Fed-SMAE: Federated-Learning Based Time Series Anomaly Detection with Shared Memory Augmented Autoencoder
Jie Huang, Danya Xu, Tao Yang · 2024
Time series anomaly detection plays a critical role in ensuring the security of Cyber-Physical Systems (CPS). However, the growing complexity of data acquired from CPS poses significant challenges to conventional anomaly detection methods. Deep learning-based anomaly detection has garnered significant attention and research interest due to its ability to discern intricate patterns and extract meaningful features from complex data. However, centralized anomaly detection methods currently are challenging due to the privacy protection issue and the difficulty of real-time data acquisition in the distributed scenario. In addition, conventional reconstruction-based methods are difficult to accurately identify anomalies due to the overfitting problem of the model. To address these issues, we proposed a novel method called federated shared memory augmented autoencoder (Fed-SMAE) to perform federated learning-based anomaly detection for multivariate time series data. First, we design the distributed training framework to address the privacy protection issues, which consists of a central server and multiple edge node devices. On the edge device, we design memory-augmented autoencoder (MAE) based on Long Short Term Memory (LSTM) to improve the performance of AE and perform local device anomaly detection. On the central server side, federated averaging algorithm (FedAvg) is employed to aggregate the model parameters from various edge devices. Meanwhile, the shared memory module mechanism is proposed to combine the MAE model with the distributed training framework. Finally, experimental results demonstrate that the proposed method outperforms the existing methods.