acceptodds
Under review as a conference paper at ICLR 2027

AdvMissing-TSC: Adversarial Missingness and Mask-Aware Defense for Time Series Classification

Abstract

Time-series classifiers can perform well under random missingness yet misclassify samples when class-discriminative observations are missing. We introduce AdvMissing-TSC to evaluate this vulnerability by selectively masking a fixed fraction of observations in each sample while leaving the remaining values unchanged. Its target-guided sparse missingness attack (TG-SMA) combines Gumbel relaxation with top-k projection to optimize masks using classifier feedback, either to reduce classification accuracy or to approach a specified target accuracy. Temporal regularization encourages contiguous missing blocks. Using masks optimized under mean or zero imputation, we evaluate the missingness–imputation–classification pipeline, examining how missingness structure affects classification accuracy and comparing imputation methods under fixed masks. We further propose MID, a mask-aware consistency defense that uses learned gating to fuse numerical and missingness features. MID penalizes prediction discrepancies between complete inputs and their masked-and-imputed counterparts more strongly at higher missing rates. Experiments across multiple datasets show that adversarial missingness yields lower accuracy than the corresponding random-missingness conditions for all evaluated classifiers. Among the four evaluated combinations of missingness patterns and imputation methods, MID raises accuracy in the most severely degraded setting to a level comparable to or slightly higher than that of the least degraded setting without defense.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series classification, Adversarial missingness, Adversarial robustness, Missing data imputation, Black-box attacks, Consistency regularization

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_kP9MOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the impact of selectively deleting time series observations on classification under a fixed missing data budget. TG-SMA uses Gumbel relaxation, cross-entropy feedback, and per-sample Top-k projection to search for missing locations, while employing structural regularization to encourage contiguous missing values; NES-Mask provides query-based search for non-differentiable classifiers. The authors generate masks on top of Mean or Zero proxy imputers, then compare different imputation methods while keeping the masks fixed, and propose MID, which fuses numerical and missing indicator features and constrains consistency between complete and missing view predictions. The experimental setup covers ten datasets, seven imputation methods, and four classifiers; in the main experiment across six datasets, adversarial missing values caused greater loss than corresponding random missing values, and the main defense experiment using Chinatown–Informer demonstrated that MID can improve accuracy under fixed attack masks.

Critical strengths

1.This paper shifts the focus from the amount of missing data to the impact of missing locations on classification performance, demonstrating the significant effects of selective information loss through adversarial attacks under a strictly controlled missingness budget, which offers a novel perspective. 2.The fixed-mask experimental design effectively eliminates the influence of varying missing locations when comparing imputation methods, while the objective discussion of methodological limitations enhances the credibility of the experimental findings.

Critical weaknesses
  1. The paper does not sufficiently validate the specific contributions of the MID mechanism or its effectiveness against adversarial attacks.
  2. The attack evaluation lacks essential experimental details and protocols needed to ensure fair comparisons and reproducibility.
Questions for authors

1.Please clarify how the missingness masks are generated during MID training and testing, and further evaluate whether MID remains effective under adaptive attacks targeting the complete defended system. 2.I suggest conducting additional ablation studies on MID, including removing the gating mechanism, consistency loss, and missing-rate weighting, to assess the individual contributions of these components. 3.To ensure fair comparisons, I recommend controlling the number, length, and shape of missing blocks under the same missingness budget for both random and adversarial missingness. 4.Please provide the complete attack configurations and query budgets for TG-SMA and NES-Mask, and clarify whether missingness masks are shared across classifiers and whether padding positions in variable-length sequences are included in the attack. 5.Please provide further details on data splitting, preprocessing, and model training, explain the discrepancy in Heartbeat–TimesNet results between Table 1 and Figure 6, and report the mean and standard deviation over multiple independent runs.

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

This paper studies adversarial missingness in time series classification: an attacker hides a fixed number of observations per sample while preserving the remaining values, and the incomplete series is imputed and classified. It proposes TG-SMA, a gradient-based mask search for reducing accuracy or approaching a target accuracy, and NES-Mask, a query-based search. MID combines value and missingness features through a learned gate and trains for prediction consistency between complete and imputed views. Experiments on UCR/UEA benchmarks evaluate the attacks across missingness patterns, imputers, and classifiers, and assess MID's effectiveness.

Critical strengths
  1. The exact-budget observation-removal threat model addresses a relevant vulnerability of time series classifiers and complements attacks that alter observed values. It makes the effect of selectively losing informative observations directly testable.
  2. Comparing imputers under fixed masks and downstream classification accuracy is useful because it controls which observations remain available. The gradient-based and query-based searches also extend the evaluation to different classifier access settings.
Critical weaknesses
  1. Robustness evidence needs validation. With matching experiments and aggregation under five-point trapezoidal integration, Table 7's Chinatown/Rand-B(avg)/Linear AURC of 0.99 exceeds the 0.9714 bound implied by Acc@30% = 0.90. Table 2's 92% versus 86% at k = 1, assuming standard Chinatown inputs, also leaves structure effects confounded with search behavior.
  2. Appendix F.5 extends MID's evaluation, but adaptive attacks, component ablations, and a baseline with the same backbone and training views are missing. Its gains cannot yet be separated from the benefit of training on missing data. The acknowledged mixed results also leave its effective conditions unclear.
Questions for authors
  1. How are AURC and Rand-B(avg) computed in Tables 7 and 9? Please provide the per-budget accuracies and aggregation details needed to reconcile the Table 7 example above.
  2. Which experiment-specific settings explain 0.758 in Table 1 versus approximately 0.87 in Fig. 7 for Chinatown–Informer–Adv-B at s = 0.30? Please report seeds, uncertainty, and whether masks are shared across classifiers.
  3. Does Table 2 use k = 1 at s = 0.05 on Chinatown? If so, how do both attacks compare with exhaustive evaluation of all 24 single deletions using the same classifier, imputer, and accuracy-minimizing objective?
  4. How does MID perform under adaptive attacks and against a baseline with matched backbone, training views, and training budget? What do gate and consistency ablations show?
Reviewer_NMz6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the robustness of time-series classification models under adversarial missingness, aiming to identify which values lead to significant performance degradation when those are missing. To formulate this setting, the authors introduce an adversarial missingness threat model for time-series classification. Specifically, given a frozen classifier, an imputer, and a fixed per-sample missingness budget (fixed missing rate), the objective is to find binary masks that minimize classification accuracy or approach a specified target accuracy. And then, they propose two mask search algorithms depending on the availability of classifier gradients : (1) TG-SMA, which is a gradient-based search algorithm to optimize learnable missingness masks using Gumbel relaxation and top-k projection and (2) NES-Mask, which is a query-based search method for settings where classifier gradients are unavailable. To evaluate the impact of missingness masks, the authors compare random and adversarial missingness patterns and examine how different imputation methods affect downstream classification performance under fixed masks. Finally, they propose MID which combines features of values and missingness masks through a gating mechanism and encourages output consistency between complete and imputed inputs. Experimental results show that adversarial missingness generally leads to greater accuracy degradation than random missingness while MID improves classification performance in the primary defense experiment on Chinatown–Informer.

Critical strengths

Well-motivated problem and reasonable formulation : The paper starts from highlighting an important limitation of evaluating time-series classification robustness under random missingness alone since the selective loss of discriminative observations can lead to substantial performance degradation under the same missing rate. As evaluating robustness under potential worst-case missingness scenarios is important in real-world settings, I think this provides a meaningful motivation for this work. To evaluate time-series classification models under such scenarios, identifying the discriminative observations is therefore required and I find it natural and reasonable that the authors formulate this identification as a constrained optimization problem over binary observation masks and propose gradient-based and query-based search algorithms to handle the combinatorially large search space of possible masks. Also, the experimental results provide empirical support for the importance of evaluating robustness under this adversarial missingness.

Critical weaknesses
  1. Limited novelty clarification : I think the paper's novelty is somewhat limited. Several prior studies have already investigated the effects of selectively masking or perturbing informative observations in the time-series domain [1,2]. Also, other prior works have explored the idea of optimizing learnable masks to identify important observations [3] and using mask-based removal to simulate the absence of observations [4]. While the proposed optimization objective and evaluation setting differ from these prior works, I think the paper does not sufficiently discuss and clarify its novelty against these closely related works. Additionally, MID also seems to be incremental rather than novel. In my opinion, incorporating missingness masks into model architectures is already common in the irregular time-series domain [5] and prediction consistency regularization is also well-established [6]. Thus, the proposed MID does not seem to offer enough methodological novelty. I think therefore the authors should more clearly position their contributions relative to existing work and provide a more explicit explanation of what is methodologically novel in both the proposed adversarial attack framework and MID.

  2. Insufficient experimental validation : While the proposed adversarial attacks show greater performance degradation than random missingness, the paper lacks comparisons with non-random mask baselines. Comparing only against random missingness makes it difficult to assess how effectively the proposed algorithm identify challenging missingness masks. In my opinion, comparisons with simpler approaches, such as gradient-based saliency or greedy mask selection, would help establish whether the proposed optimization methods offer meaningful advantages over existing alternatives. Also, the evaluation of MID lacks sufficient ablation studies to isolate the contributions of missingness-aware gating and prediction consistency regularization.


 [1] Ismail et al., Benchmarking Deep Learning Interpretability in Time Series Predictions, NeurIPS 2020.
 [2] Turbé et al., Evaluation of post-hoc interpretability methods in time-series classification, Nature Machine Intelligence, 2023.
 [3] Crabbé and van der Schaar, Explaining Time Series Predictions with Dynamic Masks, ICML 2021.
 [4] Moebus et al., Contimask: Explaining Irregular Time Series via Perturbations in Continuous Time, NeurIPS 2025.
 [5] Che et al., Recurrent Neural Networks for Multivariate Time Series with Missing Values, Scientific Reports, 2018.
 [6] Ismail et al., Improving Deep Learning Interpretability by Saliency Guided Training, NeurIPS 2021.
Questions for authors
  1. Novelty clarification : As mentioned above, could the authors clarify the key methodological differences between the proposed adversarial missingness framework and existing other mask-based approaches? Beyond the optimization objective and evaluation setting, what are the main methodological contributions of the proposed approach? Also, the authors mentioned MID is one of the paper's main contributions, so could the authors clarify the novelty of MID beyond simply combining existing missingness-aware gating and prediction consistency regularization techniques?

  2. Additional empirical evidence : Could the authors provide comparisons with other non-random mask baselines such as gradient-based saliency or greedy mask selection? If possible, I think comparisons with other optimization-based methods would also be helpful. Also, could the authors include ablation studies to isolate the contributions of missingness-aware gating and prediction consistency regularization in MID? I think it could be helpful to show whether combining these two components provides meaningful improvements to better justify the design choice.

  3. Evaluation under adaptive attacks : Could the authors clarify how the adversarial missingness masks used in the MID experiments were generated? Were the same masks used to the naive classifier and MID-trained classifier? As I understand, the reported MID results do not evaluate masks optimized directly against the MID-trained model. Although the authors acknowledge this limitation, I think evaluating MID against masks optimized directly for the defended MID-trained model is very crucial to properly assess its effectiveness. Could the authors evaluate whether MID remains effective when the missingness masks are optimized directly against MID itself?

Reviewer_3QPSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The manuscript presents AdvMissing-TSC, a framework for evaluating and mitigating the vulnerability of time series classifiers to adversarial missingness—where an attacker selectively hides a fixed proportion of observations [0, 1] while leaving retained values unaltered. The paper addresses two core problems:

  1. Adversarial Missingness Threat Model & Attack Generation: TG-SMA (Target-Guided Sparse Missingness Attack): A white-box gradient-based attack using continuous relaxation coupled with hard Top-k budget projection. It optimizes binary masks to either maximize accuracy degradation or hit a pre-specified target accuracy level a*. Temporal, channel, and synchronization total variation (TV) regularizers are included to induce contiguous missing blocks. NES-Mask: A black-box, query-based zeroth-order search using Natural Evolution Strategies (NES) to optimize mask logits when model gradients are unavailable.

  2. Fixed-Mask Evaluation & Defense:Fixed-Mask Protocol: Evaluates downstream classification conditional on fixed generated masks across seven imputation algorithms (Zero, Mean, Linear, CDRec, BRITS, GAIN, TKCM) to decouple mask generation from imputer selection. MID (Mask-aware single-Imputation consistency Defense): A defense mechanism combining value features H_x and mask features H_m via a reliability gate. MID is trained on clean complete views and imputed incomplete views using cross-entropy and a missingness-weighted Jensen-Shannon (JS) divergence consistency loss. During inference, it requires only a single imputation and one forward pass.

The authors evaluate AdvMissing-TSC across ten UCR/UEA datasets, four classification backbones (TimesNet, Informer, MiniROCKET, 1NN-DTW), and seven imputers.

Critical strengths
  1. Formulation of Observation Availability as an Adversarial Surface: Shifting the adversarial paradigm in time series from additive value perturbations to pure observation availability masks is realistic for healthcare, IoT, and industrial monitoring where sensor failures or packet drops can be manipulated or occur selectively.
  2. Differentiable Top-k Projection Mechanics: The combination of continuous Gumbel-Softmax relaxation with non-differentiable Top-k index projection provides an exact budget enforcement mechanism per sample without relying on soft penalty heuristics during final evaluation.
  3. Rigorous Fixed-Mask Evaluation Protocol: Testing multiple imputers under identical, pre-computed missingness masks prevents the imputer from confounding the attack search process. This exposes a key insight: reconstruction quality (e.g., RMSE) does not correlate directly with downstream classification performance under selective information loss.
  4. Theoretical Foundations for Consistency and Gating: Proposition 2 correctly connects discrete TV penalties to the upper bound on missing segment counts. Lemma 1 proves local perturbation attenuation through feature gating. Theorem 1 provides a tight bound linking JS divergence thresholds directly to class decision stability.
Critical weaknesses
  1. Failure to Evaluate Against Adaptive Attacks (The "Obfuscated Gradients" Fallacy) The Argument: The paper evaluates MID by testing it against masks generated against undefended, clean-trained baseline models using proxy Mean/Zero imputers. Why it is Weak: In security and adversarial ML standards (Athalye et al., 2018; Tramèr et al., 2020), a defense cannot be claimed robust if it is only tested against non-adaptive, transfer attacks. Because MID modifies the network architecture by introducing gated feature pathways and consistency training, an end-to-end white-box adversary targeting MID directly could easily bypass the defense by backpropagating through the gated fusion module.

  2. Surrogate Imputer Mismatch & Conditional Attack Fragility The Argument: TG-SMA uses a simple, differentiable proxy imputer during mask optimization, and subsequently applies these fixed masks to evaluate complex imputers like CDRec, BRITS, or GAIN. Why it is Weak: An adversarial mask optimized to fool a classifier under Mean imputation targets features that Mean imputation fails to restore. Advanced imputers (e.g., matrix completion or generative models) reconstruct temporal patterns far more effectively. Consequently, an attack generated under Mean imputation is a sub-optimal surrogate for CDRec or GAIN, leading to an overestimation of robustness for complex imputers.

  3. Selective Superiority and Non-Universal Performance of MID The Argument: Section 5.6 highlights significant MID accuracy recovery on Chinatown with Informer (increasing accuracy from 65.0% to 76.0% at s=0.40. Why it is Weak: Chinatown is an extremely short univariate time series T=24. As shown in Appendix F.5 (Figure 8), on other datasets such as Heartbeat and Handwriting, MID offers marginal, inconsistent, or even negative performance trade-offs compared to undefended models. The manuscript explicitly concedes that "MID gains are not universal", yet bases its primary defense claims on Chinatown.

  4. Arbitrary Regularization Coefficients Without Sensitivity Analysis The Argument: Structural regularizers are fixed across all datasets. Why it is Weak: Different time series domains (e.g., high-frequency ECG vs. low-frequency traffic data) have vastly different temporal resolutions and channel dependencies. Applying static weights across all datasets without hyperparameter sensitivity or ablation studies makes it unclear how much each component contributes to generating realistic block missingness.

Questions for authors

Listed in comments above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.