acceptodds
Withdrawn from ICLR 2027

Ordered Top-K UAPs: Benchmarking, Diagnosing, and a Set-then-Order Attack

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Ordered top- universal adversarial perturbations (UAPs) are a demanding and under-explored form of attack: *can one perturbation impose a prescribed ordered top- list on held-out images, for an ensemble of models?* We benchmark four ordered top- objectives, developed for per-image single-model attacks, in this setting over 3,180 UAPs spanning single models to an 18-model ensemble. The capability is real, but its reach shrinks as the ensemble grows: one perturbation controls ordered lists of up to eight classes on a single model, but only short lists across an ensemble. Fitted to many models, it also transfers: on seven unseen models, foundation-scale ones included, it imposes the attacker's class on up to two-thirds of images on average, and exact ordered pairs on fewer. No objective wins everywhere, and a single attack success rate (ASR) cannot say why. We therefore decompose ordered success into *set membership, ordering given membership, and cross-model agreement*. Ordering is set by how explicitly the loss encodes order (a control that encodes none sits on the chance line), while membership and agreement track attack strength; and each failure, whether under transfer, at large , under adversarial training or against input defences, is one factor collapsing. Acting on the diagnosis, we propose *Set-then-Order*, which establishes membership with a distribution-matching loss and then orders it with hinges, at matched compute. With one schedule frozen across settings, it is never meaningfully worse than the hinge loss alone, and gains substantially where the membership bought in the first phase survives the second.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
universal adversarial perturbations, adversarial examples, targeted attacks, ordered top-K attacks, adversarial robustness, 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_5sJgOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the novel problem of ordered top-k universal adversarial perturbations (UAPs). The authors show that the ASR of ordered top-k UAPs can be decomposed into membership ASR and ordering ASR, and hypothesizes that these two components are controlled by different factors. The authors use this decomposition to evaluate and analyze different attack losses and propose a novel attack method StO by combining existing ones. StO outperfoms existing attacks in certain models under per-model ordered ASR.

Critical strengths

The paper proposes a novel decomposition of ordered-set attacks and performs extensive evaluation and analysis of different attack losses on the ordered top-k UAPs task. Based on this analysis, the authors propose a novel attack method, StO, that outperforms baselines in certain models under per-model ordered ASR.

Critical weaknesses
  • The motivation for studying ordered top-K UAPs is not sufficiently clarified. Providing a concrete scenario where an attacker would need to enforce the same exact top-K ranking across different inputs and multiple models would be helpful.
  • The proposed joint ASR metric is central to the analysis, but is not used in evaluating the proposed StO attack.
  • The claim that “the two factors are controlled by different things”/“membership and ordering are bought separately” is not sufficiently supported by the experimental results. The results have shown that membership and ordering respond differently under different losses, but this does not establish that the two components are independently controllable or governed by different factors.
Questions for authors

None

Reviewer_f2DvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores whether one universal adversarial perturbation can impose a prescribed ordered top-K class list across images and models. It benchmarks several objectives and separates ordered success into set membership, ordering given membership, and cross-model agreement. Based on these observations, the proposed Set-then-Order attack first encourages membership using distribution matching and then applies an ordering objective. The experiments cover single models, ensembles, and transfer to unseen models.

Critical strengths

The benchmark involves 18 models. The decomposition of attack success is useful for inspecting different failure modes. The paper explicitly reports the visibility of the perturbations and the limitations of the chosen budget.

Critical weaknesses
  1. I still do not quite see the practical need for controlling an exact ordered top-K list. Targeted attacks, universal attacks, and transferable attacks each have a fairly clear motivation. Here, what concrete attacker objective requires the second, third, and later classes to appear in one prescribed order? The introduction mentions retrieval, triage, and moderation, but the experiments only manipulate ImageNet classification logits.

  2. The perturbation visibility considerably weakens the application argument. Appendix B reports a mean realized L-infinity magnitude of 103/255 and a maximum of 223/255. These are large pixel changes. The authors correctly state that the L2 budget controls energy rather than imperceptibility, but why is this the appropriate operating point for the proposed threat? The additional L-infinity projection reduces ASR by 27.6 percentage points on average, suggesting that the capability depends substantially on the allowed distortion.

  3. Most conclusions come from ImageNet-1K under one main L2 budget, and the half-budget experiments do not cover the 18-model ensemble. It is therefore difficult to know how generally the diagnosis and Set-then-Order schedule hold.

Questions for authors

See Weaknesses.

Reviewer_qc2kOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies ordered top-K universal adversarial perturbations (UAPs): can one fixed perturbation force a chosen list of K classes, in a chosen order, onto held-out images, for one model or for a whole ensemble at once?

The authors benchmark four existing ordered top-K objectives (CW-K, QuadAttacK/QP, RisingAttacK lifted to the universal setting, and KL/Adversarial Distillation) plus a no-order control (PGD-CE). They train 3,180 UAPs across 19 configurations, from single models to an 18-model ensemble, measure transfer on 7 unseen models (including CLIP and DINOv3 models), and also attack 3 adversarially trained models.

The main tool is a decomposition: ordered success = set success x P(order | set), and multi-model success is an intersection, summarized by three statistics (r, a strength-controlled dependence score D, and a mean pairwise correlation phi). Using it, they argue that (i) ordering is mostly set by how explicitly the loss encodes order, with the PGD-CE control landing on the 1/K! chance line; (ii) membership and cross-model agreement track attack strength; and (iii) transfer, input defences and adversarial training all mainly break membership, while ordering largely survives.

Building on this, they propose Set-then-Order (StO): optimize KL for 35 of 50 epochs, then switch to CW-K. With the schedule chosen on odd-class images and frozen, StO gains +38 points on ViT-B/16 and +6.7 on the 4-model ensemble at K=3 and is never more than 1.2 points worse than CW-K across 10 settings.

Critical strengths
  1. The experimental work is unusually careful and honest. This is the main reason I lean toward acceptance. The authors report the final-epoch perturbation rather than the best epoch, and they show how much best-epoch selection would have inflated results. They store per-image success indicators so joint and agreement statistics can be recomputed exactly. They select the StO switch point on odd-class images and report on the even half. They include a no-order control (PGD-CE) that lands on the analytic 1/K! line, which gives a measured baseline for what ordering is worth. They are also upfront that the l2 budget is not an imperceptibility constraint. With 3,180 runs and multiple seeds per cell, the benchmark is a useful resource for this area.

  2. The decomposition is simple but genuinely explains the results. Eq. 3 is just an identity, but separating membership, ordering and agreement turns confusing outcomes into clear ones. For example, KL's apparent collapse at K=2 (joint ordered ASR from 75% to 0.2%) turns out to be purely an ordering failure, since KL is actually the strongest objective on membership and agreement. Likewise, defences and transfer remove membership while ordering survives. These are practical takeaways for both attackers and defenders that a single ASR number would hide.

Critical weaknesses
  1. The threat model is weaker than the headline numbers suggest. The budget is ||delta||_2 <= 24.34, but by the authors' own measurement the mean l_inf is about 103/255 and the maximum 223/255, so the perturbations are clearly visible (Fig. 1). The authors also report that adding an l_inf projection costs 27.6 ASR points on average. Most UAP work uses l_inf budgets around 10/255 or 16/255, so the headline results ("8-class lists on DINOv3", "two-thirds transfer") are not comparable to that literature and overstate a realistic threat. The paper is open about this, which I appreciate, but it remains unclear whether the main findings (e.g., "ordering survives transfer") hold in the regime people actually care about.

  2. The Set-then-Order method is thin, and the explanation of when it works is after the fact. StO switches between two existing losses and clearly helps in only 2 of 10 settings (ViT-B/16 and the 4-model K=3 case); elsewhere differences are within noise, with several settings at 1/4 or 1/5 seeds won. The "kept" (retention) column is measured after training, so it describes where StO worked rather than predicting it, and the authors admit they cannot explain why retention is high on ViT-B and low on Swin-B. There is also a logical issue: the "ceiling" is described as an upper bound, yet StO exceeds it on ViT-B (+38.1 vs +34.2). This suggests something else drives the ViT-B gain. One candidate is that CW-K is simply under-optimized on ViT-B (23.1 vs 33.0 for KL), so any reasonable warm start might produce the same jump. Without a baseline such as CW-K warm-started from PGD-CE, it is hard to tell whether the KL phase specifically matters.

A relatively minor related concern: the lifted RisingAttacK and QP get higher per-model ASR on the 18-model ensemble (88.5% and 81.8% at K=1) than on the 4-model ensemble (54.9% and 71.7%). A likely cause is that the 18-model runs sample 4 models per step. If so, part of the "per-image leader ranks third" conclusion may reflect the lifting and implementation rather than the original method, and the lifting is never checked against the published per-image results.

Questions for authors
  1. Why do lifted RisingAttacK and QP achieve higher per-model ASR on the 18-model ensemble than on the 4-model ensemble at K=1? Is this caused by sampling 4 of 18 models per step? Running the 4-model ensemble with 2-of-4 model sampling would help settle this.

  2. With N=1 and a single image, does your lifted RisingAttacK reproduce the published per-image results? Without this check, it is hard to accept the claim that the lifting preserves the method.

  3. StO exceeds the stated ceiling on ViT-B/16. How should the ceiling be interpreted if it can be exceeded? Does the KL phase specifically matter, or would any membership warm start work? A run of CW-K warm-started from PGD-CE for 35 epochs would answer this directly.

  4. The caption of Table 19 notes that the objectives differ in more than how explicitly they encode order. Could you sweep the KL target margin (e.g., 0, 0.3, 1, 3)? This would directly test whether order explicitness is the driver of the ordering term, and it is a cheap experiment.

  5. Do the key findings (objective ranking, "membership breaks while ordering survives", and StO's gains) hold under a standard l_inf budget such as 10/255 or 16/255, even for a few configurations like the 4-model ensemble at K=1-3?

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

Related papers

Discussion (0)

Sign in to comment.