부산대학교 도서관

As a cyber-physical attack targeting power systems, False Data Injection Attack (FDIA) has raised widespread concern in recent years. Many FDIA detection approaches in the literature train learning models using historical data to distinguish attacked measurements from normal ones. These approaches typically suffer from severe generalization problems. Although they can achieve good detection performance when the data has a similar distribution in the offline training and online application phases, they often perform poorly when an online operating condition is unforeseen during offline training. To address this problem, this paper proposes a two-layer Dynamic Neural Network (DyNN) framework for FDIA detection. The lower layer is like a traditional FDIA detection model, which uses a Convolutional Long Short-Term Memory (ConvLSTM) network to extract latent features, and employs a Multi-Layer Perceptron (MLP) to make detection decisions based on the extracted features. However, unlike traditional approaches where the internal parameters of the learning model are fixed after training, the internal parameters of the MLP are dynamically generated according to the online data by an upper-layer DyNN adapter. An interactive co-training mechanism is also proposed to coordinate both layers and ensure that the dynamically generated MLP parameters are best suited for FDIA detection in the current operating condition. Extensive experiments demonstrate that the two-layer DyNN framework can achieve satisfactory FDIA detection performance even though the training and test data are collected under completely different operating conditions. In the most extreme cases, the average accuracy and F1 score of the proposed DyNN-based FDIA detection framework are 4.91% and 4.24% higher than that without DyNN, respectively.