New ARTA framework makes time-series anomaly detection systems resilient to adversarial attacks and corrupted data

ARTA uses joint adversarial training with a detector and mask generator to improve robustness against localized input corruptions and structured noise

The sparsity-constrained mask generator identifies minimal temporal perturbations that stress-test the anomaly detector during training

Adversarial masks serve dual purposes: hardening the detector against attacks while providing explainable insights into the detector's decision-making process

Addresses critical vulnerability of modern deep learning-based time-series anomaly detectors in complex system monitoring applications