acceptodds
Under review as a conference paper at ICLR 2027

FUSE: Adaptive Frequency-Temporal Unified Semantic Attacks for Time Series Forecasting

Abstract

Time series forecasting (TSF) models are widely deployed in critical applications, yet their adversarial robustness remains insufficiently explored. Existing attacks mainly rely on time-domain perturbations or untargeted attacks. They overlook the structured vulnerability of forecasting models across spectral and temporal dimensions, and focus only on numerical prediction errors, without analyzing model sensitivity across different representation spaces. In this paper, we propose a Frequency-temporal Unified SEmantic attack framework for time series forecasting (FUSE). We first reveal that TSF models do not exhibit uniform sensitivity across the frequency spectrum, but instead exhibit varying vulnerability across multiple frequency bands. Based on this observation, we design an adaptive full-spectrum perturbation strategy to identify and exploit critical frequency components. However, frequency-domain perturbations alone are insufficient, as they induce global and stationary disturbances, while TSF models rely on localized and non-uniform temporal dependencies. To address this mismatch, we introduce a frequency-temporal alignment mechanism that couples spectral sensitivity with temporal importance. Finally, we extend adversarial targets from numerical deviation to semantic-level manipulation of trends and periodic patterns. Extensive experiments on four datasets demonstrate that FUSE consistently outperforms existing methods in attack effectiveness, transferability, and semantic controllability across diverse forecasting models and datasets. Our code is available at https://anonymous.4open.science/r/FUSE-3353/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Adversarial Attack, Frequency Domain

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6uvEOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes FUSE, an adversarial attack framework for time-series forecasting that combines adaptive frequency-domain perturbations with temporal importance to exploit non-uniform model sensitivity. Its main contributions are full-spectrum perturbation adaptation, frequency–temporal alignment, and targeted manipulation of forecast trends, seasonality, and spectral components. Experiments on four datasets and four forecasting architectures compare FUSE with eight attack baselines, reporting improvements in targeted manipulation and transferability, supported by component ablations.

Critical strengths
  1. The paper evaluates attacks that alter trends, seasonality, and spectral components, extending the assessment beyond aggregate prediction error. This makes the attack objectives more interpretable and provides a useful perspective on forecasting vulnerabilities.

  2. FUSE connects non-uniform frequency sensitivity with temporal importance, motivating an adaptive attack rather than a fixed frequency restriction. This is a substantive methodological merit, supported by component ablations.

Critical weaknesses
  1. The incremental benefit of the proposed components is modest and insufficiently established. In Table 3, FUSE improves ASR over BIM by only 1.63 percentage points in the white-box setting and 0.90 points in transfer; removing adaptive frequency weighting reduces ASR by just 0.71 and 0.40 points, respectively. Without uncertainty estimates, these differences provide limited support for the importance of the central spectral component. Moreover, frequency-weight optimization requires additional model evaluation and backpropagation, so matching iteration counts alone does not establish an advantage under comparable computational budgets.

  2. The semantic-attack evaluation relies on future information unavailable in an online attack. Section 2.7 constructs targets by decomposing the ground-truth future sequence into trend and seasonal components. This is valid as an offline benchmark, but the manuscript does not sufficiently distinguish this privileged-information setting from practical forecasting attacks, where the future is unknown. Evidence using targets constructed from historical observations or attacker-defined trajectories would materially strengthen the practical significance of the semantic-controllability claim.

Questions for authors

Section 2.7 constructs semantic targets from the ground-truth future sequence. Is this information assumed available to the attacker, and are there results using targets derived solely from historical observations or independently specified trajectories? This distinction materially affects the practical interpretation of semantic controllability.

The semantic attack appears to rely on privileged future information: Section 2.7 constructs attack targets from the ground-truth future sequence, which would ordinarily be unavailable at attack time. Although this is a valid oracle-based evaluation, it leaves the applicability to online forecasting unclear unless the authors explicitly justify this assumption or demonstrate effectiveness with targets constructed without future observations.

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

The paper proposed an iterative gradient based adversarial attack framework on TS forecasting models which combines frequency and time domain information. The FUSE introduced learnable frequency weights optimized by GD which could adjust the contribution of each frequency component in the Fourier domain to the perturbation. It also utilized frequency and temporal alignment and semantic target attack, such that the attack could manipulate specific semantic features.

Critical strengths

The proposed method from the paper tried to optimize the attacks through both frequency weighting and time perturbation which has a little bit novelty. It also considers the manipulation of multiple semantic features as attack objectives which helped to analyze how different types of structural perturbations make forecasting models vulnerable.

Critical weaknesses

The method's applicability in real-world can be limited due to lack of assessment of the perturbations' perceptibility. The paper should also analyze the number of gradient computations, the run time and attack effectiveness. Overall, there is no substantive innovation, and the proposed method is more like a combination of existing gradient-based attack techniques.

Questions for authors

Is the attack feasible in practice is a question. The gap between an algorithmic attack and a real security threat to apply the method is not clear and needs to be addressed.

Reviewer_P1QKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FUSE, an adversarial attack on time-series forecasting models that builds on sign-gradient iterative attacks (BIM) and adds three modules: Adaptive Frequency Weighting, Frequency-temporal Alignment, and Semantic Target Attack. Experiments cover ETTh1/ETTm1/ECL/Weather and DLinear/PatchTST/TimesNet/Informer, in white-box untargeted and targeted settings and black-box transfer. Anonymous code is provided.

Critical strengths
  • The motivation of the paper is reasonable and the application scenario is broad.
  • The paper compares different baselines in different settings, the extensive comparison adds the credibility to the paper.
Critical weaknesses
  • The improvements over a basic baseline are small, and the evaluation lacks the standard strong baseline and any variance estimates, so the claim of consistent superiority is not very established. Through Table 1 to Table 3, improvement of the proposed method FUSE over the standard baseline BIM is marginal.

  • There is no analysis or visualization of perturbed inputs.

  • No defenses are evaluated (low-pass filtering, smoothing, adversarial training), which is essential for a frequency-aware attack.

Questions for authors

The proposed "Semantic Target Attack" looks like a general objective that can be applied widely to baselines, but not a specific attack approach?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.