Invoking an Efficient Deep Learning Approach for Real-Time Detection of False Data Injection Attacks

With the increasing complexity of power systems and their vulnerability to cyber-attacks, robust detection mechanisms are crucial for ensuring the grid's transient and steady-state stability and resiliency. This study introduces an efficient deep learning framework to detect false data injection attacks (FDIAs) in power systems. Leveraging the MATPOWER environment, we simulate a realistic power grid scenario to generate datasets comprising normal operations and various FDIA scenarios. Utilizing advanced deep learning techniques, including convolutional neural networks enhanced with channel attention mechanisms and transformer encoders, our model processes sequential power system data to discern subtle attack signatures from normal fluctuations. The proposed methodology employs a stratified k-fold cross-validation approach on resampled datasets to address class imbalance, ensuring robust model evaluation. The model's performance is meticulously assessed using accuracy metrics and confusion matrices, demonstrating its capability to effectively identify and differentiate between normal and compromised states in power systems. Results on the IEEE 39-bus and 300-bus systems highlight that this approach significantly enhances detection capabilities, achieving high accuracy and practical detection times of less than 2 milliseconds for the IEEE 39-bus system, which addresses the real-time requirements. These promising results suggest that such models can be pivotal for preemptive security measures and real-time monitoring of power grids.