Enhanced Anomaly Detection in Industrial Control Systems using Hybrid Deep Learning Architectures and Federated Learning

Abstract

Industrial Control Systems (ICS) are becoming more susceptible to cyber attacks, requiring effective anomaly detection mechanisms. In this paper, we introduce a novel anomaly detection framework for ICS networks based on hybrid deep learning models and federated learning. We integrate Long Short-Term Memory (LSTM) networks to analyze temporal sequences with Variational Autoencoders (VAEs) to reconstruct features and detect outliers. In order to meet data privacy requirements and the decentralized environment of ICS deployments, we achieve this using a federated learning paradigm, where model training is conducted across sites without exchanging raw data. The introduced framework is tested on a publicly available ICS dataset and shows better performance than individual deep learning models and conventional anomaly detection methods. The outcomes present enhanced accuracy, precision, and recall in detecting different types of attacks without compromising data privacy. The approach provides a viable and efficient solution towards improving the cybersecurity stance of contemporary ICS ecosystems.