acceptodds
Under review as a conference paper at ICLR 2027

RepAlign: A New Regularizer for Adversarial Training

Abstract

Deep neural networks are known to be highly vulnerable to adversarial examples, making robust training a practical necessity. However, robust training often leads to severe accuracy-robustness trade-offs. Existing methods like TRADES and Adversarial Logit Pairing (ALP) align softmax or logit outputs of adversarial inputs with those of clean samples, but are limited to layers at the end of the network. We propose RepAlign, a framework which uses representation similarity (RS) metrics such as Centered Kernel Alignment (CKA) to optimize for alignment between arbitrary features of clean and adversarially perturbed data. The use of general RS metrics allows RepAlign to align features having different properties. Evaluations across CIFAR-10, CIFAR-100, and Imagenette demonstrate that RepAlign consistently pushes the accuracy-robustness Pareto frontier forward. Soft-NN on CIFAR-10 achieves 86.90% clean accuracy (a gain over TRADES) while maintaining 49.37% robust accuracy ( over PGD-AT). On CIFAR-100, Soft-NN strictly outperforms TRADES on both fronts, achieving 60.42% clean and 26.10% robust accuracy (gains of and over TRADES, respectively). On Imagenette, CKNNA closes the gap to clean training () more than any other defense, reaching 88.46% clean accuracy ( over PGD-AT and over TRADES) while sustaining 61.89% robust accuracy. CKA heatmaps show that RepAlign mitigates representation drift and restores the layer specializaion disrupted by standard adversarial training.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
robust training, representation similarity, regularization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UnADOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes RepAlign, an adversarial training objective that replaces TRADES's output-level KL term with a representation-similarity (RS) term between clean and adversarial features: L=LCE(f(x′),y)−λ LRS(g(x),g(x′))L = L_{CE}(f(x'), y) - \lambda\, L_{RS}(g(x), g(x')), where gg is the penultimate layer and LRSL_{RS} is one of linear CKA, linear HSIC, CKNNA, or a differentiable Soft-kNN alignment computed per mini-batch. The inner maximization is standard PGD on cross-entropy, so training cost is close to PGD-AT. All RepAlign models are warm-started from a clean checkpoint and trained for 200 epochs with WRN-28-10 (CIFAR-10/100) and ResNet-50 (Imagenette). Evaluation uses the individual AutoAttack components (APGD-CE, APGD-DLR, FAB-T, Square), ART-PGD-20 and SPSA, with the minimum across AutoAttack components reported as "AA". The paper reports improved clean accuracy at comparable robustness (e.g., Soft-kNN on CIFAR-10: 86.90% clean / 49.37% "AA"; on CIFAR-100: 60.42% / 26.10% vs. TRADES 55.79% / 25.33%), and CKA heatmaps suggesting that RepAlign preserves the layer-wise block structure of clean models.

Critical strengths

S1. Simple transparently reported. RepAlign only adds a batch-level similarity term on penultimate features and keeps the PGD inner loop on cross-entropy, so training time is essentially that of PGD-AT and about 25–30% below TRADES (Table 13). The authors report full λ\lambda sweeps, both PGD step counts, both initializations, a temperature ablation, a middle-layer ablation and a reimplemented AFA baseline across seven attacks, including two black-box attacks. This transparency is what makes the issues below checkable, and it is to the authors' credit.

S2. A plausible gain on CIFAR-100, and a useful local-vs-global observation. On CIFAR-100, Soft-kNN (60.42% clean, 26.10% APGD-DLR) is better on both axes than every PGD-AT and TRADES (β=6\beta=6) configuration listed in Appendix Table 2 (best TRADES clean 57.09%, best TRADES APGD-DLR 25.44%). Across datasets, the local metrics (CKNNA, Soft-kNN) consistently outperform the global ones (CKA, HSIC), which is a concrete, testable insight into what kind of feature alignment helps adversarial training.

Critical weaknesses

W1. The evaluation and baseline reporting do not support the claimed Pareto improvement.

a. "AA" is not AutoAttack. Table 1 reports, as "overall AutoAttack robustness", the minimum of the per-attack accuracies of APGD-CE, APGD-DLR, FAB-T and Square, each run separately. AutoAttack (Croce & Hein, 2020) counts a sample as robust only if it survives all attacks, so its accuracy is the per-sample worst case and is generally lower than the minimum of the individual accuracies. All headline robust numbers in the abstract and Figures 1/3 are therefore optimistic, and they are not comparable with RobustBench or with the literature.

b. The main table does not use the best baselines, despite saying so. Appendix B states that Table 1 reports the best-performing baseline models. On Imagenette, Table 1 uses PGD-AT with 20-step training (86.65% clean, 62.52% APGD-DLR), but Appendix Table 2 shows PGD-AT, pretrained, 10-step reaching 88.15% clean, 65.81% APGD-CE, 63.95% APGD-DLR, 64.99% FAB-T, 66.55% ART-PGD, 74.17% Square and 84.31% SPSA. This configuration is better than Soft-kNN (Table 1) on six of seven metrics and tied on SPSA, and its APGD-DLR is higher than any RepAlign configuration in Tables 7–8. The Imagenette claim ("+1.81% over PGD-AT", "closes the gap to clean training more than any other defense") does not hold against the authors' own baseline.

c. CIFAR-100 baseline numbers cannot be traced. Table 1 reports TRADES at 55.79% clean / 25.33% APGD-DLR and PGD-AT at 56.78% / 22.90%. None of the four TRADES or four PGD-AT rows in Table 2 match these values (TRADES clean ranges 56.07–57.09%; PGD-AT clean ranges 49.50–54.77%), and Table 11 reports the same baselines as 56.07% and 54.77%. The abstract's "+4.63% over TRADES" depends on the 55.79% figure; against TRADES scratch/10-step (57.09%) the gap is 3.33.

d. Asymmetric tuning, apparently on the test set. For each dataset, RepAlign's reported result is the best of 4 RS metrics × 4 values of λ\lambda × 2 step counts (32 configurations), while TRADES is fixed at β=6\beta=6. No validation split is described, so the selection appears to use test accuracy. TRADES's own sweep (Table 10) shows it reaches 87.46% clean at β=1\beta=1 (warm-started), so much of the "+4.05% clean over TRADES" on CIFAR-10 is a different point on TRADES's trade-off curve, not a frontier shift. To be fair to the authors, Table 10 also suggests that at matched clean accuracy Soft-kNN may still be ahead under PGD-20 (≈51.8% vs. ≈48.5–48.9%), but this needs to be shown with full AutoAttack on all three datasets, not single operating points. On CIFAR-10, Soft-kNN is below TRADES β=6\beta=6 on every white-box attack (e.g., APGD-DLR 49.37 vs. 50.01; ART-PGD 51.82 vs. 53.10).

e. Variance and baseline strength. Table 1 reports standard deviations of ±0.01–0.06 pp "across three seeds" for adversarial accuracy, which is far below typical seed-to-seed variation in adversarial training, and clean accuracy has no deviation at all. This suggests the ± reflects attack randomness on a single model rather than training seeds. The baselines are also weaker than standard references (TRADES WRN-34-10 reaches about 84.9% clean / 53.1% AutoAttack on RobustBench, versus 82.85% / 50.0% here), training uses batch size 512 and step decay at epochs 100/150, and checkpoint selection is not specified. Robust overfitting (Rice et al., 2020) penalizes baselines without early stopping more than a regularized method, so part of the gap may be overfitting mitigation rather than the proposed alignment.

Together, these points mean the central claim, that RepAlign "consistently pushes the accuracy-robustness Pareto frontier forward" across all three datasets, is not established: it fails on Imagenette against the paper's own baseline, is a change of operating point on CIFAR-10, and rests on untraceable baseline numbers on CIFAR-100.

W2. Limited novelty against uncited prior work, and weak support for the mechanism.

a. Closely related work is missing. Regularizing hidden representations with HSIC for adversarial robustness was proposed in "Revisiting Hilbert-Schmidt Information Bottleneck for Adversarial Robustness" (Wang et al., NeurIPS 2021). Anchoring a robust model to a clean model to improve the accuracy–robustness trade-off is the idea of LBGAT (Cui et al., ICCV 2021), and HAT (Rade & Moosavi-Dezfooli, ICLR 2022) targets the same trade-off. None is cited or compared, and standard trade-off baselines such as MART and AWP (both cited) are not evaluated. Pairing-style regularizers have a history of overstated robustness: ALP, which the paper uses as motivation, was shown to provide much less robustness than claimed under stronger attacks (Engstrom et al., 2018). This makes the missing full-AutoAttack evaluation in W1 more serious.

b. The HSIC objective is ill-posed as written. Section 3.1 states LRS∈[0,1]L_{RS}\in[0,1], but Section 2.3 notes that linear HSIC is unbounded and scales with feature norm. Maximizing it (−λ HSIC-\lambda\,\mathrm{HSIC}) can be achieved by inflating feature norms, so the loss is unbounded below. This is consistent with the collapse of HSIC at λ≥5\lambda \geq 5 on Imagenette (19.21% clean in Table 7). More generally, CKA is invariant to rotation and isotropic scaling and measures batch-level relational structure, so it does not constrain the per-sample feature drift that adversarial robustness depends on. The paper offers no analysis of why it should help, and acknowledges the lack of theory.

c. The representation analysis is confounded. RepAlign models are warm-started from the clean model and regularized at the penultimate layer, so higher CKA similarity to the clean model is expected from the initialization and the objective itself. Section 5.2 does not state whether the TRADES model in Figure 2 is also warm-started; the comparison is only informative against warm-started PGD-AT/TRADES under the same analysis.

d. Hyperparameter fragility. The best λ\lambda differs by dataset (1 on Imagenette, 5–10 on CIFAR), the CIFAR-100 optimum lies at the edge of the grid (λ=10\lambda=10), and CKA, HSIC and Soft-kNN collapse on Imagenette at λ≥5\lambda\ge5 (Tables 7–8, as low as 19–42% clean). There is no recommended default, which limits the "drop-in regularizer" claim.

Questions for authors

Clarifications (could change my assessment):

Q1. Please report standard AutoAttack (the ensemble with per-sample worst case, version='standard') for every model in Table 1. Does the claimed frontier shift survive?

Q2. Why does Table 1 report Imagenette PGD-AT with 20 steps when Table 2 shows the pretrained 10-step configuration is better on almost every metric, and better than all RepAlign variants on robustness? Where do the CIFAR-100 PGD-AT (56.78%) and TRADES (55.79%) numbers in Table 1 come from, given that they match no row of Table 2 or Table 11?

Q3. How were the RS metric, λ\lambda and step count chosen for Table 1? Was a held-out validation split used, or the test set?

Q4. What do the ± values in Table 1 represent: training seeds or attack restarts on one model? Please report clean and robust accuracy as mean ± std over at least three independently trained models.

Q5. Which checkpoint is evaluated (last epoch or best by robust validation accuracy)? How do the baselines compare when early stopping is used (Rice et al., 2020)?

Q6. For HSIC, how do feature norms evolve during training, and is the objective bounded in practice? Is the Imagenette collapse caused by norm inflation?

Experiments needed to support the claims:

Q7. A matched-clean-accuracy comparison: TRADES (e.g., β∈{1,2,3,6}\beta\in\{1,2,3,6\}) evaluated with full AutoAttack on all three datasets, plotted as a curve against RepAlign's λ\lambda curve, rather than single operating points.

Q8. Comparisons with the closest prior work and standard trade-off methods under the same protocol: HSIC bottleneck (Wang et al., 2021), LBGAT (Cui et al., 2021), HAT (Rade & Moosavi-Dezfooli, 2022), and MART/AWP.

Q9. The CKA heatmap analysis for warm-started PGD-AT and TRADES, to separate the effect of initialization from that of the regularizer.

Q10. Sensitivity to batch size, since the RS terms are computed on b×bb\times b Gram matrices and training used b=512b=512 (e.g., results at b=128b=128).

Minor: Table 13 reports RepAlign epochs as faster than PGD-AT even though RepAlign adds a clean forward pass and the RS computation; please explain. Typos: "specializaion" (abstract), "arbitary" (Sec. 2.2, 3.2); Figure 2 caption refers to columns 4–6 and SVHN.

References -

Croce, F., Hein, M. Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse Parameter-free Attacks. ICML 2020.

Wang, Z., Jian, T., Masoomi, A., Ioannidis, S., Dy, J. Revisiting Hilbert-Schmidt Information Bottleneck for Adversarial Robustness. NeurIPS 2021.

Cui, J., Liu, S., Wang, L., Jia, J. Learnable Boundary Guided Adversarial Training (LBGAT). ICCV 2021.

Rade, R., Moosavi-Dezfooli, S.-M. Reducing Excessive Margin to Achieve a Better Accuracy vs. Robustness Trade-off (HAT). ICLR 2022.

Engstrom, L., Ilyas, A., Athalye, A. Evaluating and Understanding the Robustness of Adversarial Logit Pairing. arXiv:1807.10272, 2018.

Rice, L., Wong, E., Kolter, J. Z. Overfitting in Adversarially Robust Deep Learning. ICML 2020.

Zhang, H., Yu, Y., Jiao, J., Xing, E., El Ghaoui, L., Jordan, M. Theoretically Principled Trade-off between Robustness and Accuracy (TRADES). ICML 2019.

2 comments

We thank the reviewer for their careful and nuanced feedback and address their concerns below.

  1. "...standard AutoAttack...": This concern is addressed in the common concerns along with those of the other reviewers. In summary, we find only minor changes in the robust accuracy values and our claims about pushing the Pareto frontier forward remain valid.

  2. "...Imagenette PGD-AT with 20 steps...": We used 20 PGD steps during training for PGD-AT, TRADES, and RepAlign in Table 1 for standardisation. We have addressed this concern in the common concerns along with those of the other reviewers. In short, we find that we find that CKNNA achieves a trade-off with TRADES which using the PGD-10 attack, with neither dominating. CIFAR-100 numbers in Table 1: For CIFAR-100, the original evaluation files and the new AA evaluations confirm the main-table clean accuracies: 56.78% for PGD-AT and 55.79% for TRADES. The appendix lists 54.77% and 56.07% erroneously; we will update those rows and make evaluation results consistent.

  3. Hyperparameter choice: Table 1 reports all four RS metrics, each trained with 20 PGD steps. We chose λ\lambda from the appendix sweeps by comparing clean and robust accuracy, and used the same λ\lambda value for all four metrics within each dataset: 5 for CIFAR-10, 10 for CIFAR-100, and 1 for Imagenette. This lets us compare the metrics with the same coefficient and attack budget; it does not mean each metric was individually optimized. Selection in all cases used the CIFAR test sets and Imagenette validation set, which were also used for the final evaluations. We did not use a separate held-out split and will make this limitation explicit, although the same procedure is used for all the baselines and our methods.

  4. Independently trained models: The ±\pm values in Table 1 are standard deviations across attack-evaluation seeds 10, 20, and 30, using the same trained checkpoint. The seeds change the attacks' random choices, while the model weights stay fixed. Clean accuracy has no deviation because the same model is evaluated on the same clean inputs. We will clarify this in the caption. We did not add this variation in the initial paper as we did not find significant variation (less than 1%). We have also started three CIFAR-10 training runs for each of the four RepAlign variants, with training seeds 10, 20, and 30. We will report clean and standard-AA accuracy as mean ± standard deviation across those models once training and evaluation are complete. Our preliminary results indicate that the variation in clean accuracy is under 1 percentage point and that for robust accuracy is under 0.3 percentage points.

  5. Checkpoint choice: This concern is addressed in the Common concerns for all reviewers. In summary, for both baselines and RepAlign, we evaluate the checkpoint with the highest robust accuracy under the PGD attack.

  6. HSIC unboundedness: We apologise for the error. It is true that HSIC is unbounded and thus can cause issues with scale and results of optimization. However in our defense, HSIC was included for completeness. It does not claim the exact Pareto frontier. We will rectify our claims about the boundedness of representation similarity scores in [0,1][0,1]. We have provided figures for Pareto frontiers in Appendix G.2 in the updated draft for your consideration.

  7. Matched clean accuracy comparisons: We ran additional TRADES β\beta variation (β=1,2,5,6\beta = 1,2,5,6) experiments for all the datasets using 20 PGD steps for the inner training loop. We compared these with RepAlign's λ\lambda variation (λ=0.1,1,5,10\lambda=0.1, 1, 5, 10). In our results, we observe that we are pushing the Pareto frontier in all 3 datasets. Details are in Appendix G.2 of the paper.

  8. Other comparable methods: We thank the reviewer for pointing out these comparable methods. From the HAT paper, we find that our methods perform much better and significantly push the Pareto frontier forward compared to the results from that paper. We have added results for MART in Updated Table 1 in Common concerns and more details in Appendix G.3, and find that it does not lie on the Pareto frontier. Finally, the HSIC bottleneck paper is indeed a close method to ours although it constrains the HSIC between the features and inputs, and the features and the labels. We will add it to our baselines as soon as possible.

  9. Heatmaps comparison: The heatmaps in all cases are for models trained with a warm start, thus the PGD-AT and TRADES models also start from a clean model trained for 200 epochs. We will clarify this in the updated version of the paper.

  10. Batch-size sensitivity: We chose batches that fit our GPUs: 512 for CIFAR-10/100 with WRN-28-10, and 128 for Imagenette with ResNet-50. For evaluation, we used a batch size of 128 for all datasets. We have not run a controlled batch-size comparison, so we cannot quantify the sensitivity yet.

  1. Minor: RepAlign adds a clean forward pass and computes alignment once per batch. This happens outside the PGD loop, and is very rapid so it does not add any overheads. For TRADES, our implementation is slightly inefficient as it recomputes clean predictions at each attack step for the regulariser. We will correct this minor issue in our updated code. We will correct “specializaion” and “arbitary,” and remove the outdated SVHN reference. Figure 2 contains six columns and shows CIFAR-10, CIFAR-100, and Imagenette. We will revise the caption to describe those panels clearly.
Reviewer_c2EXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RepAlign adds a representation-alignment regularizer to adversarial training, encouraging clean and perturbed inputs to preserve similar feature relationships using CKA, HSIC, CKNNA, or Soft-kNN. Experiments on CIFAR-10, CIFAR-100, and Imagenette report improved clean accuracy with competitive robustness, supported by hyperparameter ablations and feature analyses.

Critical strengths

1.A useful and practical regularization idea. RepAlign aligns relationships among clean and adversarial representations rather than requiring pointwise feature equality. Its integration into standard adversarial training is straightforward, with reported training costs comparable to PGD-AT. This makes the contribution practically relevant and easy to investigate further.

2.Promising evidence of improved accuracy–robustness trade-offs. On CIFAR-100, Soft-kNN reports 60.42% clean accuracy and 26.10% APGD-DLR accuracy, improving over the main-table TRADES baseline by 4.63 and 0.77 percentage points, respectively. This simultaneous improvement is the strongest empirical argument for acceptance, although its weight depends on resolving the evaluation and baseline inconsistencies.

Critical weaknesses

1.The reported AutoAttack metric does not establish combined robustness. Table 1 defines overall AutoAttack accuracy as the minimum accuracy across individual attacks. Standard combined evaluation instead counts a sample as robust only if every attack fails; these quantities need not coincide. Consequently, the reported metric can overestimate robustness, leaving the paper’s central accuracy–robustness claims unverified.

2.The main comparison omits stronger baseline configurations reported in the appendix. On Imagenette, Appendix Table 2 reports pretrained PGD-AT with 88.15% clean and 63.95% APGD-DLR accuracy, exceeding Soft-kNN’s 87.87% and 62.50% on both measures.Yet the main table uses a weaker PGD-AT configuration. This directly weakens the claim of consistently superior trade-offs and raises concerns about how representative configurations were selected.

Questions for authors

1.Were the reported AutoAttack results computed through per-sample aggregation of attack successes, or by taking the minimum of individual attack accuracies? If the latter, what are the combined AutoAttack results, and do the claimed improvements persist?

2.What selection rule determined the configurations in Table 1? In particular, why was the stronger Imagenette PGD-AT configuration in Appendix Table 2 omitted, and does RepAlign improve the Pareto frontier when all evaluated baseline configurations are included?

3 comments

We thank the reviewer for their positive evaluation of the paper and their pertinent concerns.

  1. "...what are the combined AutoAttack results ...": This concern is addressed in the common concerns along with those of the other reviewers. In summary, we find only minor changes in the robust accuracy values and our claims about pushing the Pareto frontier forward remain valid.

  2. "...stronger Imagenette PGD-AT configuration...": This concern is addressed in the common concerns along with those of the other reviewers. In short, we find that CKNNA achieves a trade-off with TRADES which using the PGD-10 attack, with neither dominating.

The authors’ rebuttal has addressed my concerns, and I will keep my original score.

AuthorsOpenReview ↗

Thank you for the response

Dear Reviewer c2EX,

Thank you for your response. We have added additional experiments and updated the paper as well. Since we have addressed all of your concerns, we wonder if you would consider increasing your score. Thanks again for considering our request.

Regards,
The Authors

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

The paper proposes RepAlign, a regularizer for adversarial training. It trains with cross-entropy on PGD adversarial examples (as in PGD-AT) and adds a term that maximizes a representation similarity metric between the penultimate-layer features of clean and adversarial batches. Four metrics are tested: CKA, HSIC, CKNNA and a differentiable Soft-kNN. Unlike TRADES, the alignment is computed at the batch level on relational similarity matrices rather than per-sample output distributions. On CIFAR-10, CIFAR-100 (WRN-28-10) and Imagenette (ResNet-50) under ℓ∞ 8/255, RepAlign improves clean accuracy by about 2–5 points over TRADES with similar robust accuracy (e.g., Soft-kNN on CIFAR-10: 86.90% clean, 49.37% worst AutoAttack component, vs. TRADES 82.85% and 50.01%). CKA heatmaps are used to argue that RepAlign preserves the layer structure of clean models, and training cost is similar to PGD-AT.

Critical strengths
  1. The paper explores a simple and practical idea: aligning intermediate representations of clean and adversarial inputs with representation similarity metrics, instead of only their output distributions. The method is easy to plug into standard adversarial training, gives a clear gain in clean accuracy at a training cost similar to PGD-AT (and about 25% lower than TRADES), and the layer ablation offers a useful finding that alignment works best at the penultimate layer.

  2. The empirical study is broad and transparent. It covers three datasets and several white-box and black-box attacks, including Square and SPSA to check for gradient masking, and reports full hyperparameter sweeps, baseline variants (scratch vs. pretrained, 10 vs. 20 steps), training time, released code, and a representation analysis that helps interpret the method’s behaviour.

Critical weaknesses
  1. The robustness evaluation is not standard AutoAttack. The paper runs the AutoAttack components separately and reports the lowest of the four accuracies. Standard AutoAttack counts a sample as robust only if it survives all attacks, so the true AutoAttack accuracy is usually lower. The reported robustness is likely overestimated and not comparable with RobustBench.

  2. The gains are mainly in clean accuracy, not robustness. On CIFAR-10, RepAlign's robust accuracy (49.37%) is lower than TRADES (50.01%), and on Imagenette it is lower than PGD-AT (62.52%). Since TRADES with a smaller β also gives higher clean accuracy (Table 10: 87.46% at β = 1), the claim of a strictly better Pareto frontier needs full Pareto curves for both methods under the same evaluation.

  3. The experimental protocol raises concerns. Baselines also seem weaker than commonly reported (TRADES AutoAttack around 50% on CIFAR-10), possibly due to robust overfitting without early stopping, and common baselines such as MART and AWP are missing.

  4. Results are very unstable across λ (e.g., Soft-kNN on CIFAR-10 drops to 81.91% clean at λ = 1 but reaches 86.91% at λ = 5; HSIC on Imagenette collapses to 19.21% at λ = 5).

  5. The novelty is limited. The core change is to replace the KL divergence in TRADES with existing metrics. CKA (Kornblith et al., 2019), CKNNA (Huh et al., 2024) and HSIC (Gretton et al., 2007) are all established representation similarity metrics, and Soft-kNN is adapted from the existing soft nearest neighbor loss (Frosst et al., 2019). The paper does not propose a new metric or a new training framework; it mainly plugs existing metrics into adversarial training as a regularizer.

Questions for authors
  1. Could the authors report standard AutoAttack accuracy (the full ensemble, per-sample worst case) for all methods?

  2. Could the authors plot full clean–robust Pareto curves for TRADES (sweeping β) and RepAlign (sweeping λ) under the same evaluation?

  3. Could the authors compare against stronger baselines (TRADES and PGD-AT with early stopping, MART, AWP) or RobustBench entries with the same architecture?

  4. Could the authors explain why aligning batch-level relational structure improves robustness, given that the metrics are invariant to rotation and scaling of the features?

4 comments

We thank the reviewer for their careful feedback and address their specific concerns below.

  1. "...report standard AutoAttack accuracy...": This concern is addressed in the Common concerns along with those of the other reviewers. In summary, we find only minor changes in the robust accuracy values and our claims about pushing the Pareto frontier forward remain valid.

  2. Pareto improvement: With standard AA, CIFAR-10 Soft-kNN gives 86.90% clean / 49.36% AA, compared with TRADES at 82.85% / 50.00%: higher clean accuracy, but lower robustness. On CIFAR-100, Soft-kNN improves both measures over the Table 1 TRADES checkpoint: 60.42% / 25.97% versus 55.79% / 25.32%, gains of 4.63 and 0.65 percentage points. On Imagenette, the PGD-10 model reported in the Appendix reaches 88.15% / 63.90%, exceeding Soft-kNN's 87.87% / 62.50% on both measures. We will include that stronger baseline and qualify the claim of consistent frontier improvement. All Pareto plots are in Appendix G.2.

  3. Checkpoint selection: This is addressed in the Common concerns. In summary, we pick checkpoints that have the highest robust accuracy against the PGD attack on the validation data during training, which works better than early stopping.

  4. "... compare against stronger baselines (TRADES and PGD-AT with early stopping, MART, AWP) or RobustBench entries with the same architecture ...": For training our models we use multi-step learning rate and determine the best model, as noted above, using robust validation accuracy, which provides a stronger baseline than early stopping. We could not provide exact numbers corraborating with the RobustBench benchmark as most of their published models use additional data or synthetic generation which is out of the scope of the paper. We would also like to thank the reviewer for pointing out MART whose results have now been integrated into our baselines. We find that MART does not lie on the Pareto frontier for any dataset. Please refer to Appendix Section G.3 for details and Updated Table 1 of the common concerns.

  5. "...why aligning batch-level relational structure improves robustness...": This is a nuanced question that we provide an explanation to in the Why does RepAlign work? section of the Common concerns.

  6. Sensitivity to λ\lambda variation: We thank the reviewer for this nuanced question. We address this in the Hyperparameter sensitivity section (Point 4) of the Common concerns. As the representation similarities are bounded, and the loss is not, we use the hyperparameter λ\lambda to scale it to match the robust loss in magnitude and find this significantly improves our baselines.

  7. "...novelty is limited...": We note that our method is not merely a direct adaptation of TRADES, which minimises the loss on benign data and adds a KL-divergence term between the softmax outputs for benign and perturbed data. We instead minimise the loss over perturbed data ('robust loss'), and add a differentiable alignment term between the representations of benign and perturbed data. We agree that we do not propose a new representation similarity term, but the primary purpose of this paper is to show that using other RS metrics can push the Pareto frontier forward.

Dear Reviewer RoCP,

Thank you once again for your helpful review. We believe we have addressed all of your concerns in the updated paper, common rebuttal and the rebuttal specific to you. Please let us know if any concerns remain. We would greatly appreciate a response before the end of the rebuttal period.

Regards,
The Authors

Dear Reviewer,

The authors have kindly provided detailed responses to your comments. We would appreciate it if you could take a moment to review their rebuttal and share any further feedback or updates to your evaluation.

Thank you for your time and thoughtful consideration.

Dear Reviewer RoCP,

We have updated the paper in response to your feedback. Please let us know if any concerns remain or if anything needs further clarification. Thank you for your time.

Regards,
The Authors

Comments on the paper

7 comments
AuthorsOpenReview ↗

Common Response: Part 1

We thank the reviewers for their detailed feedback on the paper. We are glad that the reviewers found RepAlign to be a simple and practical idea that is easy to add to standard adversarial training (RoCP, c2EX, UnAD), with a training cost similar to PGD-AT (RoCP, c2EX, UnAD). We appreciate that the reviewers recognize the clear gain in clean accuracy (RoCP) and the improvement in both clean accuracy and robustness on CIFAR-100 over every PGD-AT and TRADES configuration (c2EX, UnAD). This simultaneous improvement was seen as the strongest empirical argument for acceptance (c2EX). We also appreciate that the reviewers found the empirical study to be broad (covering three datasets and various white-box and black-box attacks, and also hyperparameter sweeps) and transparent (RoCP, UnAD). Finally, we thank the reviewers for highlighting the following useful findings from our analyses:

  1. Alignment works best at the penultimate layer. (RoCP)
  2. Local metrics (CKNNA, Soft-kNN) consistently outperform the global ones (CKA, HSIC). (UnAD)
  3. Representation analysis helps interpret the method's behaviour. (RoCP)

List of updates in the paper: Note: We have added to the manuscript a rebuttal section (Appendix G: Rebuttal) with only relevant rebuttal figures at the very end after the current appendix. The main manuscript will be updated at a later stage.

  1. Added a Autoattack column for Tables 2,3,4,5,6,7,8,9,11 in the Appendix.
  2. Appendix G to be used for Rebuttal:
    • G.1. Updated Table 1 with Autoattack column (requested by Reviewers RoCP, c2EX and UnAD)
    • G.2. TRADES vs RepAlign parameter sweep plots (requested by Reviewers RoCP and UnAD)
    • G.3. Additional baseline MART (requested by Reviewers RoCP and UnAD)
    • G.4. Training Dynamics with respect to different λ\lambda
AuthorsOpenReview ↗

Common Response: Part 2

1. Standard AutoAttack (RoCP, c2EX, UnAD): As all reviewers noted, the paper used a slightly non-standard AutoAttack, reporting the lowest robust accuracy across attacks instead of checking whether a sample was misclassified under any attack. We have revised the paper with standard AutoAttack results, shown below.

We ran ℓ∞\ell_\infty, ϵ=8/255\epsilon=8/255 on the full evaluation sets for PGD-AT, TRADES, and all four RepAlign variants. On CIFAR-10, Soft-kNN achieves the best robustness-accuracy tradeoff, raising clean accuracy by over 4 percentage points while losing only 0.64 points in robust accuracy. On CIFAR-100, its 60.42% clean / 25.97% AA exceeds the TRADES checkpoint's 55.79% / 25.32%, gains of 4.63 and 0.65 percentage points. These are means over three attack seeds on fixed checkpoints, not independent training runs.

Updated Table 1: Clean & robust accuracy (%) on CIFAR-10, CIFAR-100, and Imagenette (ℓ∞\ell_\infty, ϵ=8/255\epsilon=8/255). Results reflect the best configuration per method with sample standard deviation over 3 attack seeds on a fixed checkpoint where available. AA denotes the sequential AutoAttack ensemble; APGD-CE, APGD-DLR (targeted), FAB-T (targeted), ART-PGD, Square, and SPSA are white-/black-box diagnostics. AFA uses Track B (joint training) with pretrained initialization. Bold and underlined‾\underline{\text{underlined}} values mark best and second-best defense results (ties bolded).

DatasetMethodCleanAAAPGD-CEAPGD-DLRFAB-TART-PGDSquareSPSA
CIFAR-10Clean Training96.1796.170.000.000.000.000.000.000.000.000.020.020.990.990.250.25
PGD-AT83.1583.1546.28±0.0146.28\pm0.0149.17±0.0549.17\pm0.0546.31±0.0146.31\pm0.0146.7646.7649.8749.8754.64±0.0554.64\pm0.0566.88±0.0166.88\pm0.01
TRADES82.8582.8550.00±0.01\bf 50.00\pm0.0152.60±0.03‾\underline{52.60\pm0.03}50.01±0.01\bf 50.01\pm0.0150.57\bf 50.5753.10‾\underline{53.10}56.35±0.0656.35\pm0.0667.63±0.0667.63\pm0.06
AFA83.6683.6648.59±0.0348.59\pm0.0351.03±0.0351.03\pm0.0348.60±0.0148.60\pm0.0149.16±0.0749.16\pm0.0751.66±0.0551.66\pm0.0555.80±0.0455.80\pm0.0467.48±0.0667.48\pm0.06
MART82.1682.1649.98±0.01‾\underline{49.98\pm0.01}54.94±0.04\bf 54.94\pm0.0450.01±0.02\bf 50.01\pm0.0250.37±0.01‾\underline{50.37\pm0.01}55.52\bf 55.5256.47±0.0456.47\pm0.0468.70±0.0768.70\pm0.07
CKA (Ours)84.3384.3348.40±0.0248.40\pm0.0250.92±0.0350.92\pm0.0348.42±0.0148.42\pm0.0148.9248.9251.6151.6156.56±0.0456.56\pm0.0468.20±0.0768.20\pm0.07
CKNNA (Ours)86.01‾\underline{86.01}48.70±0.0148.70\pm0.0150.56±0.0250.56\pm0.0248.70±0.0248.70\pm0.0249.5349.5351.4951.4957.65±0.0757.65\pm0.0769.71±0.04‾\underline{69.71\pm0.04}
HSIC (Ours)85.9285.9249.14±0.0349.14\pm0.0351.56±0.0551.56\pm0.0549.17±0.0149.17\pm0.0149.9149.9152.1752.1757.67±0.03‾\underline{57.67\pm0.03}69.63±0.0969.63\pm0.09
Soft-kNN (Ours)86.90\bf 86.9049.36±0.0249.36\pm0.0251.19±0.0551.19\pm0.0549.37±0.01‾\underline{49.37\pm0.01}50.0850.0851.8251.8257.95±0.07\bf 57.95\pm0.0770.19±0.02\bf 70.19\pm0.02
CIFAR-100Clean Training80.9880.980.000.000.000.000.000.000.040.040.020.020.430.430.190.19
PGD-AT56.7856.7822.90±0.0222.90\pm0.0226.22±0.0226.22\pm0.0222.90±0.0122.90\pm0.0123.2023.2026.7826.7828.64±0.0428.64\pm0.0439.87±0.0239.87\pm0.02
TRADES55.7955.7925.32±0.0125.32\pm0.0128.85±0.0528.85\pm0.0525.33±0.0125.33\pm0.0125.6825.6829.2029.2029.34±0.0629.34\pm0.0640.39±0.0440.39\pm0.04
AFA56.3556.3523.46±0.0223.46\pm0.0226.92±0.0126.92\pm0.0123.45±0.0323.45\pm0.0323.73±0.0123.73\pm0.0127.20±0.0227.20\pm0.0227.54±0.1527.54\pm0.1539.59±0.0639.59\pm0.06
MART55.0555.0526.07±0.01\bf 26.07\pm0.0130.62±0.07\bf 30.62\pm0.0726.07±0.01‾\underline{26.07\pm0.01}26.29±0.00‾\underline{26.29\pm0.00}31.02\bf 31.0230.18±0.08‾\underline{30.18\pm0.08}41.41±0.0341.41\pm0.03
CKA (Ours)57.0257.0224.03±0.0024.03\pm0.0027.61±0.0227.61\pm0.0224.02±0.0124.02\pm0.0124.2324.2328.3028.3029.68±0.1329.68\pm0.1340.88±0.0540.88\pm0.05
CKNNA (Ours)60.82\bf 60.8224.64±0.0124.64\pm0.0127.46±0.0527.46\pm0.0524.71±0.0124.71\pm0.0125.0025.0028.1128.1129.95±0.0529.95\pm0.0541.42±0.04‾\underline{41.42\pm0.04}
HSIC (Ours)60.0560.0522.96±0.0222.96\pm0.0225.13±0.0525.13\pm0.0522.99±0.0222.99\pm0.0223.36±0.0223.36\pm0.0225.72±0.0125.72\pm0.0128.30±0.0728.30\pm0.0739.36±0.0439.36\pm0.04
Soft-kNN (Ours)60.42‾\underline{60.42}25.97±0.00‾\underline{25.97\pm0.00}29.43‾\underline{29.43}26.10±0.01\bf 26.10\pm0.0126.49±0.01\bf 26.49\pm0.0130.36±0.02‾\underline{30.36\pm0.02}31.73±0.04\bf 31.73\pm0.0442.97±0.04\bf 42.97\pm0.04
ImagenetteClean Training91.8291.820.000.000.000.000.000.000.360.360.000.005.735.7339.5739.57
PGD-AT86.6586.6562.52±0.03\bf 62.52\pm0.0364.55±0.04‾\underline{64.55\pm0.04}62.52±0.03\bf 62.52\pm0.0362.93±0.03‾\underline{62.93\pm0.03}64.8764.8771.92±0.0171.92\pm0.0182.79±0.0382.79\pm0.03
TRADES84.6984.6961.66±0.0061.66\pm0.0065.06±0.01\bf 65.06\pm0.0161.6661.6661.94±0.0661.94\pm0.0665.43\bf 65.4370.06±0.1370.06\pm0.1380.36±0.0780.36\pm0.07
AFA77.1277.1247.95±0.0047.95\pm0.0051.70±0.0351.70\pm0.0347.97±0.0147.97\pm0.0148.75±0.4548.75\pm0.4551.99±0.0351.99\pm0.0356.73±0.1556.73\pm0.1573.72±0.0573.72\pm0.05
MART78.6878.6856.48±0.0056.48\pm0.0060.44±0.0460.44\pm0.0456.48±0.0056.48\pm0.0056.84±0.0056.84\pm0.0060.6660.6662.51±0.0462.51\pm0.0475.47±0.0675.47\pm0.06
CKA (Ours)87.5787.5761.88±0.0161.88\pm0.0163.75±0.0163.75\pm0.0161.89±0.0161.89\pm0.0162.35±0.0362.35\pm0.0364.4364.4372.97±0.0372.97\pm0.0383.52±0.0383.52\pm0.03
CKNNA (Ours)88.46\bf 88.4661.89±0.0161.89\pm0.0163.60±0.0163.60\pm0.0161.89±0.0161.89\pm0.0162.33±0.0462.33\pm0.0464.4864.4873.82±0.04‾\underline{73.82\pm0.04}84.31‾\underline{84.31}
HSIC (Ours)87.7587.7561.65±0.0161.65\pm0.0163.37±0.0463.37\pm0.0461.65±0.0161.65\pm0.0162.12±0.0362.12\pm0.0363.9063.9072.62±0.1072.62\pm0.1083.92±0.0183.92\pm0.01
Soft-kNN (Ours)87.87‾\underline{87.87}62.50±0.00‾\underline{62.50\pm0.00}64.54±0.0164.54\pm0.0162.50‾\underline{62.50}63.05±0.05\bf 63.05\pm0.0565.17‾\underline{65.17}73.87±0.07\bf 73.87\pm0.0784.33±0.04\bf 84.33\pm0.04
AuthorsOpenReview ↗

Common Response: Part 3

  1. Stronger Imagenette PGD-AT configuration (c2EX, UnAD): Table 1 reports robust accuracy values for our standard configuration of using 20 PGD steps during training for PGD-AT, TRADES, and RepAlign. We then chose a single λ\lambda per dataset after doing sweeps (see Appendix Section C). We used λ=5\lambda=5 for CIFAR-10, λ=10\lambda=10 for CIFAR-100, and λ=1\lambda=1 for Imagenette and used it across all four RS metrics. Each run's checkpoint was selected by its highest robust (PGD) accuracy. The 10-step Imagenette baseline was outside this common training setting. After doing standard AutoAttack ensemble as mentioned above we obtained 88.15% clean / 63.90% AA on PGD-10 on Imagenette using TRADES. This outperforms the selected Soft-kNN model on both measures. However, CKNNA remains on the observed frontier: it achieves 88.38% clean / 62.55% AA with 10-step training and 88.46% / 61.89% with 20-step training. PGD-10 offers higher robustness, while these CKNNA configurations offer higher clean accuracy; neither dominates the other. Considering the available baseline configurations and our sweeps over hyperparameters, RepAlign pushes frontier points on CIFAR-10 and CIFAR-100 at both training-step budgets (10 & 20). We will update the tables and Pareto plots in the paper to reflect this.

  2. Checkpoint selection (RoCP, UnAD): For PGD and TRADES models we evaluate the checkpoint with the highest robust accuracy under PGD attacks during training. For example, for CIFAR-10, TRADES gets the highest PGD robust accuracy of 53.21% at epoch 101 and decreases to 47.64% at last epoch. Thus, we use the model at epoch 101 for all black-box and white-box evals. The selected checkpoint gives 82.85% clean / 50.00% standard AA. We will document this selection rule and believe it addresses the concerns about evaluating final-epoch baselines as well as those around early stopping. We will clarify this in the updated draft.

  3. Hyperparameter sensitivity (RoCP, UnAD): Because most of the representation similarity metrics are bound between [0,1][0,1], the adversarial loss dominates the total loss, thus it is necessary to scale the representation similarity score being used so as to have a reasonable term to control the feature drift. In order to find a suitable hyperparameter we ablated with different values of λ\lambda to find the ones on the Pareto frontier. We have also added plots for comparing our methods and TRADES with the Pareto frontiers clearly marked. See the training dynamics figures in Appendix Section G.4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.