TimeGuard: Channel-wise Pool Training for Backdoor Defense in Time Series Forecasting
TimeGuard: Channel-wise Pool Training for Backdoor Defense in Time Series Forecasting
Authors: Quang Duc Nguyen, Siyuan Liang, Yiming Li, Fushuo Huo, Dacheng Tao Date: 2026-05-21 Paper ID: openalex:2605.22365
Summary
This paper presents a systematic evaluation of backdoor defenses in time series forecasting, revealing that existing methods struggle due to channel-level signal dilution and training-loss degeneration. To address these issues, the authors introduce TimeGuard, a training-time defense mechanism that employs a channel-wise pool training strategy. By initializing a high-confidence data pool and refining it through distance-regularized loss selection, TimeGuard effectively sanitizes the training process without sacrificing accuracy on clean samples. Experimental results confirm significant robustness gains across diverse forecasting architectures and attack scenarios.
Key Contributions
- Identified two fundamental failure modes of existing backdoor defenses in TSF: channel-level signal dilution and training-loss degeneration.
- Proposed TimeGuard, a training-time defense that utilizes channel-wise pool training and distance-regularized loss selection to improve robustness.
- Demonstrated that TimeGuard improves MAE_P by 1.96x over leading baselines while maintaining high performance on clean data.
Key Concepts
- channel-wise-pool-training: A training-time backdoor defense paradigm for time series that isolates channel-specific signal patterns to prevent the dilution caused by contaminated windows.
Archivist Review
I approved ‘Channel-wise Pool Training’ because it captures a distinct, reusable paradigm for addressing backdoor threats in multivariate time series where channel dependency often complicates sample-level filtering. The overarching framework ‘TimeGuard’ was rejected as it is a specific implementation of these principles, and I followed the policy to favor more fundamental, reusable sub-mechanisms.
Approved Concepts
- Channel-wise Pool Training: It addresses the specific challenge of channel-level signal dilution in time series backdoor defense by isolating signals within channels to prevent poisoning contamination.
Rejected Candidates
- [concept] TimeGuard (
timeguard) - subcomponent_of_broader_mechanism: This is the name of the overall framework; the individual core mechanisms (channel-wise pool training) are more reusable as concepts.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.22365
- paper_id
- 2605.22365
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- channel-wise-pool-training
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-24T06:18:23Z