- 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: , where is the penultimate layer and 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 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 () 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 × 2 step counts (32 configurations), while TRADES is fixed at . 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 (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 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 , but Section 2.3 notes that linear HSIC is unbounded and scales with feature norm. Maximizing it () can be achieved by inflating feature norms, so the loss is unbounded below. This is consistent with the collapse of HSIC at 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 differs by dataset (1 on Imagenette, 5–10 on CIFAR), the CIFAR-100 optimum lies at the edge of the grid (), and CKA, HSIC and Soft-kNN collapse on Imagenette at (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, 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., ) evaluated with full AutoAttack on all three datasets, plotted as a curve against RepAlign's 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 Gram matrices and training used (e.g., results at ).
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
- 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.
We thank the reviewer for their careful and nuanced feedback and address their concerns below.
"...standard AutoAttack...": This concern is addressed in the
common concernsalong 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."...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 concernsalong 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.Hyperparameter choice: Table 1 reports all four RS metrics, each trained with 20 PGD steps. We chose λ from the appendix sweeps by comparing clean and robust accuracy, and used the same λ 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.
Independently trained models: The ± 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.
Checkpoint choice: This concern is addressed in the
Common concernsfor all reviewers. In summary, for both baselines and RepAlign, we evaluate the checkpoint with the highest robust accuracy under the PGD attack.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]. We have provided figures for Pareto frontiers in
Appendix G.2in the updated draft for your consideration.Matched clean accuracy comparisons: We ran additional TRADES β variation (β=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 variation (λ=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.2of the paper.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 concernsand more details inAppendix 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.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.
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.