Detecting False Data Injection Attacks in Smart Grids

Abstract

The high rate of transformation of traditional power grids into smart grids has facilitated enhanced monitoring, control, and automation, however, it has also provided a wide cyber-physical attack frontier. The False Data Injection Attacks (FDIAs) are among the most challenging cyber threats, because they can tamper with the measurement data undetected to deceive a state estimation and provoke the appearance of bad data, without causing the presence of the traditional residual-based bad-data detection. Smart grids combine both cyber systems and physical systems to provide efficient power delivery, but leaves the state estimation vulnerable to stealthy FDIAs that cannot be detected by conventional residual-based methods. This review is a detailed and systematic review of attacks of smart grids with an in-depth analysis of FDIAs. Detection methods based on classical and deep learning are reviewed and they are support vector machine, random forests, convolutional neural network, long short term memory network, autoencoders, transformers, and graph neural networks, and their performance is compared. Methods of optimization of features selection and hyperparameter optimization are studied: particle swarm optimization, grey wolf optimization, genetic algorithm, architectural frameworks are considered centralized, distributed, federated, and edgebased. Important performance metrics, assessment challenges, and outstanding problems, such as lack of data, computational effectiveness, adaptability, explainability, and standardization, are determined. The future trends focus on hybrid physicsmachine learning models, lifelong learning, and benchmarks. The work offers researchers and practitioners with a systematic basis of coming up with resilient FDIA defenses to achieve nextgeneration smart grids.