acceptodds
Withdrawn from ICLR 2027

Enhancing Robustness through Repair-Oriented Honeypots and Selective Knowledge Aggregation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A model with established adversarial robustness can remain vulnerable to an encountered attack. Targeted improvement requires useful repair data and a learning procedure that limits damage to existing capabilities. We introduce HoneyTeacher, a general framework that couples repair-oriented honeypot induction with selective knowledge aggregation. The production model's estimated repair benefit guides honeypot configuration, shaping attack interactions to yield training material with greater downstream learning value. A material teacher learns from these data; aggregation then selects supervision from this teacher and the original model according to their joint correctness, consolidating complementary corrections while retaining existing strengths. We evaluate score-based black-box and gradient-based white-box instantiations on all 10,000 CIFAR-10 test images. Under Square and PGD attacks, respectively, target robust accuracy improves by 12.4% and 8.7% relative to the original model, and by 17.1% and 13.5% relative to ordinary-material direct learning. On the secondary Square evaluation, the white-box model improves accuracy by 24.0% relative to the original model and 34.2% relative to ordinary-material direct learning. Clean accuracy changes by +0.17 and −1.26 percentage points relative to the original model in the black-box and white-box settings, respectively, with all monitored capabilities meeting the adopted retention criteria. Induced material improves robustness under matched direct learning, and aggregation adds further gains; together, they reverse the net target-robustness losses observed with ordinary direct learning in both settings.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Robustness, Repair-Oriented Honeypots, Adversarial Material Acquisition, Selective Knowledge Aggregation, Capability Retention.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ULZsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Even adversarially trained models can remain vulnerable to specific attacks. Further training to improve robustness against such an attack may, in turn, reduce performance on images and attacks the original model handled correctly. To address this problem, the paper proposes HoneyTeacher. The method first collects adversarial images using a honeypot, a model configured by the defender for attackers to interact with. It then compares the outputs of the original model with those of a model trained on the collected examples and assigns new teacher outputs to the images. A new model is trained on this additional data, allowing it to incorporate new knowledge into its predictions while retaining its original capabilities.

Critical strengths

It is interesting that the method uses the honeypot and feedback manipulation to obtain attack examples useful for repairing the original model, and incorporates estimated repair utility into the design of its responses.

Critical weaknesses

Overall, the experimental validation is insufficient.

  1. CIFAR-10 is the only dataset used. The evaluation should include at least two datasets substantially larger than CIFAR-10.

  2. ResNet-18 is the only model evaluated. The evaluation should include deeper CNNs and more recent architectures, such as WideResNet. It should also examine modern transformer-based predictors, such as Vision Transformers.

  3. The evaluation considers only Square Attack and PGD with CE and DLR losses. A credible evaluation of an adversarial defense must include AutoAttack (Croce & Hein, 2020). Even if the paper aims to improve robustness against one particular attack, it should expand the set of target attacks to include each component of AutoAttack and examine how improving robustness against one component affects robustness against the others.

  4. The comparisons in this study consist only of ablations of the proposed framework. The authors should compare it with existing methods that have similar goals. For example,
    S. Dai et al., Adapting to Evolving Adversaries with Regularized Continual Robust Training (2025)
    should be considered as a baseline, although its objective is not identical. RCAT, discussed in the related work, should also be compared with the knowledge integration component of the proposed method. In addition, straightforward baselines such as further adversarial training or adversarial fine-tuning using the target attack should be evaluated. These methods generate adversarial examples adaptively during training, unlike training on fixed adversarial images as in fordf_{ord}.

Questions for authors

I would like the authors to substantially expand the experimental evaluation to allow a more rigorous assessment of whether the proposed method outperforms existing approaches.

Reviewer_6XgtOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes HoneyTeacher, a framework for enhancing a production model’s robustness against encountered adversarial attacks while preserving its existing capabilities. HoneyTeacher consists of two stages: (1) repair-oriented honeypot induction, which guides adversarial interactions toward generating training samples with higher estimated repair value, and (2) selective knowledge aggregation, which combines supervision from the original model and a material teacher to incorporate new corrections while mitigating capability degradation. Experimental results show that HoneyTeacher improves target robust accuracy by 12.4% and 8.7% relative to the original model under Square and PGD attacks, respectively, while satisfying the authors’ adopted capability-retention criterion.

Critical strengths
  1. The paper presents an interesting repair-oriented approach to honeypot-based defense, using estimated repair utility to guide adversarial interactions toward acquiring more valuable training material for robustness improvement.
  2. Beyond target robust accuracy, the paper reports clean accuracy, robustness against additional attacks, repair and damage statistics, providing a more comprehensive assessment of the proposed method.
Critical weaknesses
  1. The empirical scope is narrow, covering only one dataset, one ResNet-18 checkpoint, and two primary target-attack settings. Moreover, the paper lacks direct comparisons with closely related robustness-preserving update methods, such as RCAT, making it difficult to assess its advantages over existing approaches.
  2. The black-box material acquisition is not matched in actual query cost. Ordinary Square attacks use an average of 169 queries, whereas honeypot-induced attacks use 256. This difference may affect the quality and diversity of collected samples, making it difficult to isolate the contribution of repair-oriented feedback.
  3. As shown in Appendix A.4, earlier CIFAR-10 test results informed method and baseline revisions, potentially introducing selection bias. The independence of the final evaluation is a concern.
  4. The framework assumes that adversarial interactions are detected and routed to a defender-controlled honeypot. However, this paper does not consider adaptive attackers that recognize or account for the manipulated feedback. The effectiveness of repair-oriented induction under adaptive attack strategies remains unclear.
Questions for authors
  1. How does selective knowledge aggregation perform when applied to ordinary adversarial material? A complete two-factor ablation is needed to isolate the effects and interaction of the two proposed components.
  2. Can the authors provide comparisons with existing robustness-preserving update methods such as RCAT?
  3. How much of the benefit of repair-oriented induction comes from the repair-utility signal itself, rather than the additional queries or extended search trajectories?
Reviewer_8EXvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a method to improve an adversarially robust model against a specific attack it is not robust enough against, encountered at deployment time. It is based on two phases. In the first, a detected attack against the model is redirected to a honeypot model that alters its outputs so that the generated adversarial examples are more useful for improving the main model’s robustness, and a teacher model is fine-tuned from the original model on those inputs. In the second, a student model, initialized from the teacher, is trained via distillation from either the teacher or the original model, depending on the correctness of each input. The method is evaluated on CIFAR-10 with a ResNet-18, against Square (black-box) and PGD (white-box) attacks.

Critical strengths
  • The paper addresses a relevant topic with an interesting solution, i.e., using collected attack examples as a source to improve the models' robustness, and altering the model output at deployment time to steer the attack in order to better achieve that
Critical weaknesses
  • The paper’s writing is very hard to follow, and I found it even difficult to understand the paper’s core motivations and contributions from the abstract and introduction. The threat model is unclear and should be better described, explicitly stating everything concerning the attacker’s goal, knowledge, and capabilities, and other implications useful to assess the practicality of the defense in real-world scenarios (e.g., how the defender knows which attack is facing? how often is the model updated?). The authors implicitly assume that deployment-time attacks are always detected, but do not consider or discuss realistic cases in which a strong attacker can evade the detection mechanism. Moreover, no adaptive attackers are considered or discussed. For instance, an attacker might notice the honeypot's behavior or try to poison the model, since its inputs are used directly for training. Finally, the aggregation stage is very similar to Angioni et al. (2025) and PCT, from which it derives. The authors should better position their work with respect to them.

  • The experimental validation does not soundly support the contributions. The experimental setup considers only one (small) dataset, one model architecture, a single perturbation budget, and no repetitions across different random seeds (with the average and standard deviation reported). The black-box attacks also use a small number of queries (Square in AutoAttack uses 5000 queries). Moreover, in the black-box evaluation setting, ordinary sessions stop after the first success (avg. 169 queries), while induced sessions always run to 256 queries. The aggregation is applied only to f_new, with extra data and adversarial views, including Square ones. This might explain the white-box "secondary Square" gain. Finally, it seems that the experimental setting does not properly assess hard in-the-wild settings.

Questions for authors
  • Is it possible for the defender to run the same induced procedure offline, on its own dataset, with no attacker?
  • Could you please evaluate the final models with the full AutoAttack suite, and run Square with its default query budget (e.g., 5000)?

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

Related papers

Discussion (0)

Sign in to comment.