Reproduce “FedTADBench: Federated Time-Series Anomaly Detection Benchmark” on OpenResearch

About this paper

Fanxing Liu, Cheng Zeng, Le Zhang, Yingjie Zhou, Qing Mu, Yanru Zhang, Ling Zhang, Ce Zhu

This paper presents FedTADBench, the first comprehensive benchmark for federated time-series anomaly detection, which evaluates the effectiveness of combining various anomaly detection algorithms with federated learning methods across different data distribution scenarios. The benchmark demonstrates that federated learning is feasible and can sometimes outperform centralized training, while showing robustness to data heterogeneity and consistent improvement over isolated training.

Key insights

• Federated learning can sometimes outperform centralized learning for time-series anomaly detection, potentially due to a regularization effect.
• Federated time-series anomaly detection shows relative robustness to data heterogeneity across clients.
• Collaboration through federated learning consistently improves performance compared to isolated training for anomaly detection.

Summary by alphaXiv.