Backdoor Attacks on Time-Series Forecasting
Stealthy backdoor attacks on multivariate forecasting under realistic missing-value scenarios.
Undergraduate thesis, DBIS Lab, Nankai University — with Dr. Xiangrui Cai · Dec 2024 – May 2025
- A backdoor-attack framework for multivariate time-series forecasting under realistic missing-value scenarios, with concealed trigger injection across data series.
- An imputation strategy based on critical time-step analysis restores data integrity while preserving trigger stealthiness.
- Invisible triggers placed at missing-value positions, strengthening the association between trigger patterns and target outputs.
- Achieved a 43% reduction in attack MAE compared with baseline methods.