Strengthening Time Series: DTW-Certified Robust Anomaly Detection

The paper proposes the first DTW-certified robust defense for time series anomaly detection, realized by adapting the randomized smoothing paradigm. Existing defenses are mostly heuristic or certified only under ℓp norm constraints, and cannot characterize the intrinsic temporal structure of time series; the authors bridge ℓp norms and DTW distance via a lower-bound transformation to construct robustness certificates. In experiments across multiple datasets and multiple models, the method improves the F1 score by up to 18.7% over traditional certified models under DTW-based adversarial attacks.