acceptodds
Under review as a conference paper at ICLR 2027

TARGETWEAVER: ADAPTIVE TARGET CONSTRUCTION FOR BACKDOOR ATTACKS ON MULTIVARIATE TIME SERIES FORECASTING

Abstract

Backdoor attacks on multivariate time series forecasting commonly assume predefined future targets, overlooking that suitable attack responses may vary across datasets, samples, and variables. We propose TargetWeaver, which treats target shape and temporal placement as adaptive decisions rather than fixed attack specifications. TargetWeaver identifies dataset-relevant target patterns, selects context-dependent targets using temporal evidence and forecasting feedback, and adopts staged optimization to stabilize shape and location learning. A historical trigger generator then learns to induce the selected future responses. Across seven datasets and three forecasting architectures, TargetWeaver achieves the lowest average target-region error on six datasets while largely preserving clean forecasting accuracy. Our results suggest that adaptive target construction is a fundamental degree of freedom in forecasting backdoors, extending attack design beyond trigger optimization alone.

Primary area:
learning on time series and dynamical systems
Keywords:
Data Poisoning, Backdoor Attacks, Adaptive Target Construction

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_2gtnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TargetWeaver, a backdoor attack framework for multivariate time-series forecasting that adaptively selects target shapes and temporal positions for individual samples and variables. Its main contribution is a hierarchical target-construction mechanism combining dataset-specific pattern screening, context-dependent pattern and offset selection, and forecasting-model feedback, trained alongside a historical trigger generator through staged optimization.

Validation covers seven datasets and three forecasting architectures—FEDformer, Autoformer, and TimesNet—with comparisons against BackTime, TDBA, and three fixed-trigger strategies. The authors report the lowest average target-region MAE on six datasets while largely preserving clean forecasting accuracy, supplemented by component ablations, sensitivity analyses, and detectability evaluation. However, each method is evaluated against its own constructed targets, so lower target-matching error does not by itself establish stronger attack effectiveness rather than easier target selection. This limits the conclusions supported by the reported improvements.

Critical strengths
  1. A relevant research question. The paper investigates adaptive target construction rather than treating attack targets as fixed. Considering target shape and temporal placement jointly is a meaningful extension of forecasting-backdoor research and provides the main conceptual merit of the work.

  2. Reasonably broad empirical coverage. Experiments span seven datasets and three forecasting architectures, with component ablations and sensitivity analyses. This provides useful evidence about the framework’s behavior across settings, beyond a single favorable example.

Critical weaknesses
  1. The main comparison does not adequately control for target difficulty. Each method is evaluated against its own constructed targets, while TargetWeaver selects targets using forecasting feedback. Consequently, lower target-region MAE may reflect easier targets rather than a more effective backdoor. Without comparisons under matched targets or comparable target difficulty, the reported improvements do not convincingly establish superiority over existing attacks.

  2. The evaluation does not sufficiently establish the trigger’s causal effect. Target-matching error alone cannot determine whether the trigger produces a substantial malicious change or whether the model would already predict something close to the selected target. The Clean reference uses a fixed cone target rather than TargetWeaver’s selected targets, and illustrative prediction plots do not replace an aggregate comparison. Evaluating triggered and untriggered inputs against identical selected targets is essential to establish meaningful attacker-induced control.

  3. Practical stealth is not consistently demonstrated. USAD achieves an AUC-ROC of 1.0000 on Weather and 0.8243 on Alibaba, indicating substantial detectability in these settings. Preserving clean forecasting accuracy does not imply that poisoned inputs evade detection. Although the paper reports these limitations, the results leave the practical trade-off between target fitting and detectability insufficiently established.

Together, these concerns limit the strength of the central empirical claim.

Questions for authors

Do the authors have controlled results showing that TargetWeaver’s advantage persists when competing methods use identical targets, or targets of comparable difficulty? This would help determine whether the reported improvement reflects stronger backdoor effects or easier target selection.

or TargetWeaver’s selected targets, how does target-region MAE differ between triggered and unmodified inputs to the same poisoned model, evaluated on identical sample–variable pairs and intervals? This comparison would establish whether the trigger materially changes predictions toward the intended targets.

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

TargetWeaver studies adaptive target construction for training-data backdoors in multivariate time-series forecasting. It screens target templates, selects shapes and temporal offsets using context and surrogate feedback, and trains selectors alongside a historical trigger generator. The constructed data are then used to train separate forecasting models from scratch. Across seven datasets and three architectures, the paper reports the lowest cross-architecture average target-region MAE on six datasets while largely retaining clean forecasting accuracy.

Critical strengths
  1. Separating future-target construction from historical-trigger generation extends forecasting backdoors by introducing an additional design dimension relative to BackTime and TDBA (§1-4). This could help explain which responses forecasting backdoors can induce.

  2. Seven datasets and three downstream architectures provide useful coverage (§5.1). The paper explicitly distinguishes the stage-two FEDformer from the construction surrogate. It also reports the ETTh1 exception, mixed ablation outcomes, and perfect USAD detection on Weather (Tables 2-4), which help establish the method's practical limits.

Critical weaknesses
  1. The experiments do not sufficiently isolate the source of the claimed advantage. Equation (8) and Table 2 evaluate each method against its own constructed targets; the Clean column uses a fixed cone at offset zero on unmodified inputs. Lower target MAE can therefore reflect easier targets rather than stronger trigger-induced behavior. Figure 3 does not provide an aggregate paired control. The §5.3/Table 3 ablations replace adaptive patterns with a fixed cone and adaptive offsets with random offsets while removing several associated mechanisms. These comparisons do not isolate sample-specific adaptation from a favorable dataset-level choice, which makes the central contribution difficult to assess. Paired triggered/untriggered evaluations on identical frozen targets would help separate target attainability from trigger effects. A validation-selected global fixed shape-offset comparator, with otherwise matched settings, would help assess the benefit of sample-specific adaptation. Independent repeat counts and available variability are also needed to assess ranking stability; architecture averages are not repeated runs.

  2. The method and comparison settings need further specification. In Figure 2 and §4.3-4.4, OffsetSelector is conditioned on an initial one-hot pattern, but the final module receives pattern logits and depicts a pattern-offset search space. The available text does not say whether the final decision can override the pattern or whether it recomputes conditional offset scores. This leaves the roles of learned selection and feedback-based search unclear. Complete losses, gating, updates, selection constraints, and baseline configurations are deferred to Appendices A-C, which were absent from the supplied package. Section 5.1 also leaves exceptions to shared baseline settings unspecified. Without these details, I cannot adequately assess the implemented algorithm or the fairness of the comparisons. Please identify any separate supplements and provide an explicit decision rule, complete objectives and parameter-specific updates, normalization/target scaling, and baseline exceptions. Basic indexing, dimensions, and MAE denominators appear consistent under half-open slicing and entrywise-L1 conventions; the missing definitions do not demonstrate equation errors. The small labels in Figure 2, dense Table 2, and large gaps on p. 8 also make the paper harder to inspect. These are secondary readability issues and do not establish a formal formatting violation.

Questions for authors
  1. What evidence separates trigger effects and individualization from target attainability? For the existing targets, can you provide paired triggered/untriggered errors using identical targets, offsets, and sample-variable pairs, together with target deviation from true futures?
Reviewer_HJTGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses backdoor attacks on multivariate time series forecasting, where existing attacks (BackTime, TDBA) use predefined, fixed target shapes for the future response. The authors argue that suitable targets vary across datasets, samples, and variables, and propose TargetWeaver, which adaptively selects both the target shape (pattern) and temporal placement (offset) per sample–variable pair.

Critical strengths
  1. Framing target shape and temporal placement as learnable, sample- and variable-dependent decisions (rather than hyperparameters) is a genuinely new angle over BackTime/TDBA, and the offline diagnostic in Figure 1 provides direct evidence that fixed targets are suboptimal. The hierarchical selector, forecasting-feedback, and staged-training design is non-trivial and internally consistent.

  2. Seven datasets and three architectures with matched settings against BackTime/TDBA/fixed-trigger baselines, component ablations, parameter sensitivity, and a stealth analysis that reports failures (Weather/Alibaba) rather than cherry-picking. The ETTh1 exception and the Weather pattern-adaptation regression are disclosed rather than hidden.

Critical weaknesses
  1. The most important figure is likely generated by GenAI.

  2. The headline metric does not establish attack effectiveness, it may only show easier targets.

  3. The proposed attack is not reliably stealthy, and key implementation details are missing.

Questions for authors
  1. Since MAE_a is computed against each method's own constructed target and your selectors minimize target error, how do we know the gains reflect stronger trigger-induced control rather than selection of easier targets?

  2. On Weather, clean MAE_c improves by ~3.2 pp under poisoning (Average 14.58 → 11.34; TimesNet 23.40 → 15.33). Why? Is this due to surrogate-selected high-error timestamps being removed from training, or run variance? Please report results across multiple seeds (e.g., 3–5) with mean ± std and significance for the main table and ablations.

  3. On Weather/Alibaba the trigger is clearly detectable (AUC 1.0000 / 0.8243). How do MAEa_a and MAEc_c change if per-variable modification is constrained (e.g., ≤1×SD, per-variable clipping)? Is the Weather gain of 8.59 pp over BackTime preserved under a stealth-constrained trigger budget?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.