Innovative defense strategies: Fusion deep learning approach to counter false data injection attacks in power systems

A pervasive threat to power systems is the occurrence of False Data Injection Attacks (FDIA), which can disrupt the underlying control infrastructure and endanger the reliability of the system. Many efforts have been made to combat FDIAs. However, most of the reported methods have some shortcomings that prevent them from being used in real-world applications, such as scalability issues, resource constraints, and privacy concerns, which hinder their practical applicability. In this paper, a novel Fusion Deep Learning (FDL) method is applied to the detection of FDIAs, enabling real-time deployment with high accuracy, low memory overhead, and fast model convergence. The proposed framework employs a confidence-aware, memory-based fusion strategy that enhances robustness to non-IID data, concept drift, and realistic disturbances. It demonstrates strong generalization by accurately distinguishing cyber-attacks from natural variations such as sensor noise and topological changes, thereby reducing false positives. More importantly, no data is transmitted to the control center during optimization of the FDIA detection model to maintain data privacy. The performance of the proposed method is compared with the Centralized Deep Learning model based on Convolutional Neural Network as a representative of centralized approaches. The evaluation of the proposed method was performed on the IEEE 118-bus power network, a real dataset of the New York Independent System Operator. Through numerical evaluation, our findings not only affirm the superiority of FDL in accuracy, data storage efficiency, and convergence speed but also underscore its potential to revolutionize security strategies in modern power infrastructures.