evasion attack is then run in a targeted manner, using the most confident of the 2 classes as the target. The discovery phase can also be repeated multiple times (excluding previously selected targets) to obtain multiple targets.

Critical strengths
  • The paper addresses a relevant topic and provides a practical solution that is simple, cheap, and easily integrable into common evasion attacks. Empirical comparison shows clear gains when applying the proposed method to standard untargeted attacks, with minimum computational overhead, and the thorough ablation studies support its main design ideas.
Critical weaknesses
  • The novelty behind the target selection method appears limited with respect to what is discussed in the paper. The authors should better discuss and position their work with respect to Yamamura et al. (2024), who apply a very similar target selection method, and Gao et al. (2022), who also tested ranking targets at adversarial points (which contradicts what the authors say).

  • The gains against the strongest competitors (multi-target attacks) are small, and the experiments are run once, without repeating them with different random seeds and reporting mean and standard deviation, potentially weakening their significance further.

Questions for authors
  • Could you please compare your method with the target selection of Yamamura et al. (2024) at the same gradient budget and using the same targeted loss, and clarify the relation with Gao et al. (2022) in re-ranking at adversarial points?
  • Could you please repeat the experiments using different random seeds and report the mean and standard deviation?
3 comments
AuthorsOpenReview ↗

Reply to Reviewer fF6s Part 1

Reply to W1 and Q1.

We thank the reviewer for highlighting these closely related studies. We agree that both target ranking at perturbed inputs and preliminary target selection before targeted optimization have been explored in prior work. We have clarified these connections in the revised paper. DCD's contribution is not ranking at a perturbed input alone, but jointly updating the observation state and refining a revisable competitor set through progressive mean-margin optimization.

1. Comparison with Yamamura et al. (2024).

Yamamura et al. first select a set of high-scoring target candidates and perform a short targeted attack for each candidate. The target achieving the highest objective value is then selected for further optimization. DCD differs in how promising competitors are identified. Rather than separately probing preselected targets, DCD jointly evaluates competitors through a single mean-margin optimization process. At each iteration, DCD re-ranks all eligible competitors, updates the active Top-KK set, and progressively reduces its size. Previously excluded competitors remain eligible for reconsideration as their rankings change.

Consequently, DCD requires TdT_d discovery-gradient updates regardless of the number of active competitors, rather than performing separate gradient-based searches for each candidate. This distinction becomes especially important in large label spaces, where a small set of preselected targets may exclude promising competitors.

Experimental results.

In APGD-EDA-TS, we combine the target-selection procedure of Yamamura et al. (2024) with the same APGD refinement and pairwise targeted-margin loss used in APGD-DCD1. In EDA-TS (K,s)(K,s), KK is the number of targets and ss is the number of gradient evaluations per target. The (9,5)(9,5) configuration uses 9×5=45 gradient for target selection and 56 for APGD refinement, while (5,10)(5,10) uses 5×10=50 gradient for target selection and 51 gradient for APGD refinement, respectively. Both therefore match APGD-DCD1's total budget of 101 gradient evaluations per input. As shown in Table 11 (reproduced below), mean robust accuracy is 31.00% for APGD-DCD1, 31.07% for EDA-TS (9,5)(9,5), and 31.10% for EDA-TS (5,10)(5,10) (lower is better). Compared with (9,5)(9,5), the stronger EDA-TS configuration on all four models, APGD-DCD1 trails by 0.05 percentage points on CIFAR-10 but leads by 0.06 on CIFAR-100, 0.14 on ImageNet-1K, and 0.12 on ImageNet-12K. Thus, DCD's advantage is most pronounced on the datasets with large label spaces.

Target-selection comparison with EDA. Robust accuracy (%) is reported, lower is better. The existing runs use 101 gradient evaluations per attacked input for APGD-DCD1 and APGD-EDA-TS variants. Avg. is the unweighted mean across the four models. The best result is bolded.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.RiceEngs.CXNAvg.↓\downarrow
APGD-DCD159.8019.0830.3614.7531.00
APGD-EDA-TS (9,5)(9,5)59.7519.1430.5014.8731.07
APGD-EDA-TS (5,10)(5,10)59.7719.1630.5814.8931.10

2. Comparison with Gao et al. (2022).

We acknowledge that Gao et al. (2022) also investigated target ranking using adversarially perturbed inputs, as reported in Appendix of their paper. Their experiments found limited improvements over clean-input ranking, which they adopted in the final STARS procedure. We have corrected our discussion to explicitly recognize this prior investigation.

DCD goes beyond replacing the clean input with a perturbed input for target ranking. It repeatedly updates the observation state using a mean-margin objective, re-ranks all eligible competitors, and progressively concentrates optimization on a smaller active set. DCD therefore continuously refines competitor selection during optimization, rather than selecting targets from a ranking obtained at a single input.

AuthorsOpenReview ↗

Reply to Reviewer fF6s Part 2

3. Experimental evidence supporting this distinction.

The experimental results further support DCD's distinct design. First, the Initial Ranking ablation fixes competitor rankings at the original input while keeping the remaining DCD procedure unchanged. This increases mean robust accuracy from 35.15% to 35.70% across eight models, demonstrating the benefit of dynamic re-ranking. Second, APGD-DCD9 achieves a lower mean robust accuracy than MM+ of Gao et al. (2022) across ten benchmark models (35.14% versus 35.23%), confirming its competitive performance against an established target-selection method.

In summary, DCD differs from Yamamura et al. and Gao et al. in how promising targets are discovered. DCD jointly evaluates competitors through mean-margin optimization, repeatedly re-ranks all eligible classes, and progressively narrows the active candidate set. This design enables adaptive target discovery without relying on separate target-wise searches or fixed initial rankings.

Reply to W2 and Q2.

We agree that improvements over strong multi-target baselines should be evaluated across multiple random seeds. We therefore repeated the comparisons using five seeds (1-5) on four representative models, covering CIFAR-10, CIFAR-100, ImageNet, and ImageNet-12K. The results are reported in Table 7 and reproduced below.

Robust accuracy (%, mean ±\pm sample standard deviation) over five random seeds. Lower is better; bold indicates the lowest mean for each model.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.RiceEngs.CXN
APGD-T59.588±0.00859.588 \pm 0.00818.992±0.00818.992 \pm 0.00830.316±0.00930.316 \pm 0.00914.490±0.02614.490 \pm 0.026
APGD-DCD959.584±0.01359.584 \pm 0.01318.928±0.008\mathbf{18.928} \pm 0.00830.144±0.017\mathbf{30.144} \pm 0.01714.248±0.019\mathbf{14.248} \pm 0.019
MultiTargeted59.812±0.01559.812 \pm 0.01519.090±0.01419.090 \pm 0.01430.384±0.01730.384 \pm 0.01714.486±0.03314.486 \pm 0.033
MM+59.566±0.017\mathbf{59.566} \pm 0.01718.964±0.01118.964 \pm 0.01130.280±0.01430.280 \pm 0.01414.450±0.05114.450 \pm 0.051
MALT59.578±0.01959.578 \pm 0.01918.944±0.01718.944 \pm 0.01730.184±0.01730.184 \pm 0.01714.290±0.03914.290 \pm 0.039

1. Performance stability across random seeds.

The five-seed results show low variability and preserve the overall performance ranking observed in the original experiments. APGD-DCD9 achieves the lowest mean robust accuracy on three of the four models, with standard deviations ranging from 0.008 to 0.019 percentage points. MM+ achieves the lowest mean on Carmon, exceeding DCD9 by only 0.018 percentage points. These results confirm that the reported performance trends are stable across repeated evaluations.

2. Performance gains and target efficiency.

Compared with APGD-T, APGD-DCD9 improves mean robust accuracy by 0.004, 0.064, 0.172, and 0.242 percentage points on Carmon, Rice, Engstrom, and ConvNeXt Nano, respectively. Although the improvements over the strongest competing methods are smaller, the repeated evaluations preserve DCD's advantage on these models, particularly on ImageNet and ImageNet-12K.

Importantly, DCD's advantage extends beyond the final robust accuracy. Figure 3 shows that DCD often matches or outperforms competing methods even when the number of targets considered is limited, highlighting its effectiveness in identifying promising competitors early.

In summary, the five-seed evaluations confirm the stability of DCD's performance, while the multi-target results demonstrate its effectiveness in identifying promising targets under limited optimization budgets.

I thank the authors for their rebuttal and additional experiments, which partially address my concerns. However, the new results confirm only small improvements with respect to prior methods, while the technical novelty appears limited. For these reasons, I keep my rating.

Reviewer_Abw9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Dynamic Competitor Discovery (DCD) for untargeted adversarial attacks, arguing that competitor rankings at a single state are unreliable, so attacks should not commit early to one target. DCD maintains a progressively shrinking Top-KK candidate set, assigns uniform weight to mean-margin gradients, repeatedly re-ranks competitors, and commits to a target only after discovery before restarting the base attack from the clean input; DCD-mm rediscovers multiple targets. Experiments across CIFAR-10/100, ImageNet, and ImageNet-12K under ℓ∞\ell_\infty and ℓ2\ell_2 show DCD strengthens multiple attacks, outperforms or matches multi-target baselines, and finds examples missed by AutoAttack and standard APGD objectives.

Critical strengths
  1. The paper identifies a genuine and well-motivated failure mode in adversarial target selection and turns it into a simple, principled mechanism. The decomposition mc(x)=mc⋆−Δc(x)m_c(x)=m_c^\star-\Delta_c(x) cleanly explains why observed competitor rankings can be state-dependent and unreliable, and the empirical rank-evolution analysis supports this concern. This is consequential because it reframes target selection from a one-shot ranking problem into a coarse-to-fine, revisable discovery process, and the resulting DCD design—Top-KK support contraction with uniform mean-margin allocation and deferred commitment—follows directly from that insight rather than being an arbitrary heuristic.

  2. The empirical validation is broad and convincingly demonstrates practical value beyond a single setting. DCD is evaluated across 16 classifiers, multiple datasets, both ℓ∞\ell_\infty and ℓ2\ell_2 threat models, and several base attack families, and it is compared against strong multi-target and target-selection baselines. The multi-target scaling results, additional examples beyond AutoAttack, exclusive-success analysis against APGD-CE/Margin/RCE, and mechanism ablations together provide strong evidence that the gains come from improved target discovery rather than incidental tuning. This is consequential because it supports both generality and complementarity, which are central to accepting a new attack mechanism.

Critical weaknesses
  1. The paper provides no evaluation of transferability, which is a central requirement for adversarial attacks in realistic black-box threat models. DCD relies on white-box gradients and logits to repeatedly re-rank competitors, contract the Top-KK support, and select a final target; all reported experiments are white-box attacks against the same model used for discovery. However, in black-box or transfer settings, the attacker cannot compute these state-dependent rankings on the target model, and a target discovered on a surrogate may be strongly overfitted to that surrogate’s local geometry. The paper does not test whether DCD-discovered targets transfer across models, whether the coarse-to-fine support contraction remains beneficial under surrogate-target mismatch, or whether DCD improves transfer attack success compared with standard transfer attacks such as MI-FGSM, DI, TI.

  2. The threat model is narrow and does not address realistic black-box query budgets or adaptive defenses, while some evidence suggests DCD’s advantage may shrink under gradient masking. The method assumes full white-box access to differentiable gradients and logits, a fixed ℓp\ell_p perturbation budget, and a standard classifier interface; it does not consider query-limited black-box attacks. The k-WTA stress test in fact shows that DCD’s gains are not robust under intentionally non-informative gradients, and the multi-target extension incurs O(mT)O(mT) cost, which is problematic when each gradient or query is expensive.

Questions for authors
  1. How transferable is DCD?

  2. How effective is it to combine DCD's dynamic ranking concept with other schemes? For example, MALT.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer Abw9 Part 1

Reply to W1 and Q1.

We thank the reviewer for highlighting the importance of transferability. We agree that cross-model transferability is important for surrogate-based black-box attacks. However, DCD addresses a distinct and well-established problem: strengthening white-box adversarial attacks under limited white-box computational budgets. We clarify its research scope and practical significance below.

1. White-box robustness evaluation is an independent and important research problem.

DCD uses the logits and gradients of the evaluated model to identify promising competitors and improve attack effectiveness. This white-box setting is also adopted by established attacks such as APGD, FAB, and MALT. In contrast, transfer-based attacks such as MI-FGSM, DI, and TI aim to generate adversarial examples on a surrogate model that also succeed against a different target model. These methods address a different attack setting and optimization objective.

Importantly, white-box attacks play a fundamental role in adversarial robustness evaluation and adversarial training. For example, RobustBench employs AutoAttack, including white-box components such as APGD and FAB, for standardized robustness evaluation (Croce et al., 2021). PGD-based adversarial training similarly generates adversarial examples using gradients of the model being trained (Madry et al., 2018). In these settings, the objective is to identify vulnerabilities of the evaluated model itself, regardless of whether the resulting adversarial examples transfer to other models.

2. DCD substantially strengthens white-box adversarial attacks.

Our experiments demonstrate the practical effectiveness of DCD for white-box robustness evaluation. Under equal or lowergradient budgets, DCD strengthens five base attacks in 49 of 50 attack-model comparisons across ten benchmark models. Furthermore, APGD-DCD9 uncovers 88 additional adversarial examples missed by AutoAttack, revealing vulnerabilities that standard robustness evaluation may overlook. These improvements directly support DCD's contribution to strengthening white-box attacks, independently of cross-model transferability.

We recognize that transferring DCD-discovered targets between different models introduces additional challenges due to surrogate-target mismatch. Such an extension would require separate investigation, and we make no claims about transferability in the present work. We have explicitly clarified the intended white-box threat model in the revised paper.

AuthorsOpenReview ↗

Reply to Reviewer Abw9 Part 2

Reply to W2.

We thank the reviewer for highlighting these concerns. We clarify the intended threat model, the implications of gradient masking, and the computational cost of DCD below.

1. Scope and practical relevance.

DCD focuses on white-box robustness evaluation, where model logits and gradients are available for identifying adversarial examples under prescribed perturbation constraints. This is the well established setting of widely used attacks such as PGD, APGD, FAB, and MALT, and is directly relevant to robustness evaluation and adversarial training, as discussed in our response to W1.

Our experiments cover four datasets, diverse CNN and transformer architectures, both standard and adversarially trained models, and ℓ∞\ell_\infty and ℓ2\ell_2 perturbations. DCD is designed to strengthen gradient-based attacks under these settings, rather than address every attack access model or defense mechanism.

2. Effect of gradient masking.

We acknowledge that DCD relies on informative gradients for competitor discovery. We deliberately include the k-WTA model as a gradient-masking stress test (Appendix C.8, Table 12), where gradients may not provide reliable optimization directions. As expected for gradient-based attacks, DCD does not consistently improve performance in this setting. Importantly, these results do not undermine the consistent improvements demonstrated across the ten standard benchmark models. Instead, they clarify the conditions under which DCD is effective. Our contribution focuses on strengthening gradient-based attacks, rather than overcoming gradient masking or arbitrary adaptive defenses.

3. Computational cost and efficiency.

APGD-DCD$m$ considers up to mm targets, with TT gradient updates allocated to each target. Its worst-case gradient cost therefore scales as O(mT)O(mT). APGD-DCD$m$ uses 15 of the 101 gradient evaluations per target for discovery and the remaining 86 for targeted optimization. This matches the gradient budgets of the corresponding multi-target baselines, including APGD-T and MM+. We additionally report end-to-end runtime measurements, including discovery and candidate re-ranking overhead. APGD-DCD9 averages 1.945 seconds per image across eight models, compared with 1.941 seconds for APGD-T, 1.948 seconds for MM+, and 2.006 seconds for MALT. Thus, APGD-DCD9 incurs only 0.2% additional runtime over APGD-T on average, while remaining faster than MM+ and MALT under the evaluated configurations.

In summary, DCD consistently improves attack performance across diverse models and attack methods under matched gradient budgets and comparable runtime. The k-WTA stress test highlights the importance of informative gradients, while the results across ten benchmark models and five base attacks demonstrate DCD's effectiveness in white-box robustness evaluation.

Reply to Q2.

We thank the reviewer for this suggestion. DCD and MALT address target selection through different mechanisms, and their potential integration is worth exploring.

1. Key differences between DCD and MALT.

MALT ranks candidate targets using class margins and gradient norms computed at the original input, based on its almost-linearity assumption. DCD instead repeatedly re-evaluates competitors during optimization and progressively narrows the active candidate set. Unlike MALT, DCD does not require separate gradient-based boundary estimates for individual competitors. The key distinction is that MALT selects targets using boundary estimates from the original input, whereas DCD refines competitor selection through successive optimization updates.

2. Potential integration and computational considerations.

The two approaches could be combined, for example, by applying MALT's distance-based ranking to candidates identified by DCD. However, this would introduce additional class-specific gradient evaluations and require careful allocation of the shared gradient budget. Whether the combined method improves attack performance would therefore need to be established experimentally.

Our experiments evaluate DCD and MALT independently to establish the effectiveness of dynamic competitor discovery. Combining them would require deciding when to introduce MALT's class-wise boundary estimates, which competitors to evaluate, and how to allocate the additional gradient computations within a fixed budget. These choices would lead to a new target-selection method requiring separate design and experimental validation. We therefore consider a DCD-MALT combination beyond the scope of the present work and leave its investigation for future research.

Although DCD cannot be applied to black-box scenarios, the author's clarification helped me better understand this work, and my questions have been resolved. I will raise the score to 3.

Thank you again for your thoughtful feedback and for raising your score. Your questions helped clarify our scope and suggested a direction for future work.

Reviewer_FE1yOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Dynamic Competitor Discovery (DCD), a coarse-to-fine target selection framework for untargeted adversarial attacks that progressively prunes an active competitor candidate set via uniform mean-margin updates, which circumvents the optimization pitfalls of premature target commitment induced by transient logit rankings and consistently boosts diverse base attacks across multiple benchmark models.

Critical strengths

The paper provides a convincing empirical analysis revealing the instability of initial logit rankings during optimization, which effectively motivates the dynamic coarse-to-fine candidate discovery framework and translates into solid, plug-and-play empirical improvements across various attack substrates and threat models without blowing up the gradient budget.

Critical weaknesses
  1. As descriped in Section 4.3 and the attack setup of Section 5.1, once target c^\hat{c} is selected at the discovery endpoint xTdx_{T_d}, the attack discards this state and restarts from the clean input x0x_0 using the remaining budget Ta=T−TdT_a = T - T_d. But the motivation of DCD is arguing that xTdx_{T_d} has a reduced margin deficit that makes target selection reliable, yet throwing away a candidate that already sits closer to the boundary feels geometrically counterintuitive. Dropping this accumulated perturbation essentially wastes TdT_d gradient steps just to fetch a target label. It would be helpful to see a direct comparison between restarting from x0x_0 versus simply continuing optimization from xTdx_{T_d}.

  2. The geometric support schedule in Algorithm 1, Kt←max⁡{Kf,⌈∣U∣1−t−1Td−1⌉}K_t \leftarrow \max\{K_f, \lceil \vert{}\mathcal{U}\vert{}^{1 - \frac{t-1}{T_d-1}} \rceil\}, drops candidates far too aggressively. Like, on ImageNet with ∣U∣=999\vert{}\mathcal{U}\vert{}=999 and Td=15T_d=15, over 90% of competitors are pruned within the first 3 to 4 steps. Considering the paper strongly criticizes prior methods for premature target commitment(section 4.1, etc.), this exponential decay seems to commit the same mistake very early on. How sensitive is this heuristic when shifting between small label spaces like CIFAR-10 and massive ones like ImageNet-12K?

  3. When uniformly averaging pairwise margin gradients over all active candidates (Eq. 7–8), it seems pose a clear gradient cancellation issue in the early iterations. When KtK_t contains hundreds or thousands of classes on datasets like ImageNet, individual gradient vectors ∇x(zc−zy)\nabla_x (z_c - z_y) point in widely divergent or even opposing directions across pixel space. Taking a simple unweighted mean risks washing out meaningful directional signals into near-isotropic noise, directly contradicting the intuition of guiding the search toward viable boundary gaps.

  4. While claimed a matched gradient budget in Table 2 by keeping T=100T=100 and N=9N=9, DCD-mm introduces noticeable runtime overheads in practice. Algorithm 2 requires re-running the entire Td=15T_d=15 discovery loop from scratch for every new target, which demands extra forward evaluations at the endpoints and repeated top-KK logit sorting over thousands of classes at every step. Could authors report end-to-end wall-clock time against standard multi-target baselines to demonstrate that sequential re-discovery remains affordable for large-scale evaluation?

Questions for authors
  1. It would be helpful to see a direct comparison between restarting from x0x_0 versus simply continuing optimization from xTdx_{T_d}.

  2. How sensitive is this heuristic when shifting between small label spaces like CIFAR-10 and massive ones like ImageNet-12K?

  3. Could authors report end-to-end wall-clock time against standard multi-target baselines to demonstrate that sequential re-discovery remains affordable for large-scale evaluation?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer FE1y Part 1

Reply to W1 and Q1.

We thank the reviewer for this important question. We agree that initializing the targeted attack from the discovery endpoint xTdx_{T_d} is a natural alternative. In fact, our early implementation used this approach. We directly evaluated both initialization strategies in Table 15, with the results reproduced below.

Comparison of clean-input and discovery-endpoint initialization (robust accuracy, %). Lower is better.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.
Standard DCD59.8066.7837.0119.0830.3636.8214.8916.4835.15
Use Endpoint59.8666.8237.0519.1130.3836.7814.9116.2935.15

1. Direct experimental comparison.

Both variants use 15 discovery iterations to identify the same target at the same endpoint, followed by 85 targeted APGD iterations. Standard DCD restarts from the clean input x0x_0, while Use Endpoint continues from xTdx_{T_d}. This comparison isolates the effect of initialization under identical gradient budgets.

Both variants achieve the same average robust accuracy of 35.15% across eight models. Clean-input initialization performs better on six models, while endpoint initialization performs better on two, particularly ViTC (16.29% versus 16.48%). These results demonstrate that retaining the discovery endpoint does not consistently improve attack performance.

2. Why the discovery endpoint is not necessarily a better initialization.

The key distinction is between target identification and targeted optimization. DCD optimizes the mean margin over an evolving competitor set, whereas the subsequent attack optimizes exclusively toward the selected target. The accumulated discovery perturbation therefore reflects multiple competitors and need not provide a better initialization for the final targeted attack. Importantly, the 15 discovery iterations are not wasted. They identify a promising competitor, allowing the remaining 85 iterations to focus on its decision boundary rather than continuing to distribute optimization effort across multiple competitors.

3. Rationale for clean-input initialization.

Restarting from x0x_0 preserves the original initialization and optimization procedures of each base attack. This is important to integrate DCD with different attack formulations, including distortion-minimizing methods such as FAB and FMN. Given the comparable average performance and the absence of a consistent advantage from endpoint initialization, we adopt clean-input initialization as the default. This choice keeps DCD independent of the base attack's optimization procedure while retaining the benefits of dynamic target discovery.

AuthorsOpenReview ↗

Reply to Reviewer FE1y Part 2

Reply to W2 and Q2.

We thank the reviewer for raising this concern. Importantly, geometric contraction reduces the number of active competitors, not the set of competitors eligible for final target selection. At each iteration, DCD re-ranks all eligible competitors and selects the current Top-KK set for gradient computation. Thus, a competitor excluded at one iteration can be reconsidered later if its rank improves. We further evaluate this design through additional ablations across different label-space sizes. 1. Why rapid contraction does not imply premature target commitment.

Unlike fixed-target methods, DCD does not permanently eliminate competitors based on early rankings. At each discovery iteration, the active Top-KK set is recomputed from all eligible competitors. A competitor excluded from the current update can therefore be reconsidered when its ranking improves. Moreover, the final target is selected after re-evaluating all eligible competitors at the discovery endpoint. Thus, the geometric schedule controls gradient concentration, not permanent competitor exclusion. Retaining a large support throughout discovery would continue distributing gradient weight across many potentially weak competitors. Progressive contraction instead focuses the updates on increasingly promising competitors, while keeping target selection revisable.

Support contraction ablation (robust accuracy, %). Lower is better.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.
Geometric59.8066.7837.0119.0830.3636.8214.8916.4835.15
Linear59.8166.8037.2019.2530.7036.8815.1216.5635.29
No contraction60.0667.0137.5219.6031.3037.3215.4116.7435.62

2. Direct evidence across different label-space sizes.

We evaluate three support schedules under identical discovery and attack budgets in Table 15 (and the Table above): the default geometric schedule, the Linear Shrink variant reported, and a newly added no-contraction variant (Kf=∣U∣K_f=|\mathcal U|). All variants retain the same mean-margin objective and base attack, isolating the effect of support contraction.

Geometric contraction achieves the best robust accuracy on all eight models, improving the average by 0.14 percentage points (pp) over linear contraction and 0.47 pp over no contraction. On CIFAR-10, geometric and linear contraction perform almost identically, with differences of only 0.01--0.02 pp. The advantage becomes clearer on larger label spaces. On ImageNet-12K, geometric contraction improves robust accuracy by up to 0.23 pp over linear contraction and 0.52 pp over no contraction. The results show that rapid support contraction does not compromise attack effectiveness. Instead, the geometric schedule outperforms both linear contraction and full-support averaging on all eight models, including ImageNet-12K with thousands of competing classes.

In summary, geometric contraction allows DCD to progressively focus gradient updates on fewer promising competitors without prematurely committing to a fixed target. Its consistent performance advantage across CIFAR-10, CIFAR-100, ImageNet, and ImageNet-12K supports its use across different label-space sizes.

Reply to W3.

We agree that competitor gradients may point in different directions and partially cancel. However, uniform averaging does not necessarily eliminate meaningful gradient information. We clarify this point and provide empirical evidence from the ablation studies.

1. Why uniform averaging remains meaningful.

The mean-margin gradient has the following decomposition: gt=1Kt∑c∈St∇x(zc−zy)=1Kt∑c∈St∇xzc−∇xzy.g_t =\frac{1}{K_t}\sum_{c\in\mathcal S_t}\nabla_x(z_c-z_y)=\frac{1}{K_t}\sum_{c\in\mathcal S_t}\nabla_x z_c-\nabla_x z_y. Although competitor-specific gradients may partially cancel, the shared true-class gradient −∇xzy-\nabla_x z_y remains intact. Importantly, gtg_t is the gradient of a well-defined mean-margin objective and provides a first-order ascent direction whenever it is nonzero. Moreover, DCD does not maintain broad gradient averaging throughout optimization. It progressively reduces the active support, concentrating subsequent updates on fewer promising competitors. The final targeted attack then optimizes exclusively toward the selected target.

AuthorsOpenReview ↗

Reply to Reviewer FE1y Part 3

2. Empirical evidence from support and weighting ablations.

We evaluate whether reducing the number of averaged gradients or assigning unequal weights improves attack performance. The results are reported in Table 15, including the additional Kf=∣U∣K_f=|\mathcal U| variant that retains all eligible competitors throughout discovery.

Progressive contraction is consistently beneficial. Standard DCD achieves an average robust accuracy of 35.15%, compared with 35.29% for Linear Shrink and 35.62% without contraction (Kf=∣U∣K_f=|\mathcal U|). These results show that progressively concentrating the mean-margin gradient improves performance over retaining a broad support.

Stronger score-dependent weighting does not improve performance. Under the same support contraction schedule, T-CE (τ=1\tau=1) achieves 35.17%, closely matching standard DCD at 35.15%. However, reducing the temperature to τ=0.2\tau=0.2 places greater gradient weight on higher-scoring competitors, but weakens attack performance, increasing robust accuracy from 35.17% to 36.03%. This provides further evidence that concentrating gradient weights on high-scoring competitors is not necessarily beneficial.

Overall, DCD combines uniform mean-margin weighting with progressive support contraction to address potential gradient cancellation. The ablations demonstrate its effectiveness over full-support averaging and strongly score-dependent weighting, even in large-class settings.

Reply to W4 and Q3.

We agree that matched gradient budgets do not necessarily imply comparable runtime. To directly address this concern, we have added end-to-end wall-clock measurements across eight models, reported in Table 8 and reproduced below. The measurements cover all discovery operations, including time for additional forward evaluations and repeated Top-KK candidate selection.

End-to-end attack runtime per image (seconds). Avg. is the arithmetic mean of the displayed runtimes across the eight models. Lower is better.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.
APGD-T0.5172.8833.4880.4731.3071.0513.0172.7911.941
MM+0.4913.1813.3230.4621.3281.0142.9732.8081.948
MALT0.5102.9133.4230.4971.3591.1803.1882.9812.006
APGD-DCD90.4952.9043.3980.4661.2971.1213.0962.7861.945

1. Direct wall-clock comparison with multi-target baselines. Table 8 reports the average runtime per image for each attack, including all computations for target selection and subsequent optimization. Across the eight models, APGD-DCD9 and APGD-T average 1.945 seconds and 1.941 seconds per image at the reported precision, compared with 1.948 seconds for MM+ and 2.006 seconds for MALT. APGD-DCD9, APGD-T, and MM+ use matched gradient-evaluation budgets, while MALT uses additional gradient evaluations for target ranking. Although runtime varies across models, we observe no substantial increase in average wall-clock time for DCD.

2. Why repeated discovery incurs limited runtime overhead.

DCD uses 15 of the 100 gradient updates per target for discovery and the remaining 85 for targeted optimization. Thus, repeated discovery does not incur extra gradient updates. Moreover, Top-KK selection operates on the logits already computed during discovery, without requiring separate gradient evaluations for individual competitors. Additional endpoint evaluations and candidate re-ranking may introduce some overhead. Their costs are included in the reported wall-clock measurements.

Overall, our results demonstrate that sequential rediscovery achieves runtime comparable to established multi-target attacks, even on models with thousands of competing classes. The additional computational cost is therefore manageable under the evaluated configurations.

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

This paper studies competitor selection in untargeted adversarial attacks. It argues that class rankings at an individual attack state can be unreliable indicators of which competitors are promising to pursue. The proposed Dynamic Competitor Discovery (DCD) method uses a short discovery phase that repeatedly updates competitor rankings, progressively reduces the active Top-K support, and averages the retained competitors’ margin gradients. After selecting a target, the base attack restarts from the clean input and uses the remaining budget for targeted optimization. DCD-m repeats this procedure while excluding previously selected targets. The experiments cover several attack families, four datasets, and both l-infinity and l-2 threat models. The paper reports improvements in 49 of 50 main l-infinity attack–model comparisons and 88 additional adversarial examples when APGD-DCD9 is applied after AutoAttack. The appendices provide conditional theoretical analyses, implementation details, and extensive ablations.

Critical strengths
  1. A clear and useful target-discovery procedure with meaningful component-level evidence. DCD provides a concrete way to revise competitor preferences before committing to a target. The separation between discovery and subsequent targeted optimization makes the procedure straightforward to combine with different base attacks. The ablations support important parts of this design: fixing the initial ranking increases average robust accuracy from 35.15% to 35.70%, and the tested fixed-width mean-margin variants are substantially weaker than the default progressively contracting support. These results support the usefulness of updating competitor rankings and refining the support within the proposed discovery framework.
  2. Broad empirical support with explicit accounting for discovery costs. The evaluation covers multiple attack families, architectures, datasets, and perturbation norms. The discovery phase is included in the per-target optimization budget, and Table 2 explicitly reports gradient and forward evaluations. The improvements across the main comparisons, the l-2 results, and the additional successes beyond AutoAttack support the method’s practical usefulness for robustness evaluation. The appendices also provide substantial ablation and sensitivity analyses, including comparisons that help distinguish target discovery from a favorable warm start. The reporting of settings where DCD does not consistently help, such as the k-WTA stress test, further clarifies its operating regime.
Critical weaknesses
  1. The multi-target evaluation does not isolate the benefit of the full discovery design. Table 12 provides evidence that dynamic re-ranking and progressive support contraction improve performance relative to the tested mean-margin controls. However, the FullSupportCE variant, which retains the discovery-then-targeted-attack procedure while using CE-style weighting over the full candidate set, achieves an average robust accuracy of 35.28%, close to 35.15% for standard DCD. For context, APGD-CE averages approximately 37.45% over the same eight models. The simpler variant therefore achieves much of the aggregate improvement observed for DCD. Because FullSupportCE changes both support selection and gradient weighting, this comparison does not isolate the contribution of either component. Figure 3 does not include a FullSupportCE-based multi-target variant. It also does not compare repeated discovery with performing discovery once and then attacking targets in the resulting fixed order. These controls would clarify whether the full discovery design is needed to obtain the reported multi-target gains, particularly in settings with many competing classes.
  2. The relative value of the additional computation after AutoAttack is not established. Table 3 demonstrates that DCD9 finds valid adversarial examples missed by the executed AutoAttack configuration. However, it does not compare alternative uses of the same additional budget, such as more restarts or targets for existing attacks, or an additional MM+/MALT run. Consequently, the 88 additional successes establish complementary coverage, but do not yet establish how advantageous DCD is as the choice of supplementary attack. This matters because the practical contribution includes improved robustness evaluation under limited computation. Comparisons on the same AutoAttack survivors, with matched additional budgets and reported execution times, would make that contribution substantially easier to assess.
Questions for authors
  1. Could you compare DCD-m with a FullSupportCE-based discovery variant under the same total gradient budget on one or two representative high-class-count models? Separately, how does repeated discovery compare with performing discovery once and then attacking targets in the resulting fixed order, with discovery and attack costs both included in the budget?
  2. On the same AutoAttack survivor sets used in Table 3, how many additional adversarial examples are found when a comparable extra budget is allocated to alternative attacks or to additional restarts/targets of existing attacks? Please also specify the AutoAttack configuration and, if available, report execution times for these supplementary attacks. These results would directly affect my assessment of DCD’s value as a practical addition to an existing robustness-evaluation pipeline.
2 comments
AuthorsOpenReview ↗

Reply to Reviewer kAdo Part 1

Reply to W1 and Q1.

We thank the reviewer for this important observation on our design ablation. We have added two sets of experiments to address these concerns directly.

Mean robust accuracy of 8 models across different support selection, gradient allocation, and stage separation.

Stage / SupportSingle-stageTwo-stageTwo-stage
Gradient allocationFull supportFull supportSupport contraction
Mean margin gradientAPGD-RCE: 43.92%$K_f=\vert \mathcal U\vert $: 35.62%Standard: 35.15%
CE-style gradientAPGD-CE: 37.45%FullSupportCE: 35.28%T-CEτ=1\text{T-CE}_{\tau=1}: 35.17%

1. Contributions of progressive support contraction and mean-margin weighting.

We agree that, FullSupportCE uses the full candidate set with CE-style weighting, whereas standard DCD progressively contracts the support and uses uniform mean-margin weighting. To isolate these two effects, Table 17 (reproduced above) compares standard APGD-DCD1 and FullSupportCE with two controlled variants. T-CE (τ=1\tau=1) retains progressive support contraction but uses CE-style weighting, while the new Kf=∣U∣K_f=|\mathcal{U}| variant retains uniform mean-margin weighting without support contraction.

Separating competitor discovery from targeted optimization already accounts for a substantial part of the performance improvement. Within this two-stage design, progressive support contraction consistently outperforms full-support optimization under both mean-margin and CE-style weighting. Although the additional gain from uniform mean-margin weighting over T-CE (τ=1\tau=1) is relatively small, mean margin offers important design advantages (Appendix A.6). It avoids excessive dependence on potentially unreliable score differences, requires no additional weighting parameters, and provides a simpler optimization objective. Although CE is widely used, it assigns greater gradient weight to competitors with higher current scores. Mean margin instead weights all retained competitors equally, regardless of their score differences.

2. Ablation comparison of DCD$m$ variants under multi-target evaluation.

We conducted an additional multi-target comparison, reported in Table 16 and reproduced below. FullSupportCE achieves performance comparable to standard DCD$m$. For example, on two ImageNet-1K models, DCD$m$ achieves robust accuracies of 30.14% and 36.68%, with FullSupportCE yielding similar results. Nevertheless, we favor uniform mean-margin weighting because it avoids additional dependence on potentially unreliable score differences among retained competitors, directly supporting DCD's central design principle of separating competitor selection from gradient allocation, as discussed further in Appendix A.6.

Ablation study of APGD-DCD$m$ design choices (robust accuracy, %). Lower is better. AVG is the mean robust accuracy across all eight reported models. The best result in each column is bolded.

CIFAR-10CIFAR-100ImageNetImageNet12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAVG$\downarrow$
Standard59.6066.6036.8218.9230.1436.6814.1115.0034.73
DCD$m$-OD-8559.6066.6036.8418.9230.2036.6814.2315.0434.76
DCD$m$-OD-10059.5966.5836.8518.9330.1636.6614.2115.0334.75
FullSupportCE59.5766.6036.8318.9330.1436.6814.1215.0634.74

3. How much do the discovery rule and repeated discovery contribute in multi-target attacks?

Table 16 (reproduced above) reports two comparisons across eight models. First, we replace the standard DCD discovery rule with FullSupportCE while keeping repeated discovery and the total gradient budget fixed. FullSupportCE closely matches standard DCD$m$ on average (34.74% versus 34.73%). The two methods achieve identical robust accuracies on both ImageNet-1K models at the reported precision. On the two ImageNet-12K models, DCD$m$ achieves 14.11% and 15.00% robust accuracy, slightly outperforming FullSupportCE at 14.12% and 15.06%, respectively.

Second, we evaluate the benefit of repeated discovery by comparing DCD$m$ with two one-discovery (OD) variants. Both perform DCD discovery once for 15 steps, then attack up to mm targets in a fixed order. OD-85 and OD-100 use 85 and 100 attack steps per target, respectively, requiring 15+86m15+86m and 15+101m15+101m gradient evaluations. OD-100 therefore uses 15 more gradient evaluations than standard DCD$m$. Standard DCD$m$ achieves a mean robust accuracy of 34.73%, compared with 34.76% for OD-85 and 34.75% for OD-100. These results show that repeated discovery improves attack performance over a fixed target order, even when the latter receives a larger gradient budget.

AuthorsOpenReview ↗

Reply to Reviewer kAdo Part 2

4. Overall, the new experiments confirm the value of DCD's core design. Progressive support contraction consistently improves single-target discovery, while uniform mean-margin weighting avoids additional dependence on uncertain score differences without sacrificing performance. In multi-target attacks, repeated discovery provides further gains over fixed-order selection, even under a smaller gradient budget. Together, these results establish the individual benefits of DCD's design choices while demonstrating the effectiveness of the overall approach.

Reply to W2 and Q2.

We thank the reviewer for suggesting this comparison. We compare APGD-DCD9 with MM+, MALT, MultiTargeted, and an additional APGD-T restart using a different random seed. All methods are evaluated separately on the same examples that AutoAttack failed to attack successfully. As shown in Table 2, APGD-DCD9, MM+, and the additional APGD-T restart use the same number of gradient evaluations, while MALT requires additional gradients for target ranking.

As shown in Table 3 (reproduced below), APGD-DCD9 finds 88 additional adversarial examples, compared with 62 for MALT, 58 for MM+, 51 for MultiTargeted, and 24 for the additional APGD-T restart. Thus, DCD9 finds 26 more examples than the strongest alternative supplement, MALT. It achieves the best or tied-best additional coverage on nine of the ten models. These comparisons show that APGD-DCD9 uncovers more adversarial examples missed by AutoAttack than the competing attacks under the reported computational budgets.

AutoAttack configuration.

We use the official AutoAttack standard configuration for the ℓ∞\ell_\infty evaluations, with the same dataset-specific perturbation bounds as the other attacks. The ensemble sequentially applies APGD-CE, targeted APGD-T, targeted FAB, and Square Attack to the remaining robust inputs. APGD and FAB use 100 iterations and one restart, with 9 target classes. Square Attack uses a maximum budget of 5,000 queries per input. APGD uses one EOT sample and ρ=0.75\rho=0.75. Our runner leaves these standard attack settings unchanged.

Additional adversarial examples beyond AutoAttack.

CIFAR-10CIFAR-100ImageNetImageNet-12K
MethodCarm.Rebu.GowalRiceEngs.Salm.MIM.CXNCXCViTCTotal
APGD-DCD$9$55835619202688
MALT01814616142162
MM+46412205132158
MultiTargeted01100304172551
APGD-T-R542112023424

Table 8 (reproduced below) further confirms the computational efficiency of DCD9 across eight models. APGD-DCD9 averages 1.945 seconds per image, compared with 1.941 seconds for APGD-T, 1.948 seconds for MM+, and 2.006 seconds for MALT. These results show that DCD9 achieves comparable runtime to established multi-target attacks while uncovering more adversarial examples missed by AutoAttack.

Average attack runtime per image (seconds, lower is better).

CIFAR-10CIFAR-100ImageNetImageNet12K
MethodCarm.Rebu.GowalRiceEngs.Salm.CXCViTCAvg.↓\downarrow
APGD-DCD90.4952.9043.3980.4661.2971.1213.0962.7861.945
APGD-T0.5172.8833.4880.4731.3071.0513.0172.7911.941
MM+0.4913.1813.3230.4621.3281.0142.9732.8081.948
MALT0.5102.9133.4230.4971.3591.1803.1882.9812.006

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.