TrojanScope: Interpretable Backdoor Detection for Time Series Forecasting

Maolin Wang, Pengzhi Zhong, Binhao Wang, Tianshuo Wei, Beining BAO, Ruocheng Guo, Xiangyu Zhao

Keywords: backdoor detection, time series forecasting, visual analytics, interpretable machine learning, adversarial robustness, temporal pattern analysis, neural network security

Abstract: Time series forecasting models are increasingly targeted by backdoor attacks, which embed hidden triggers into parameters while preserving accuracy on clean data. Existing defenses from vision and NLP fail to capture the temporal complexity of triggers and offer limited interpretability. We present TrojanScope, a visual analytics framework for reliable and explainable backdoor detection in time series models. TrojanScope integrates (i) block-level substitution to localize contaminated components, (ii) residual filtering with correlation analysis to isolate abnormal temporal channels, and (iii) candidate extraction and optimization to reconstruct the hidden trigger. Our method is supported by theoretical guarantees on detection consistency and recovery accuracy, and it produces intuitive visual evidence of Trojan propagation. Across multiple datasets (including ESA-ADB satellite telemetry) and architectures (PatchTST, N-HiTS), TrojanScope achieves superior accuracy, efficiency, and interpretability compared with state-of-the-art baselines. This work highlights the importance of combining formal guarantees with visual diagnostics for trustworthy and practical backdoor defense in high-stakes time series forecasting.

Primary Area: learning on time series and dynamical systems

Submission Number: 11474