acceptodds
Under review as a conference paper at ICLR 2027

Spectral and Gradient Alignment for Time-Series Adversarial Training

Abstract

Deep neural networks (DNNs) are vulnerable to small adversarial perturbations, which is concerning in high-stakes domains such as healthcare. Adversarial training (AT) is a standard defense, but it often degrades performance on clean test data. For time-series, unconstrained perturbations can introduce frequency content that is inconsistent with the acquisition process and sensor bandwidth, yielding a poor training prior. We propose an adaptive filtering method that generates adversarial perturbations regularized toward each input’s spectral distribution without requiring application-specific frequency cutoffs. Across nine datasets spanning five health-related applications, our method improves clean-test performance over baselines while retaining competitive robustness.

Primary area:
learning on time series and dynamical systems
Keywords:
Adversarial training, time series, robustness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VioaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an adaptive spectral filtering method for time-series adversarial training to improve clean-data performance while preserving robustness. It uses KL divergence for spectral alignment and zero-phase filtering to preserve gradient direction. Experiments on nine datasets demonstrate improved clean-test performance and competitive adversarial robustness.

Critical strengths
  1. Addressing the trade-off between clean accuracy and adversarial robustness is particularly important for time-series applications. By introducing input-adaptive spectral filtering, this work offers a practical way to generate adversarial perturbations that better preserve the spectral characteristics of the original signals.

  2. A notable strength of this work lies in its extensive experimental evaluation across nine datasets and five health-related applications. The consistent improvements in clean-test performance, together with comprehensive ablation studies and robustness evaluations, provide convincing evidence of the method’s effectiveness.

  3. Beyond empirical performance, the authors establish a theoretical justification for zero-phase filtering and conduct physician-based assessments of perturbation realism. These complementary analyses strengthen the reliability of the proposed method and support its potential applicability to physiological time-series data.

Critical weaknesses
  1. The Introduction does not adequately establish the rationale behind the proposed method. It remains unclear why adaptive filtering is preferred over conventional filtering approaches and how spectral alignment contributes to improving adversarial training. These aspects should be more clearly motivated to strengthen the logical foundation of the work.

  2. Table 3 shows that the proposed method performs worse than some baselines on certain datasets, particularly Chapman, but these results are not sufficiently discussed. Moreover, the statement “Our method improves clean-test performance over baselines” in Section 4 appears overly general, as the proposed method does not consistently outperform all baselines. The authors should provide further analysis of these limitations and make their claims more precise.

  3. The Conclusion mainly highlights the advantages of the proposed method without adequately discussing its limitations. A more balanced discussion of potential weaknesses and scenarios where the method may be less effective would strengthen the paper.

  4. Several inconsistencies can be found in the reference list, including variations in citation styles and missing publication details. A more consistent and standardized bibliography would improve the paper’s presentation and readability.

  5. The novelty of the proposed method remains somewhat unclear, particularly in comparison with existing approaches such as FA and IFAP, which also consider input-dependent frequency constraints. While the authors employ a learnable spectral filter and KL-based regularization, it is not sufficiently explained how these designs offer fundamentally new contributions beyond prior work. Further clarification of the methodological differences would help establish the originality of the approach.

  6. While Theorem 2.2 demonstrates that zero-phase filtering preserves the original perturbation direction by minimizing angular deviation, it remains unclear how this property translates into improved adversarial effectiveness. The authors should further explain the relationship between gradient-direction preservation and adversarial loss maximization to provide stronger theoretical support for the proposed method.

  7. A gap remains between the theoretical formulation and the practical optimization procedure. While Equations (1) and (2) define a spectrally constrained minimax objective, the implemented method relies on a soft KL regularization term without explicitly enforcing the spectral constraint. The authors should further clarify how the practical algorithm approximates the proposed constrained objective.

Questions for authors
  1. The authors should further clarify the methodological novelty of the proposed approach compared with existing input-dependent frequency methods, particularly the contributions beyond learnable filtering and KL-based regularization.

  2. Further analysis is needed to explain why the proposed method performs worse than PGD on certain datasets and whether these results reveal potential limitations of the spectral alignment strategy.

  3. It remains unclear how the minimum-angle property contributes to adversarial loss maximization and to what extent the practical soft KL regularization approximates the proposed constrained optimization objective.

Reviewer_f4aCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies adversarial training (AT) for health-related time series. The authors argue that norm-bounded attacks such as PGD place perturbation energy at frequencies that are nearly absent from the input, which reduces clean accuracy. They therefore filter the input gradient with a zero-phase filter whose magnitude response is produced by a small MLP from the normalized input spectrum and trained to maximize the classifier loss, with a KL penalty that pulls the perturbation spectrum towards the input spectrum. Proposition 2.1 gives a Pinsker bound on the spectral mass, and Theorem 2.2 shows that zero phase minimizes the angle to the gradient for a fixed magnitude response. Experiments cover nine datasets, twelve baselines, several attacks, ablations and a realism rating by two physicians. The main claims are better clean performance than all baselines and "competitive" robustness.

Critical strengths
  1. A broad and relevant empirical study with nine datasets and many baselines, including the recent spectral methods IFAP, F-SAT and f-VAT. Clean performance improves on most datasets. The gains are not universal (PGD is slightly better on Chapman, Aug. has the best F1 on PhysioNet), but for health applications this is potentially meaningful, provided that the gain can be attributed to the method (Weakness 2).
  2. The method is easy to implement and, unlike DARG, needs no generative model, and the physician study attempts to measure plausibility directly.
Critical weaknesses
  1. The theory gives little support for the central claim, and the novelty is incremental. Theorem 2.2 follows in one line from Parseval's theorem and says nothing about the iterative, rescaled and projected procedure that is actually run. Proposition 2.1 is Pinsker's inequality and applies only if the KL divergence is below κ\kappa, which is never enforced (the constraint is replaced by a soft penalty with weight 1). Moreover, DKL(p∥q)D_{\mathrm{KL}}(p\|q) is mode-seeking, so concentrating the perturbation on one strong input frequency already yields a small value, and the filter sees only the input spectrum, not the gradient spectrum. Compared with FA (Jia et al., 2022), IFAP (Yoshida & Okuda, 2025) and the other spectral baselines, the new elements are the learned magnitude response and the zero-phase choice.
  2. The effect of spectral alignment is not convincingly isolated, and the robustness claim is overstated.
    • Ablations: on IEEE SPC22 (Table 6a), all ablated variants are far below standard training and PGD (e.g., "Sign only" reaches ρ=46.10\rho=46.10 vs. 80.79 for PGD in Table 1), the DaLiA "Sign only" row repeats the values of the SPC22 "Sign + Filter" row, and no standard deviations are given.
    • Budgets: the methods use different ϵ\epsilon, norms and numbers of steps (Appendix C.4.2). The matched comparisons (Appendices D.2, D.5 and E.4) either omit the spectral baselines or show differences mostly within one standard deviation, so the clean gain is hard to attribute to the spectrum.
    • Robustness: TRADES is more robust under ℓ2\ell_2-PGD on Chapman and PhysioNet (Table 4), and TRADES or DARG is slightly ahead under CW on all four datasets. The advantage appears mainly under f-PGD/f-CW, whose passband is close to the method's own prior. The abstract says "competitive" robustness, whereas the contributions claim "improved" robustness.
Questions for authors
  1. What does the theory guarantee for the procedure that is actually run (Weakness 1)? Can you report the achieved DKL(p∥q)D_{\mathrm{KL}}(p\|q) for all datasets next to PGD, SAP, DARG and IFAP (Table 27 and Appendix E.2.1 cover only some datasets and methods), and explain why the mode-seeking KL is the right alignment measure?
  2. Why are the ablated variants in Table 6 so much worse than PGD and standard training in Table 1, and is the DaLiA "Sign only" row in Table 6(a), which repeats the values of the SPC22 "Sign + Filter" row, a typo? Were the budget and the number of steps the same as for Table 1? Please add standard deviations.
  3. Does the clean-accuracy gain hold at a matched ℓ2\ell_2 budget, with the strength sweep of Appendix D.5, against DARG, IFAP, FA, F-SAT and f-VAT? Given Table 4, do you claim improved robustness or competitive robustness together with better clean accuracy?
Reviewer_jHMhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes an input-conditioned spectral filter for adversarial training on physiological time series data. A small MLP maps each input’s magnitude spectrum to a filter with zero phase. The filter is applied to the loss gradient during iterative l2 ascent and trained with a negative cross-entropy loss plus a KL term between the perturbation spectrum and the input spectrum. Two theoretical results are given: a Pinsker-based bound on spectral mass, and a proof that zero phase minimizes the angle to the gradient. Across nine datasets in five health tasks, the method gives the best average clean performance, competitive robustness, and the highest realism ratings in a blinded study with two physicians.

Critical strengths
  1. The paper targets a well-motivated problem: losing clean accuracy under adversarial training matters in health applications, and the spectral mismatch of PGD perturbations is clearly shown.

  2. The method is simple and cost-effective. There's no generative model and no hand-set frequency cutoffs per application.

  3. Empirical studies covers broad evaluation (9 datasets, 12 baselines) with useful extra analyses, especially the controlled energy-allocation experiment.

Critical weaknesses
  1. Robustness evaluation is weak: PGD-l2 adversarial training is missing from the robustness tables; The “white-box” A-PGD attack is often weaker than plain PGD, even against the undefended model, and it uses the proposed method’s own filter. It therefore does not support the claim that the advantage persists under an adaptive attack; and AutoAttack results are not reported.

  2. Ablations look under-tuned. On SPC22, most ablated variants reach 50, far below standard training (81). Table 6 has no standard deviations.

3 . Other experiment issues: The method is effectively a data augmentation, so it should be compared with frequency-domain augmentations, AdvProp, and random noise shaped to the input spectrum. There is no control condition (e.g. unperturbed signals), raters knew the study objective, and the perturbation strength is not reported. The gains are small. With only 3 seeds, many margins are within one standard deviation, and PGD beats the method on Chapman. The significance tests treat correlated metrics as independent and ignore seed variance. Only generation time is reported - filter training needs second-order backpropagation through the unrolled attack, so the training cost should be reported.

  1. The theory part is trivial. Prop. 2.1 is Pinsker’s inequality, and Thm. 2.2 is close to trivial. The method does not solve the constrained objective in Eqs. 1–2. At the KL values actually reached (k ˜ 0.36–1.38), the bound is nearly vacuous.

  2. Possible copy errors. DARG on UCIHAR is identical to Aug. (Table 2) for the three metrics.

Questions for authors
  1. Could you report per-example worst-case accuracy over all attacks, including AutoAttack-l2?
  2. Does the zero-phase choice matter? How does a random-phase filter with the same magnitude response perform?
  3. How does the method compare with random noise shaped to the input spectrum at the same norm?
  4. What perturbation strength did the physician study use, and how would raters label unperturbed signals?
  5. What are the full training time and peak memory compared with PGD adversarial training?

Comments on the paper

1 comment
Area_Chair_MiSPOpenReview ↗

Discussion encouraged during this period!

Hi all reviewers,

We’re about halfway through the discussion period. This is a friendly reminder to check the authors' replies. If the authors have answered questions or responded to critiques in the reviews, please leave a comment for the authors!

Thank you so much!

Best, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.