Defense against Power System Time Delay Attacks via Attention-based Multivariate Deep Learning

Time delay attacks pose a threat to power systems that conventional cybersecurity methods do not adequately address. Conventional methods analyze the contents of network packets to identify threats; this is not effective against time delay attacks, which do not alter packet contents. To detect and identify time delay attacks, a new method is needed. In this paper, a novel and data-driven deep learning (DL) approach is developed to detect time delay attacks on power systems and simultaneously identify both the time of attack and attack magnitude. While conventional DL networks struggle with multivariate long time series data generated by power systems, this can be improved using attention mechanisms. In this paper, dual attention mechanisms (DA) are used to focus and improve a gated recurrent unit (GRU) network for detecting and identifying time delay attacks. A comparative analysis shows the proposed GRU-DA approach outperforms conventional DL, machine learning (ML), and statistical methods while maintaining low model complexity.