Light-weight federated learning-based anomaly detection for time-series data in industrial control systems

The paper proposed an anomaly detection method that applies the Federated Learning technique to make use of both advantages of distributed learning in different local areas and global updating for all local learning models - FATRAF.

Authors: Huong Thu Truong, Bac Phuong Ta, Quang Anh Le, Dan Minh Nguyen, Cong Thanh Le, Hoang Xuan Nguyen, Ha Thu Do, Hung Tai Nguyen, Kim Phuc Tran.

Link: https://doi.org/10.1016/j.compind.2022.103692

Abstract:

With the emergence of the Industrial Internet of Things (IIoT), potential threats to smart manufacturing systems are increasingly becoming challenging, causing severe damage to production operations and vital industrial assets, even sensitive information. Hence, detecting irregularities for time-series data in industrial control systems that should operate continually is critical, ensuring security and minimizing maintenance costs. In this study, with the hybrid design of Federated learning, Autoencoder, Transformer, and Fourier mixing sublayer, we propose a robust distributed anomaly detection architecture that works more accurately than several most recent anomaly detection solutions within the ICS contexts, whilst being fast learning in minute time scale. This distributed architecture is also proven to achieve lightweight, consume little CPU and memory usage, have low communication costs in terms of bandwidth consumption, which makes it feasible to be deployed on top of edge devices with limited computing capacity.

Keywords: Anomaly detection; ICS; Federated learning; Autoencoder; Transformer; Fourier.

Highlights:

  • A fast-learning model in 20-min time scale that can cope with frequent updating.
  • A light-weight detection scheme in terms of CPU, Memory usage, and running time.
  • Faster system response upon attacks since detection is implemented near the sources.
  • An accurate anomaly detection scheme for time-series data.
  • Federated learning to reduce bandwidth consumption on the link from Edge to Cloud.

Fig 1. Training process

Fig 2. ATRAF Learning Model operation during the testing process

Citation: Huong Thu Truong, Bac Phuong Ta, Quang Anh Le, Dan Minh Nguyen, Cong Thanh Le, Hoang Xuan Nguyen, Ha Thu Do, Hung Tai Nguyen, Kim Phuc Tran, Light-weight federated learning-based anomaly detection for time-series data in industrial control systems, Computers in Industry, Volume 140, 2022, 103692, ISSN 0166-3615.