A Survey of Research Advances in Federated Learning for Anomaly Detection

Abstract

In fields such as industrial control systems, cybersecurity, and finance, detecting abnormal behaviors is of utmost importance to ensure operational reliability and prevent potential losses. In recent years, the advantage of Federated Learning (FL) in addressing privacy protection challenges and enabling collaborative model training across distributed data sources has made it an increasingly important solution for anomaly detection tasks. This paper provides a comprehensive review of the application of FL in various anomaly detection scenarios, including cybersecurity, intelligent industry, and financial services. It systematically analyzes the performance of popular models such as RNN/LSTM, AE/VAE, GNN, and Transformer-based methods under non-IID data distributions common in federated settings. Furthermore, this study identifies and discusses several critical challenges faced in FL-based anomaly detection, such as data heterogeneity, communication overhead, robustness, and security threats. Corresponding strategies to mitigate these issues are also summarized. Experimental evidence and case studies demonstrate that FL can achieve efficient, privacy-preserving, and scalable distributed anomaly detection, and thus holds significant promise for deployment in real-world applications.