acceptodds
Withdrawn from ICLR 2027

The Mean Is Not Enough: Defeating Query-Based Attacks with Learned Disagreement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Many defenses against query-based adversarial example (AE) attacks are vulnerable to Expectation over Transformation (EOT): by averaging repeated responses, an attacker can estimate the defense's mean response and optimize directly against it. But recovering the mean is useful only when AEs crafted against the mean transfer reliably across the predictors. We present DISSENT, which suppresses this transferability by learning a distribution of predictors that agree on clean inputs but disagree along adversarial directions. Its learning objective discourages AEs crafted against one sampled predictor from transferring to another, while preserving agreement on benign inputs. At inference, each query is answered by one randomly selected predictor in a single forward pass, avoiding the computational cost of evaluating multiple predictors for voting or averaging. Our theoretical analysis explains how learned disagreement can sustain robustness even when the attacker knows the mean. We evaluate DISSENT on CIFAR-10 and ImageNet with different architectures. On CIFAR-10 with ResNet-50, for example, DISSENT retains 59.5% robust accuracy under EOT, compared with 23.1% for the strongest evaluated baseline. Further evaluations against adaptive attacks, including those crafted on method-aware surrogates, demonstrate its robustness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial robustness, Adversarial machine learning, Black-box adversarial attacks, Query-based attacks, Adaptive attacks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xH1tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DISSENT, a randomized defense against query-based adversarial attacks to lower down the effectiveness of adversarial example crafted against the mean predictor thorough EOT. The method uses lora adapters and cross-vulnerability training to encourage agreement on clean input and reduce adversarial transferability. The method is evaluated on CIFAR-10 and ImageNet against various attack methods and compared with many defense strategies.

Critical strengths
  1. The motivation for mitigating EOT-based attacks by distinguishing between recovering the mean prediction and achieving reliable adversarial transfer is insightful, and the proposed cross-vulnerability training directly addresses this problem.
  2. The evaluation is very comprehensive. The paper evaluates multiple adaptive attacks and provides extensive ablation studies supporting the effectiveness of learned disagreement across adapters. The comparison with other defenses also highlights the advantages of the proposed method over existing baselines.
Critical weaknesses
  1. Although the paper evaluates the inference latency and storage overhead, it doesn't quantify the training cost of generating adversarial examples and optimizing across sampled predictors. A fair comparison with conventional adversarial training under the same computational budgets would help establish whether the additional complexity is need. 2.The evaluation focuses on conventional image classifiers and dataset, while modern vision-language models increasingly support more complex, and the adversarial attacks have evolved to question-answering scenarios such as jailbreak or prompt injection. It is unclear how useful and how much overhead cost would the proposed method introduce. This limits the applicability of the proposed framework.
Questions for authors
  1. Could the author report the training FLOPs or GPU-hours of DISSENT and other defenses or compare their effectiveness under matched computational budgets?
  2. How could the proposed method be extended to vision-language models, particularly open-ended visual question answering tasks?
Reviewer_ZjqvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces DISSENT, a defense against query-based adversarial attacks that estimates the mean prediction. DISSENT trains a distribution of predictors as stochastic low-rank adapters on a frozen base network that agree on clean inputs but disagree along adversarial directions. During inference, a randomly chosen predictor returns the output. The paper theoretically shows that disagreement can preserve single-query robustness even when the attacker knows the mean exactly. Experiments on CIFAR-10 and ImageNet show strong results under standard, EOT, hard-label, transfer, and method-aware surrogate attacks.

Critical strengths
  • The central idea is clear and well motivated. The paper aims to solve a key challenge in randomized defense.
  • The experiments cover several defenses and different attack scenarios.
  • The paper provides theoretical analysis to demonstrate the role of disagreement between predictors.
Critical weaknesses
  • The experiments are limited. The paper only tests Wide Resnet and Resnet50, while there are many other models/architectures. The experiments only evaluate one hard label attack, HSJA, and only the EOT version.
  • Although the theoretical results show that disagreement makes the attack difficult, but it's not clear how the method increases disagreement. The paper only explains the intuition in Line 216-222, but low loss on adversarial samples and attack-aligned disagreement are not well-connected.
  • The method looks like using multiple adversarial trained models (last term in (1)), making the novelty insignificant. Also, the paper only performs ablation study for λc\lambda_c but not λv\lambda_v.
  • The method decreases the clean accuracy significantly compared to other defenses.
Questions for authors

What is the difference between this method and adversarially training multiple models independently?

Reviewer_4FV7OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper defends image classifiers against attackers that repeatedly query a model to construct adversarial examples. Randomizing the response alone may be insufficient because an attacker can average repeated responses and attack the mean. The authors propose DISSENT, which trains low-rank adapters on a frozen network so that the resulting predictors agree on clean images but share fewer adversarial errors. Each query is answered by one randomly selected predictor. The idea is that an example that fools the mean may still be classified correctly by the selected predictor. Experiments on CIFAR-10 and ImageNet report improved robustness, including 59.5% robust accuracy under EOT on CIFAR-10.

Critical strengths
  1. The motivation is clear: estimating the average response and reliably fooling the next sampled predictor are different problems. Training the predictors to reduce shared adversarial errors directly addresses this distinction.

  2. The evaluation is substantial, including EOT, transfer attacks, surrogates built using the defense design, DVERGE comparisons, and training ablations. The runtime measurements also support the practical benefit of answering each query with one forward pass.

  3. I appreciate that the theoretical analysis states its assumptions and checks related behavior in the trained networks. This provides useful context for the method.

Critical weaknesses
  1. The related-work discussion misses prior defenses that randomly select one model per input, including MTDeep [1], randomized classifier mixtures [2], and BARRE [3]. These papers should be discussed to clarify how DISSENT’s learned adversarial disagreement differs from existing approaches to constructing robust randomized ensembles.

  2. The method is motivated by changing the “mean” predictions. However, this is essentially aiming to change the model's behavior, hurting the utility when there is no attack in most cases. Table 2 supports this concern: the defended model’s accuracy drops noticeably. If the accuracy is significantly dropped, then the model’s calibration will be hurt a lot, see AAA’s discussions.

  3. Why not train an independent surrogate with exactly the published configuration? In Table 17, even the closest surrogate changes the loss weights. Since configuration mismatch is used to explain weak transfer, this missing control is particularly important.

[1] MTDeep: Boosting the Security of Deep Neural Nets Against Adversarial Attacks with Moving Target Defense, GameSec’19

[2] Randomization matters: How to defend against strong adversarial attacks, ICML’20

[3] On the Robustness of Randomized Ensembles to Adversarial Perturbations, ICML’23

Questions for authors

See weaknesses

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

Related papers

Discussion (0)

Sign in to comment.