acceptodds
Under review as a conference paper at ICLR 2027

BoundaryWarp: Revealing and Exploiting Patch-End Vulnerabilities in Time-Series Foundation Models

Abstract

State-of-the-art time-series foundation models divide input sequences into fixed, non-overlapping patches before forecasting, yet the robustness consequences of this design are not well understood. We identify a recurring tokenizer-relative patch-end vulnerability in which model sensitivity concentrates at the final position of a patch. To distinguish this pattern from general recency effects, we shift tokenizer alignment and patch length and find that the sensitivity maximum moves with the active patch grid rather than remaining at a fixed temporal position. Internal-boundary, unpatched, recency-controlled, and content-matched analyses further support a tokenizer-relative explanation. We then introduce BoundaryWarp, an untargeted temporal deformation attack that exploits this structural sensitivity through bounded resampling while preserving temporal order, endpoints, and context length. Across Chronos-Bolt, Moirai, and TimesFM on four forecasting benchmarks, BoundaryWarp consistently degrades forecasts under sub-sample temporal constraints. With a maximum displacement of half a sampling interval, it increases median mean squared forecasting error by up to 12.76% relative to clean predictions. Our results show that fixed patch geometry is not only a computational design choice but also a robustness-relevant architectural property that can create a systematic adversarial attack surface in time-series foundation models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Time-Series Foundation Models, Adversarial Robustness, Patch Tokenization, Temporal Deformation Attacks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uAoQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether fixed patching makes time-series foundation models more sensitive to input perturbations at patch ends. It first examines loss gradients on clean inputs, then changes tokenizer alignment and patch length to test whether sensitivity peaks move with patch positions. BoundaryWarp uses this observation to introduce a boundary prior into monotone, displacement-bounded temporal resampling, increasing forecasting error while preserving temporal order and endpoints. Experiments on Chronos-Bolt, Moirai, and TimesFM across four datasets compare random warps, boundary-unaware optimized warps, and BoundaryWarp. The results support a relationship between sensitivity and tokenizer position and show that optimized temporal warps can increase error, although the extra gain from the boundary prior varies across models.

Critical strengths
  1. The alignment experiment tests a falsifiable prediction. The descriptive pooled alignment profiles track the predicted tokenizer-relative phase, and the patch-length intervention provides complementary evidence. Series-balanced exact peaks agree in 43 of 48 configurations, while the stricter joint direct-peak criterion supports 26 of 48. The internal-boundary, recency-detrended, and content-matched comparisons help distinguish the observed pattern from the recency and local-content explanations tested in the paper. This finding is the main reason I support acceptance.

  2. The matched controls separate optimization from the boundary prior. Optimized Warp shares the objective, parameterization, optimizer, and budget with BoundaryWarp and differs only in the prior. This separates the benefit of optimizing a warp from the extra benefit of knowing the patch grid, and it lets the paper report that the prior gives a recurring rather than universal gain.

Critical weaknesses
  1. The link from value sensitivity to time-warp vulnerability is assumed. Equation (2) measures sensitivity to observed values, while the attack in Section 3.3 changes sampling times. A sub-sample warp changes the value at a position roughly by the local slope times the displacement, so vulnerability depends on the signal’s local variation as well as on s_t. The paper does not show that high s_t predicts where optimized warps place displacement or how much error they add. The “exploiting” part of the contribution therefore rests on the prior’s added effect.

  2. That added effect is uneven. In Table 1, the gain of BoundaryWarp over Optimized Warp is 2.31, 1.25, and 1.65 points, and the TimesFM interval includes zero; 20 of 36 settings remain significant after Holm correction. Under output-only access in Section 4.3, boundary-weighted SPSA does not beat uniform SPSA in any setting. The evidence therefore supports tokenizer-relative sensitivity more strongly than an attack that needs the true grid, which limits my enthusiasm for the method.

Questions for authors
  1. Without the prior, where does Optimized Warp place its displacement? If it already concentrates near patch boundaries, the prior mainly speeds up a ten-step search; if not, the sensitivity analysis and the attack may be exploiting different things.
  2. Does the gain over Optimized Warp in Table 1 come with larger realized displacement or roughness, given that the prior reweights intervals before the L∞ rescaling?
  3. Correct and shifted boundary priors (e.g., shifted by half a patch) compared at the same optimization budget and similar realized displacement. If the shifted prior performs as well, the gain does not depend on knowing the true boundary locations.
  4. Per-window rank correlation between attack gain and s_t at the displaced positions, alone and weighted by the local slope.
Reviewer_85PaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the robustness consequences of fixed, non-overlapping patch tokenization in time-series foundation models. It reports a tokenizer-relative "patch-end" vulnerability: gradient-based input sensitivity on clean, undeformed inputs concentrates at the terminal position of each patch (φ = P−1). This is distinguished from recency/absolute-position/content effects by controlled interventions: shifting the tokenizer alignment over the same sequence moves the sensitivity peak with the grid (predicted phase φ(o) = (P−1−o) mod P matched in 28/28 Chronos-Bolt settings), and changing patch length relocates the peak to the new P−1 (16/16 Moirai settings).

Critical strengths
  1. The demonstration that the sensitivity maximum tracks the active tokenizer grid, moving with alignment shifts and relocating with patch length, together with the unpatched, recency-detrended, internal-boundary, and content-matched controls, convincingly rules out the main confounds. This is a substantive mechanistic finding about patch-based forecasting architectures, and the quantitative prediction (φ(o)) matched exactly in all 28 alignment settings is strong evidence.

  2. Series-cluster bootstrap inference, explicit multiplicity control, deformation-validity audits, matched Random/Optimized/BoundaryWarp controls, restricted-access SPSA with query budgets, NRMSE robustness checks, and full per-setting tables. The paper reports where patch information does not help (uniform-vs-boundary SPSA) and separates the existence of the vulnerability from the efficiency of any particular exploit, a commendable level of transparency.

Critical weaknesses
  1. Only 12/28 Chronos-Bolt alignment settings pass the joint strict criteria, and the alignment-shift and patch-length interventions each rest on one model (Chronos-Bolt and Moirai, respectively). Because the strongest causal claims (grid tracking) are demonstrated on a single model per intervention, the generality of "sensitivity is organized by tokenizer geometry" across architectures is less certain than the pooled descriptive profiles suggest.

  2. BoundaryWarp is an untargeted, oracle-based attack (retrospective scoring against the realized future) whose degradation is modest at the primary budget, and more importantly the patch-specific component does not reliably beat a boundary-unaware optimizer: 20/36 significant white-box comparisons, and 0/36 under restricted access with pooled differences of only 0.25–0.34 pp. This means the paper establishes a vulnerability (patch-end sensitivity) much more strongly than it establishes that the vulnerability is practically exploitable through the proposed mechanism; an optimized warp without any patch knowledge captures most of the attack. As written, the third contribution ("exploitation") overstates the role of the identified structure.

Questions for authors
  1. In the series-balanced direct-peak analysis, only 12/28 Chronos-Bolt alignment settings, 13/16 Moirai settings, and 1/4 TimesFM native settings pass the joint (argmax + CI + Holm) criteria, versus 28/28 and 16/16 descriptive. Does the terminal-maximum claim hold per dataset for every offset and patch length under the strict criteria? What drives the gap—power, or genuinely non-terminal peaks in specific settings?

  2. The attack requires the realized future continuation y for retrospective scoring and is untargeted. How would BoundaryWarp behave with a self-supervised surrogate objective (e.g., predicting the observed window from a warped prefix, or a learned critic) in place of the oracle? What degradation level is practically consequential for downstream applications such as energy or traffic management?

1 comment
Reviewer_85PaOpenReview ↗

Thanks for your response.

Thanks for the responses. I maintain my score. However, the quality of the text and fighures need to be improved. For example, Figure 1 need to be polished; Table 1 is hard to understand.

Reviewer_phjzOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how fixed-patch tokenization affects the robustness of foundation models for time series. We observe that the sensitivity of the prediction loss often peaks at the final position of each patch. By adjusting patch alignment and varying patch lengths—while employing several control conditions to disentangle this effect from the influences of "recency" and local signal content—we validate this observation.

Building on this finding, we propose "BoundaryWarp," an attack method that modifies the sampling time points of an input sequence while preserving its temporal order and limiting the magnitude of the displacement. Experiments conducted on Chronos-Bolt, Moirai, and TimesFM models across four datasets demonstrate that BoundaryWarp effectively increases prediction errors. Its performance generally surpasses that of random warping and other comparable optimization-based attacks that do not leverage boundary information.

Critical strengths
  1. This paper not only observes that vulnerability varies with position, but further reveals that peaks in sensitivity often coincide with the "patch boundaries" of the tokenizer.

  2. The authors compare BoundaryWarp against "random warping" and an optimizer that lacks boundary information (while being otherwise identical), which helps to disentangle the respective contributions of resampling, the optimization process, and a "patch-aware prior."

Critical weaknesses
  1. Shifting 496 observations within 512 positions changes the number of partially observed patches and the masking pattern. Since this experiment is central to the paper's main claim, the authors should clarify whether these changes affect the measured sensitivity peaks.
  2. Optimized Warp, which does not use boundary information, already achieves most of the attack effect. BoundaryWarp shows statistically significant improvements in only 20 of 36 comparisons after correction, with no significant overall improvement for TimesFM or under output-only access.
Questions for authors
  1. In the tokenizer-alignment experiment, how are partially masked patches handled? What evidence rules out changes in masking patterns and the number of active patches as an explanation for the sensitivity peaks?

  2. How can the authors confirm that BoundaryWarp's additional gains come from patch-boundary information rather than differences in optimization? Would a prior shifted away from the actual boundaries, or a longer optimization run for Optimized Warp, produce similar results?

1 comment

Could you please respond to the questions in my original review? Also, the paper is missing a diagram showing how the full method works. Figure 1 only illustrates time warping. The main results table is too brief; please include the key per-dataset comparisons in the main text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.