Optimizing and evaluating deep learning techniques for stealthy false data injection attacks on smart grids

The smart grid, as a critical cyber–physical system, is highly susceptible to False Data Injection Attacks (FDIAs), which pose significant threats to its stability and security. This paper introduces an advanced deep learning framework designed to generate stealthier FDIAs targeting state estimation (SE) in power systems. Our approach incorporates enhanced Autoencoders (AE), Variational Autoencoders (VAE), and Conditional Generative Adversarial Networks (cGANs). These models are optimized and enhanced with physics-informed constraints specific to the power system’s SE process. The developed models are evaluated based on bypass rates, convergence rates, and data diversity, highlighting their ability to evade detection mechanisms, such as bad data detectors (BDD) and similarity-based metrics like Jensen–Shannon Divergence (JSD). Simulations on IEEE 14-bus and 57-bus systems using real-world load data demonstrate the models’ ability to generate highly covert FDIAs while adhering to the physical principles of the grid. The results highlight the substantial risks posed by these advanced attacks and provide critical insights into developing more resilient detection strategies for smart grid security.