TimeGuard: New Defense Against Backdoor Attacks in Time Series Forecasting
Researchers have identified vulnerabilities in Time Series Forecasting (TSF) to backdoor attacks and proposed a new defense mechanism called TimeGuard. The study systematically evaluated thirteen existing backdoor defenses across the TSF life cycle, revealing two fundamental issues: data entanglement causing channel-level signal dilution, and task-formulation shift leading to training-loss degeneration. These issues render sample-filtering and trigger-synthesis defenses ineffective. TimeGuard introduces channel-wise pool training as its core paradigm, initializing a high-confidence pool to mitigate backdoor threats during training. The research is published on arXiv with ID 2605.22365.
Key facts
- Time Series Forecasting is vulnerable to backdoor attacks.
- Thirteen representative backdoor defenses were evaluated across the TSF life cycle.
- Data entanglement causes channel-level signal dilution.
- Task-formulation shift leads to training-loss degeneration.
- Sample-filtering and trigger-synthesis defenses are ineffective.
- TimeGuard uses channel-wise pool training.
- TimeGuard initializes a high-confidence pool.
- The paper is available on arXiv with ID 2605.22365.
Entities
Institutions
- arXiv