FedSW-TSAD: SWGAN-Based Federated Time Series Anomaly Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Decentralized time series data collection presents challenges for model training and data privacy.
• Existing federated anomaly detection methods suffer from unstable training and poor generalization due to client heterogeneity.
• Single-path detection methods lack the expressiveness to handle diverse anomalies effectively.

Purpose of the Study:

• To propose FedSW-TSAD, a novel federated time series anomaly detection method.
• To enhance training stability and generalization in federated anomaly detection.
• To ensure robust anomaly detection and privacy preservation in decentralized sensor networks.

Main Methods:

• Utilized Sobolev-Wasserstein GAN (SWGAN) to stabilize adversarial training.
• Combined discriminative signals from reconstruction and prediction modules for improved robustness.
• Implemented a differential privacy mechanism with L2-norm-constrained noise injection for privacy preservation.

Main Results:

• FedSW-TSAD demonstrated superior performance, achieving an average F1-score improvement of 14.37% over existing methods.
• The method showed enhanced robustness against diverse anomalies in real-world sensor datasets.
• Gradient privacy was significantly improved under the differential privacy mechanism.

Conclusions:

• FedSW-TSAD offers a practical and effective solution for privacy-preserving anomaly detection in federated settings.
• The proposed method addresses key challenges in decentralized time series analysis.
• FedSW-TSAD has significant implications for industrial IoT, remote diagnostics, and predictive maintenance.