Leveraging Deep Learning to Increase the Success Rate of DoS Attacks in PMU-Based Automatic Generation Control Systems

The stability of modern power grids relies heavily on effective frequency control, mainly achieved through automatic generation control (AGC) systems. However, AGC systems have become increasingly vulnerable to cyber-physical attacks due to their dependence on communication infrastructure and cyber-physical devices. In this article, we present a novel attack strategy that leverages the learning capabilities of a convolutional neural network and long short-term memory on encrypted data to predict the imminent operating state of a power system. By accurately forecasting vulnerable operational conditions through encrypted traffic analysis, attackers can orchestrate timely denial-of-service (DoS) attacks on AGC systems, significantly amplifying the physical impact of their cyber-attacks. Despite the implementation of network traffic protection measures through the utilization of the IPsec/ESP protocol, which offers encryption at the network layer to ensure the confidentiality of the original packet content and enhances the complexity of traffic analysis, our research demonstrates that attackers can still extract spatiotemporal features from high-entropy encrypted synchrophasor data packets. This finding underscores the inherent limitations of IPsec/ESP protocols in completely obstructing malicious analysis of network traffic and achieving absolute prevention of such nefarious activities. Consequently, the confidentiality of the power system's operational state remains compromised. To gauge the effectiveness of the proposed attack strategy, we conduct various test cases and simulations, revealing a significant increase in the success rate of DoS attacks. These findings underscore the urgency of implementing countermeasures to thwart attackers from exploiting traffic analysis techniques and emphasize the necessity for strengthened security measures in power grid infrastructure.