Data-Driven Cyber Anomaly Detection Using RBM and Features-Based Residual LSTM with Multi-Head Attention in Cyber Physical Systems

Cyberattacks on cyber-physical systems (CPS) might result in complicated situations, it is crucial to anticipate them early on. The ability of security systems to assess unknown threats is one of the many problems that need to be resolved in this situation. Machine Learning (ML) methodologies based on semi-supervised and system behavior analysis techniques address these problems. A novel hybrid model called RLSTMMHA, which combines Residual Long Short-Term Memory (RLSTM) with Multi-Head Attention (MHA) for classification and Restricted Boltzmann Machine (RBM) for feature extraction, is presented in this paper as a data-driven method for cyber anomaly and attack detection in Cyber-Physical Systems (CPS). The suggested approach tackles the difficulties in identifying unknown assaults in CPS when there is a shortage of real-world labeled data. Experimental evaluations on three public CPS datasets demonstrate superior performance in accuracy with 95.17% and more robustness compared to classical models.