TimeGuard: First Backdoor Defense Tailored for Time Series Forecasting

arXiv·low signal

TimeGuard addresses an underexplored threat: backdoor attacks on time series forecasting models used in finance, energy, and healthcare. Systematic evaluation of 13 existing backdoor defenses across the TSF lifecycle reveals two fundamental failure modes — data entanglement diluting channel-level signals and task-formulation shifts breaking classification-oriented defenses. TimeGuard introduces channel-wise pool training to isolate and neutralize backdoor triggers without degrading forecasting accuracy.