acceptodds
Under review as a conference paper at ICLR 2027

Spectral-temporal Decoupled Sparse Attack on Parameter-efficient Time Series Forecasting

Abstract

Recent advances in time series forecasting (TSF) have embraced architectural minimalism and parameter efficiency, but still achieved competitive accuracy. However, extreme structural sparsity and drastically reduced parameter spaces are vulnerable to adversarial perturbations. In this work, we propose a spectral-temporal decoupled sparse attack on parameter-efficient (PE) TSF models, which simultaneously disrupts localized morphological shapes and dominant periodic trajectories. Extensive experiments show that our method successfully sabotages five state-of-the-art PE TSF models for a small perturbation budget (SNRdB) without any structural or parameter queries, even under six representative adversarial defense measures. It can be applied to non-query attacks on remote edge-devices, stress testing in critical infrastructure domains, and structural diagnosing on PE edge-device deployment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Spectral-temporal decoupled sparse attack, parameter-efficient, time series forecasting

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_xpdsOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes STDSA, a transfer-based adversarial attack on parameter-efficient time-series forecasting models that requires no queries to the target model and aims to disrupt both local temporal structures and dominant periodic patterns. Its main contributions are complementary temporal and frequency-domain surrogate pathways, segment-wise spectral gradient filtering, scale-preserving gradient blending, and theoretical analysis of forecasting perturbations and filtering defenses. Validation covers five forecasting models on ETTh1, Weather, and Exchange under six defense configurations.

Critical strengths
  1. The paper studies attacks that require no queries to the target forecaster and combines temporal and spectral surrogate pathways. This provides a meaningful approach to investigating transferable vulnerabilities when direct model access is unavailable.

  2. The paper provides bounds on forecasting perturbations and an analysis of how low-frequency perturbations can survive frequency filtering. This offers an analytical motivation for the spectral design beyond empirical performance, although its applicability to the evaluated architectures requires stronger validation and does not change my rejection recommendation.

Critical weaknesses
  1. The connection between parameter efficiency and vulnerability is not convincingly established. The paper attributes susceptibility to restricted parameter spaces and structural sparsity, but attacking five selected models does not isolate these factors as the cause. Controlled comparisons across model sizes or sparsity levels, with forecasting accuracy and attack budgets accounted for, are needed to support this central claim.

  2. The attack comparisons are too limited to establish a substantial methodological advance. The experimental section identifies VMI-FGSM with the proposed surrogate loss and random perturbations as comparisons. This leaves uncertainty about whether the gains arise from the surrogate design, the optimization procedure, or spectral filtering. Stronger transfer-attack baselines and simple low-frequency perturbations under matched computational budgets would provide a more convincing assessment.

  3. The defense results do not fully substantiate the proposed explanation. The paper motivates its spectral design partly through resistance to frequency filtering, whereas the six listed experimental defenses primarily involve smoothing, training, and regularization. Direct evaluation of explicit frequency-filtering defenses across cutoff settings, while measuring their clean-accuracy cost, would better connect the theoretical argument to the empirical findings.

  4. The claims of imperceptibility and stealth exceed the evidence provided. An ℓ∞ budget of 0.1 and an average SNR of 21.52 dB quantify perturbation magnitude, but do not establish that changes are operationally plausible or difficult to detect. In particular, smooth low-frequency shifts can still violate sensor constraints or be identified by temporal anomaly detectors. Without such validation, the practical stealth claim remains insufficiently supported.

  5. The evaluation does not adequately support the breadth of the deployment claims. Experiments on ETTh1, Weather, and Exchange provide useful benchmark evidence, but do not demonstrate attacks on remote edge devices or realistic critical-infrastructure workflows. Constraints such as access to only a subset of sensors, causal modification of streaming observations, and downstream decision impact are not established by the reported benchmark results. The practical conclusions should therefore be narrowed or supported by more representative experiments.

Together, these concerns leave the claimed structural insight and practical advantage insufficiently established for acceptance.

Questions for authors
  1. Could the authors clarify what evidence isolates parameter efficiency or structural sparsity as the cause of increased vulnerability, rather than general sensitivity to temporal and low-frequency perturbations? Controlled comparisons across model sizes or sparsity levels would materially affect my assessment of the central claim.

  2. Under identical surrogate models, perturbation budgets, and optimization budgets, how does STDSA compare with standard transfer attacks and simple low-frequency perturbations? Please clarify whether the reported ablations establish the additional benefit of spectral-temporal decoupling and adaptive gradient blending.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

Thank you for recognizing the importance of target-query-free attacks and the motivation for spectral-temporal surrogates. The revised manuscript adds experiments that directly examine the mechanisms underlying transfer, trained-surrogate attack comparisons, explicit frequency-filtering defenses, and restrictions on the information and channels available to the attacker. The resulting evidence supports a concrete finding: structured perturbations generated without target-model access transfer across forecasting architectures, remain effective across model capacities, and survive the evaluated input defenses.

1. Structural vulnerability and model capacity (Weakness 1; Question 1)

The capacity experiment identifies a vulnerability shared across scales of the same forecasting architecture. The revised Appendix A.10 evaluates three iTransformer widths containing 501,472, 1,789,280, and 6,724,192 parameters. Their Weather RMIs are 31.40%, 32.51%, and 29.78%, respectively. The attack therefore produces approximately 30% forecasting-error increases over a 13.4-fold parameter-count range. This result establishes persistence across capacity. The three values do not imply a monotonic relationship between parameter count and vulnerability.

The structural interpretation is further supported by the perturbation controls in Table A9. On the same supplementary Weather evaluation suite, random low-frequency perturbations give 0.12% mean RMI, random channel-constant offsets give 6.24%, informed channel-constant offsets give 24.36%, and the full attack gives 31.86%. Thus only concentrating energy at low frequencies does not reproduce the attack effect, and an informed mean shift leaves a further 7.50 percentage-point improvement available to the full construction. Parameter efficiency defines the studied forecasting family. The experiments identify its transferable sensitivity to particular structured directions. The capacity sweep, control experiments, and architecture-level propagation analysis together substantiate that mechanism-oriented finding.

2. Stronger transfer comparisons and the contribution of the construction (Weakness 2; Question 2)

The revised comparison includes trained-surrogate transfer, an optimizer comparison using the TSP loss, and explicit low-frequency and constant-offset controls. Tables A5 and A7 compare STDSA with Transfer-PGD from a trained DLinear surrogate. STDSA achieves the higher RMI in 14 of the 15 dataset–target pairs, including all five targets on ETTh1 and Weather. The five-target mean RMIs are:

DatasetDLinear Transfer-PGDSTDSAGain
ETTh19.7877%17.0963%7.3086 pp
Exchange29.0154%62.4299%33.4145 pp
Weather8.2502%30.0217%21.7716 pp

Table A6 additionally reports STDSA/VMI-FGSM mean RMIs of 17.10%/13.18% on ETTh1, 62.43%/56.16% on Exchange, and 30.02%/23.71% on Weather. These results establish a transfer advantage across three datasets and multiple target architectures.

The supplementary Weather control suite uses 10444 windows and an ℓ∞\ell_\infty budget of ϵ=0.1\epsilon=0.1. Table A9 isolates perturbation mechanisms: its 31.86% full-attack result exceeds both random low-frequency perturbations and the 24.36% informed-offset control, with improvements over the latter on four of five targets. Table A12 separately examines source construction: the fixed uniform Direct source obtains 21.15%, fixed Direct + MixLinear obtains 30.33%, the full trained dual source obtains 31.86%, and the adapted dense control of Liu et al. (2023) obtains 9.59%. These source comparisons and the original ETTh1 component ablations answer different aspects of the mechanism and are reported under their respective configurations. In Table A4, Full STDSA and Direct Pathway Only have similar ETTh1 means (17.09% and 17.07%), while their cross-target standard deviations are 1.42 and 4.70. Removing LPF gives 16.99%. The pathway combination therefore improves cross-target consistency in that experiment, and the isolated LPF contribution to its mean is 0.10 percentage points.

Appendix A.13 also directly measures the effect of raw/projected gradient blending: changing its coefficient from 0 to 0.75 changes 14.64% of gradient signs in the 16-window first-step diagnostic. This records a directional change in that raw/projected diagnostic. Together, the new controls show that the observed transfer effect depends on the constructed directions and source, beyond perturbation magnitude or random low-frequency energy alone.

AuthorsOpenReview ↗

Reply to Reviewer [2/2]

3. Explicit frequency-filtering defenses and clean-accuracy cost (Weakness 3)

Appendix A.10 directly evaluates full-window FFT low-pass defenses at six cutoffs. Each filter is applied to both clean inputs and fixed transferred adversarial inputs. The attack effect for each defended target is normalized by that target's own defended clean MSE:

=100\left(\frac{\mathrm{MSE}_{\mathrm{adv,def},m}} {\mathrm{MSE}_{\mathrm{clean,def},m}}-1\right).$$ Table A11 reports the mean of these five target-wise percentages. Across $K\in\\{12,24,48,96,192,360\\}$, defended RMI remains **29.09%–31.86%**. At $K=12$, the attack retains **30.95% RMI** while clean MSE increases by **36.23%**. At $K=24$, the corresponding figures are **29.09%** and **22.08%**. The aggressive filters therefore impose substantial clean forecasting costs while retaining substantial attack effects. Median/IQR preprocessing produces essentially the same mean RMI, changing 31.8576% to 31.8531%. These are direct measurements on the final adversarial inputs. The input filter uses full-lookback frequency coordinates. the attack's gradient processing uses folded-segment coordinates, as explicitly distinguished in Appendix A.10. # 4. Measured perturbation structure and the meaning of stealth (Weakness 4) **The revised evidence characterizes the actual attack output in both time and frequency.** Figure A2 presents a preselected Weather example's clean and adversarial inputs, their additive difference, and the final perturbation spectrum. The lowest twelve full-lookback rFFT bins contain **70.99%** of the Parseval-weighted perturbation energy in the displayed example. The all-window Weather evaluation gives approximately **70.05%** in the same bins (Appendix A.11). This measurement verifies low-frequency concentration after sign updates and clipping. The final perturbation is dense in time and retains energy outside those bins. The $\ell_\infty$ bound and SNR quantify amplitude, while the spectrum quantifies concentration. These are the properties demonstrated by the evaluation. Operational detection performance additionally depends on the sensor constraints, anomaly detector, and application-specific acceptance thresholds. Those quantities are distinct from amplitude and spectral concentration. # 5. Restricted sensor access and information available at attack time (Weakness 5) **The new evaluation explicitly imposes restricted channel access during attack generation.** Table A10 applies channel masks inside every iterative update and evaluates the resulting attacks on the complete test sets. With access to one versus three channels, the five-target mean RMIs are 0.2682%/0.9188% on ETTh1, 3.3568%/10.4291% on Exchange, and 0.9356%/2.7035% on Weather. Thus average transfer remains positive under partial access and increases with the number of controllable channels. Table A8 also evaluates attack objectives using only information available before the prediction horizon. Replacing the future trajectory with the last observed value gives mean RMIs of **2.47%, 21.51%, and 14.25%** on ETTh1, Exchange, and Weather. Using a detached clean-surrogate forecast gives **1.64%, 6.74%, and 7.78%**, respectively. Weather and Exchange use reconstructed MixLinear + Linear sources. ETTh1 uses MixLinear only. In these experiments, future ground truth is used exclusively for evaluation, and generation uses neither target queries nor target gradients. The experiments therefore establish forecast-level transfer under partial channel control and future-label-free objectives. Remote-device deployment and downstream critical-infrastructure decisions introduce separate system-level conditions. The presented evidence supports using STDSA to stress-test the forecasting component under these explicitly defined access conditions.
AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in One Day]

Dear reviewer xpds,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only one day, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in 12 hours]

Dear reviewer xpds,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only 12 hours, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Reviewer_7SzUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Spectral-Temporal Decoupled Sparse Attack (STDSA), a non-query black-box ℓ∞\ell_\infty adversarial attack targeting parameter-efficient (PE) time series forecasting (TSF) models. STDSA optimizes perturbations against a dual-pathway surrogate loss combining a uniform linear mapping (Wdir=1L1H×LW_{dir} = \frac{1}{L}\mathbb{1}_{H \times L}) and a periodic low-pass filtered (LPF) pathway.

Critical strengths
  1. The paper identifies how parameter-saving mechanisms—such as channel-shared basis weights (WbaseW_{base} in TimeBase), position-only Gumbel-Softmax routing, and residual connections in periodic folding (SparseTSF)—preserve and amplify low-frequency perturbations.

  2. STDSA is evaluated across five distinct PE TSF paradigms and six defense mechanisms (TV purification, Randomized Smoothing, Adversarial Training, MSE Consistency, KL Consistency, and a combined defense), demonstrating that standard temporal defenses fail against low-frequency drift perturbations.

Critical weaknesses
  1. The Figures are likely generated by GenAI, especually for Figures 1 and 2.

  2. Does STDSA use the ground-truth future sequence YY at test time in Eqs. (5) and (8)? How does STDSA perform if YY is replaced by the surrogate's clean prediction Y^cln\hat{Y}_{cln} so that no future ground truth is leaked?

  3. Compare against standard black-box transfer attacks optimized on trained lightweight TSF surrogates (e.g., DLinear or FITS) and prior TSF attacks (e.g., Liu et al., 2023).

Questions for authors

Given the minimal gap between Full STDSA (17.09%17.09\%), Direct Pathway Only (17.07%17.07\%), and Without LPF (16.99%16.99\%) on ETTh1 (Table A4), do the periodic pathway and LPF provide larger gains on Weather and Exchange?

Evaluate defenses that use median/IQR normalization to mitigate simple lookback mean shifts.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/2]

We thank the reviewer for recognizing the structural attack analysis and the transfer evaluation across five forecasting architectures. The revised manuscript directly addresses future-label availability, adds trained-surrogate and prior-work comparisons, and evaluates both structured perturbation controls and median/IQR preprocessing. The principal evidence is reported in Appendices A.9–A.13.

1. Figures 1 and 2 and the disclosed use of generative AI

As disclosed in the manuscript's AI Use Statement, generative AI was used only for grammar correction, minor language editing, and collecting and summarizing experimental data. It was not used for other tasks, including generating Figures 1 and 2. In fact, we draw them with .pptx and convert them into .pdf. Figure 1 depicts the attack computation, while Figure 2 depicts the propagation of perturbations through TimeBase. Their roles are specified in their captions and developed in Sections 3.1–3.4. The revision also provides the actual clean and adversarial input sequences, their difference, and the final perturbation spectrum in Figure A2 and Appendix A.11. Thus the method diagrams are accompanied by measurements of the generated perturbations themselves.

2. Attack generation without future ground-truth labels

The added experiments establish positive average transfer when future ground truth is unavailable to the attacker. The original benchmark optimizes its MSE objective against the future sequence YY. Appendix A.9 and Table A8 additionally evaluate the two generation-time alternatives requested by the reviewer: repeating the last observed value over the forecast horizon, and using a detached clean-surrogate prediction. In these experiments, ground-truth future labels enter only the evaluation of forecasting errors. Attack construction uses the observed lookback and the source model, with zero target-model queries and zero target-gradient access.

DatasetLast-observation proxy: mean RMI (%)Detached clean-surrogate proxy: mean RMI (%)
ETTh12.471.64
Exchange21.516.74
Weather14.257.78

Each entry averages five target-wise RMI values. Table A8 identifies the source configuration for each dataset: MixLinear only on ETTh1 and reconstructed MixLinear + Linear sources on Exchange and Weather. These results substantiate future-label-free attack generation under those configurations and quantify its transfer strength directly.

3. Trained lightweight surrogates and the forecasting-attack baseline

The revised comparison includes transfer-PGD from a trained DLinear surrogate and an adapted dense forecasting-attack control based on Liu et al. (2023). Tables A5 and A7 evaluate trained-DLinear transfer across all fifteen dataset–target pairs. STDSA achieves higher RMI in fourteen pairs, with the following arithmetic means over five targets:

DatasetTrained-DLinear transfer-PGD RMI (%)STDSA RMI (%)Gain (percentage points)
ETTh19.787717.09637.3086
Exchange29.015462.429933.4145
Weather8.250230.021721.7716

The complete pairwise table establishes the breadth of this advantage across architectures. Table A6 also reports higher dataset-level mean RMI than the evaluated VMI-FGSM configuration on all three datasets.

For the requested prior forecasting-attack comparison, Appendix A.13 and Table A12 report the adapted dense control citing Liu et al. (2023). On the supplementary Weather suite, the full trained dual source reaches 31.86% mean RMI, compared with 9.59% for this control. Table A12 explicitly identifies these as supplementary source-construction comparisons. Its 31.86% value belongs to that evaluation suite, while Tables A5–A7 report the main cross-architecture benchmark. Together, the comparisons evaluate transfer from a trained lightweight forecaster, a prior-work-inspired dense control, and the previously included optimizer baseline.

AuthorsOpenReview ↗

Reply to Reviewer [2/2]

4. What the pathway and spectral controls establish

The added Weather experiments show a substantially larger advantage over direct-only and channel-offset constructions. Table A12 reports 21.15% mean RMI for the fixed uniform Direct source, 30.33% for Fixed Direct + MixLinear, and 31.86% for the full trained dual source. Thus, the fixed Direct + MixLinear construction exceeds the fixed Direct-only source by 9.18 percentage points, and the full trained dual source exceeds it by 10.71 points. These are source-construction comparisons, with the configurations identified in Table A12.

Table A9 provides an independent control on the alternative explanation that the attack is simply a channel-constant shift. On the supplementary Weather test set of 10444 windows with ϵ=0.1\epsilon=0.1, random low-frequency perturbations achieve 0.12% mean RMI, random channel-constant offsets achieve 6.24%, and Linear-gradient-informed channel-constant offsets achieve 24.36%. The full attack reaches 31.86%, exceeding the informed-offset control by 7.50 percentage points and improving on four of the five targets.

The ETTh1 result also contains information beyond its nearly equal means. In Table A4, Full STDSA and Direct Pathway Only achieve 17.09% and 17.07% mean RMI, while their cross-target standard deviations are 1.42 and 4.70, respectively. Full STDSA improves RMI on TimeBase, iTransformer, and SparseTSF, producing more consistent transfer across the five architectures.

The revision's component comparisons cover ETTh1 and the supplementary Weather setting. Exchange provides the cross-architecture transfer and future-label-free evaluations in Tables A5–A8. These results support the measured contribution of the full construction and the reported source combinations. The isolated LPF-removal result remains the ETTh1 comparison in Table A4.

5. Median/IQR normalization and explicit frequency filtering

Median/IQR preprocessing leaves the measured attack effect essentially unchanged. Appendix A.10 reports Weather mean RMI of 31.8576% before this preprocessing and 31.8531% afterward, a change of only 0.0045 percentage points. This directly evaluates the robust-normalization defense requested by the reviewer.

The revision additionally evaluates full-window FFT low-pass preprocessing over cutoffs K∈12,24,48,96,192,360K\in\\{12,24,48,96,192,360\\} in Table A11. Each defended model is normalized by its own defended clean MSE:

RMIdef,m=100(MSEadv,def,mMSEclean,def,m−1).\mathrm{RMI}_{\mathrm{def},m}=100\left(\frac{\mathrm{MSE}_{\mathrm{adv,def},m}}{\mathrm{MSE}_{\mathrm{clean,def},m}}-1\right).

Mean defended RMI remains between 29.09% and 31.86%. At K=12K=12, attack RMI is 30.95% while clean MSE increases by 36.23%. At K=24K=24, the corresponding values are 29.09% and 22.08%. These measurements establish substantial residual transfer under both robust normalization and explicit frequency filtering, with the clean-forecasting cost reported for the two strongest tested cutoffs.

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in One Day]

Dear reviewer 7SzU,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only one day, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

AuthorsOpenReview ↗

Your Comments are Highly Valuable to Us [Deadline in 12 hours]

Dear reviewer 7SzU,

We have carefully considered all the valuable points and concerns you raised in your initial review through the initial rebuttal. As the deadline for discussion is approaching in only 12 hours, we respectfully ask whether our response has addressed your concerns, or you still have questions to raise. Your comments are highly valuable to us. Thank you once again for your professional time and insightful effort in reviewing our submission.

Best regards,

Authors

Reviewer_HQRtOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes STDSA, a query-free transfer attack on parameter-efficient (PE) forecasting models. The surrogate loss mixes two fixed pathways: a direct pathway with the uniform averaging map WdirW_{dir}, and a periodic pathway that low-pass filters the input across period segments. At each step, the gradient is folded by a period P\mathcal{P}, low-pass filtered along the segment axis, rescaled and blended with weight α\alpha, and a sign step is taken within an ℓ∞\ell_\infty ball of radius ϵ\epsilon. Theorem 1 bounds the forecast error of a linear target, and Theorem 2 states that a frequency-filtering defense does not remove a low-pass perturbation. On ETTh1, Weather and Exchange, a single adversarial set per dataset raises the forecast MSE of five target models. The paper reports that STDSA outperforms Gaussian noise and VMI-FGSM with the same loss, and that its effect persists under six defenses.

Critical strengths
  1. The problem is relevant: PE forecasters are intended for edge deployment, and an attack that needs neither queries nor a trained surrogate is cheap to mount. A single adversarial set per dataset transfers to five architecturally different targets, and it is far stronger than Gaussian noise at a matched ℓ∞\ell_\infty budget or SNR (Table A3).
  2. The results are reported in detail. Clean, attacked and defended MSE are given for five targets, six defenses and three datasets (Tables 1, A1, A2), together with an ablation and a sensitivity analysis (Table A4, Fig. A1). The tables also show openly where defenses help, e.g., adversarial training (AT) recovers 13.04% of the attacked MSE on Exchange (Table A2).
  3. The attack-chain derivations (Sec. 3.4, App. A.3) make explicit how an input perturbation propagates through each target.
Critical weaknesses
  1. The ablations do not support the main contribution, the spectral-temporal decoupling. Since WdirW_{dir} is a fixed averaging map (Eq. 6), the direct-pathway gradient is constant over time, so this pathway alone yields a per-channel offset of ±ϵ\pm\epsilon whose sign depends on the true future YY. This direct-only variant matches the full method in mean RMI on ETTh1 (17.07% vs. 17.09%, Table A4), removing the low-pass filter or varying α\alpha changes little (Table A4, Fig. A1), and, as written, Eq. 13 makes the sign step independent of α\alpha. Four of the five targets normalize by the instance mean, so the offset passes into the forecast (Eqs. 40, 45), yet this simple baseline is not tested.
  2. The theorems do not cover the attack as implemented. Theorem 2 assumes that δ\delta is band-limited along time (Eq. 25), but the mask in Eq. 11 acts across segments, the sign step and clipping (Eq. 14) break band-limitedness, and none of the six defenses is a frequency filter. Theorem 1 assumes a linear, injective target, which RevIN, attention and gating violate; injectivity is also impossible on the full input space when L=720>H=96L=720>H=96. Both bounds hold for any perturbation, so they do not explain why STDSA in particular works.
  3. The threat model and baselines do not match the claims. The attack uses the true future YY (Eq. 8), which an attacker on an edge device does not have at forecast time. The only attack baseline is VMI-FGSM with the same loss and unspecified settings; white-box PGD, transfer from a trained surrogate and published forecasting attacks (e.g., Liu et al., 2023) are missing. The SNR of about 20 dB follows directly from ϵ=0.1\epsilon=0.1 (App. A.5.2), and apart from a schematic in Fig. 1, the paper shows no adversarial inputs or spectra of δ\delta to support the stealthiness claim.
  4. The defense evaluation is underspecified. The paper gives neither the attack used to generate training perturbations for the trained defenses (AT, MSE consistency, KL, combined) nor any defense hyperparameter. In the defended rows, RMI mixes the undefended model's attacked MSE with the defended model's clean MSE (e.g., Weather, iTransformer, AT: 38.90%), so it does not measure the attack on the defended model. On ETTh1 and Exchange, the trained defenses recover part of the error (e.g., AT AMR: 7.88–16.48% on Exchange, 20.30% for FITS on ETTh1), so the claim that STDSA persists under the defenses is overstated.
Questions for authors
  1. How does STDSA compare with a constant per-channel offset of ±ϵ\pm\epsilon (direct-pathway sign and random sign), white-box PGD on each target, and transfer from a trained surrogate? Is Eq. 13 meant to blend in the raw gradient G~(k)\tilde G^{(k)}?
  2. How does STDSA perform when YY in Eq. 8 is replaced by information available at attack time, e.g., a naive or surrogate forecast?
  3. Which attack generated the defenses' training perturbations, and what are the defense hyperparameters, P\mathcal{P} and the VMI-FGSM settings? Please report each defended model's RMI against its own clean MSE and show adversarial inputs with the spectrum of δ\delta. In what sense is the attack sparse, given that δ\delta is only ℓ∞\ell_\infty-bounded and may be nonzero at every entry?
6 comments
AuthorsOpenReview ↗

Reply to Reviewer [1/3]

Thank you for recognizing the relevance of query-free transfer attacks and the breadth of our evaluation. The additional experiments in Appendices A.9–A.13 address the central empirical questions with explicit offset controls, trained-surrogate transfer, future-ground-truth-free objectives, full-window filtering, final-perturbation spectra, and defended-clean-normalized errors. The evidence establishes strong transfer beyond a channel-constant shift and substantial residual forecasting degradation after defense. We address the mathematical and protocol questions separately below.

1. Constant offsets, structural pathways, and gradient blending

The revised Appendix A.13 characterizes the fixed uniform direct pathway exactly. With μc(X)=L−1∑tXt,c\mu_c(X)=L^{-1}\sum_t X_{t,c} and MSE averaged over the HCHC outputs,

=\frac{2}{LC}\left(\mu_c(X)-\frac{1}{H}\sum_hY_{h,c}\right).$$ This gradient is constant along time within each channel. The new Weather controls therefore directly test channel-constant perturbations. On 10444 windows with $\epsilon=0.1$, Table A9 reports mean target-wise RMI of **0.12% for random low-frequency perturbations, 6.24% for random channel-constant offsets, 24.36% for Linear-gradient-informed channel-constant offsets, and 31.86% for the full attack**. The full attack exceeds the informed offset by **7.50 percentage points**, with gains on four of five targets: 9.61 points on iTransformer, 8.07 on SparseTSF, 8.28 on FITS, and 12.29 on LSINet. These results identify a substantial transfer contribution beyond the channel-constant mechanism. The original ETTh1 ablation also contains a relevant cross-target distinction: the full and direct-only mean RMIs are 17.09% and 17.07%, but their cross-target standard deviations are **1.42 and 4.70**. The full attack raises the direct-only RMI from 12.22% to 15.80% on TimeBase, from 14.00% to 15.51% on iTransformer, and from 13.64% to 16.99% on SparseTSF (Table A4). The benefit is broader target coverage, alongside the larger Weather improvement. Table A12 further compares source constructions under the supplementary Weather settings: the full trained dual source attains **31.86%**, the fixed Direct–MixLinear combination 30.33%, and the fixed uniform direct operator 21.15%. These supplementary trained-source results are distinct from the original fixed-source ETTh1 ablation. Yes: the raw/projected blend described in Section 3.3 combines the raw gradient $\tilde G^{(k)}$ with the magnitude-restored projected gradient. **The Eq. (13) instead combines two collinear projected tensors and requires correction.** The separate first-step diagnostic in Appendix A.13 records a change in **14.64% of gradient signs** when the raw/projected coefficient changes from 0 to 0.75 across 16 windows. This gives direct directional evidence for the blend evaluated in that diagnostic. The equation must express those two distinct terms. # 2. Theorems: the precise mathematical scope and the measured spectrum Theorem 1 assumes injectivity of the **restricted linear operator** $\mathcal T_{\mathrm{tgt}}|_{\mathcal X}$. For $L>H$, that assumption can hold only on an appropriate subspace with $\dim\mathcal X\le HC$. It does not establish injectivity on the full $LC$-dimensional input space. Its conclusion is a conditional bound on prediction displacement under the stated linear assumptions. The architecture-specific analysis and transfer experiments provide the evidence for the evaluated nonlinear models and for increased forecasting MSE. Comparative effectiveness of STDSA is established by the transfer and control experiments. For a full-window ideal low-pass projection $P_K$, the general filtering identity is $$P_K(X+\delta)-P_KX=P_K\delta.$$ Exact preservation requires $P_K\delta=\delta$ in those same frequency coordinates. Segment-axis filtering of a folded gradient and full-window filtering of a final perturbation are different operations. Sign updates and clipping can also introduce higher frequencies. **The empirical property supported by the final attack is strong low-frequency concentration.** Appendix A.11 and Figure A2 measure the final perturbation itself: the lowest twelve full-window rFFT bins contain **70.99%** of its Parseval-weighted energy in the preselected Weather example and approximately **70.05%** across the evaluation set. Figure A2 displays the clean and adversarial inputs, their difference, and the actual final spectrum.
AuthorsOpenReview ↗

Reply to Reviewer [2/3]

The new explicit filtering experiment tests the resulting attack directly. Across full-window FFT cutoffs K∈12,24,48,96,192,360K\in\\{12,24,48,96,192,360\\}, mean defended-clean-normalized RMI remains 29.09%–31.86% (Table A11). At K=12K=12, residual RMI is 30.95% while clean MSE increases by 36.23%. At K=24K=24, the respective values are 29.09% and 22.08%. Thus substantial attack effects survive explicit spectral preprocessing, including settings with a large clean-accuracy cost.

3. Transfer when future ground truth is unavailable

The original objective uses YY as the benchmark reference. The revised Appendix A.9 separately evaluates two objectives constructed from information available before forecasting: the last observed value repeated across the horizon, and a detached clean-source prediction. Future ground truth is used only for evaluating these experiments.

DatasetSource configurationLast-value targetClean-source-prediction target
ETTh1MixLinear only2.47%1.64%
WeatherMixLinear + Linear14.25%7.78%
ExchangeMixLinear + Linear21.51%6.74%

These are arithmetic means of the five target-wise RMIs in Table A8. Positive average transfer remains achievable without future-label access, target prediction queries, or target gradients. The table records each source construction explicitly, so these results establish the available-information setting under their stated configurations.

4. Trained-surrogate, prior-attack, and white-box comparisons

The expanded benchmark transfers attacks from a trained DLinear source to all five targets on three datasets. STDSA wins 14 of 15 dataset–target pairs, with mean target-wise RMI increasing from 9.7877% to 17.0963% on ETTh1, from 8.2502% to 30.0217% on Weather, and from 29.0154% to 62.4299% on Exchange (Tables A5 and A7). Table A6 separately reports higher dataset-level mean RMI than the evaluated VMI-FGSM configuration on all three datasets. The adapted dense control associated with Liu et al. (2023) reaches 9.59% on the supplementary Weather suite, compared with 31.86% for the full trained dual source (Table A12).

Appendix A.13 also reports the requested target-specific white-box PGD reference: 242.65%, 790.78%, 50.37%, 52.46%, and 107.05% on TimeBase, iTransformer, SparseTSF, FITS, and LSINet, respectively. These attacks require each target's parameters and gradients, so their results are not directly comparable with our zero-query black-box transfer results because the two attacks operate under fundamentally different access assumptions. They quantify the effect of privileged target-gradient access; the trained-source and structured-control experiments assess transfer effectiveness.

5. Residual error measured against each defense's own clean performance

The relevant within-defense quantity is

=100\left(\frac{\mathrm{MSE}_{\mathrm{adv,def},m}} {\mathrm{MSE}_{\mathrm{clean,def},m}}-1\right).$$ For the cited Weather iTransformer AT example, the defended clean/adversarial MSE pair is $0.1576\rightarrow0.2033$, giving **29.00% residual RMI**. The 38.90% entry instead uses the undefended attacked MSE 0.2189 and should not be read as this within-defense metric. Recomputing from the published four-decimal cells gives mean target-wise AT residual RMIs of **16.03%, 27.17%, and 42.63%** on ETTh1, Weather, and Exchange, versus **12.54%, 20.97%, and 33.45%** for the evaluated VMI variant (Appendix A.12). Extending the same calculation to all six defenses gives **positive within-defense RMI in all 90 dataset–defense–target combinations**. Section 7 below reports every per-target value. These are arithmetic recalculations from the published MSE cells, with the precision of those cells. For the specific FITS example on ETTh1, AT lowers attacked MSE to 0.4115, but its own clean MSE is 0.3539, leaving **16.28%** residual degradation. For Exchange TimeBase, the corresponding pair $0.1059\rightarrow0.1551$ leaves **46.46%**. Defense-induced error reduction and substantial residual attack effect are therefore simultaneously present, with each quantified against its appropriate reference.
AuthorsOpenReview ↗

Reply to Reviewer [3/3]

6. Configuration details, observed inputs, and “sparse”

The revised Appendices A.5.1–A.5.2 specify the main attack: L=720L=720, H=96H=96, batch size 16, ϵ=0.1\epsilon=0.1, 200 steps of size 0.01, w1=0.9w_1=0.9, Kkeep=12K_{\mathrm{keep}}=12, α=0.75\alpha=0.75, and a clean-input initialization. Appendix A.12 records three-step PGD with ϵ=0.1\epsilon=0.1 and step size 0.02 for ETTh1 AT on TimeBase and iTransformer, seed 2023, and checkpoint selection by clean validation MSE. Weather iTransformer MSE consistency uses weight 3.0 and the same inner PGD budget, with no random start. Test examples are held out from training. These training settings apply to the named runs. Table A6 identifies the evaluated VMI configuration by its reported parameter α=0.008\alpha=0.008. This is separate from the STDSA blending coefficient α=0.75\alpha=0.75.

Figure A2 provides the requested actual inputs and perturbation spectrum. Although the final perturbation is dense in time, approximately 70% of the final perturbation energy is placed in the lowest twelve full-lookback rFFT bins, which reflects the sense of sparsity. Moreover, the structural sparsity concerns the lightweight forecasting mechanisms and the restricted spectral processing. Appendix A.10 additionally evaluates genuine channel-restricted attacks with masks enforced throughout generation. SNR describes perturbation magnitude. The final spectrum and forecasting-error measurements provide the direct evidence about the constructed attack.

7. Per-target residual RMI for all six defenses

All entries below are percentages computed as 100(MSEadv,def/MSEclean,def−1)100(\mathrm{MSE}_{\mathrm{adv,def}}/\mathrm{MSE}_{\mathrm{clean,def}}-1) from the four-decimal MSE cells in revised Tables A1 (ETTh1), 1 (Weather), and A2 (Exchange). The final column is the arithmetic mean of the five target-wise ratios.

ETTh1

DefenseTimeBaseiTransformerSparseTSFFITSLSINetTarget-wise mean
TV15.69%15.39%16.85%17.70%19.29%16.98%
Randomized Smoothing14.52%14.61%15.63%17.34%17.92%16.00%
AT15.71%13.22%16.09%16.28%18.87%16.03%
MSE Consistency15.86%13.44%16.86%17.09%19.23%16.49%
KL Temporal Consistency15.99%12.95%16.71%17.94%19.43%16.60%
AT + MSE Consistency + TV15.41%13.38%15.95%16.12%18.69%15.91%

Weather

DefenseTimeBaseiTransformerSparseTSFFITSLSINetTarget-wise mean
TV24.34%27.38%31.84%32.62%34.06%30.05%
Randomized Smoothing24.10%27.31%31.80%32.17%34.34%29.94%
AT23.26%29.00%26.07%26.94%30.56%27.17%
MSE Consistency24.64%30.72%29.69%30.22%32.84%29.62%
KL Temporal Consistency24.53%29.66%31.84%32.68%33.13%30.37%
AT + MSE Consistency + TV23.26%27.37%25.90%26.64%29.51%26.54%

Exchange

DefenseTimeBaseiTransformerSparseTSFFITSLSINetTarget-wise mean
TV58.35%39.46%71.22%76.38%67.12%62.51%
Randomized Smoothing58.61%38.33%71.33%76.27%66.02%62.11%
AT46.46%26.16%48.45%51.53%40.57%42.63%
MSE Consistency46.85%33.70%50.05%53.33%47.17%46.22%
KL Temporal Consistency45.93%33.36%53.81%69.31%65.82%53.65%
AT + MSE Consistency + TV46.79%23.01%48.34%51.65%39.00%41.76%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.