Locational detection of the false data injection attacks via semi-supervised multi-label adversarial network

Abstract

False data injection attacks (FDIAs) pose a significant threat to the healthy and safe operation of smart grids. Traditional fdia detection methods are difficult to deal with complex and harsh scenarios. Considering the increasing demand for attack location services in practical power grid management, this paper proposes a new adversarial scheme called Semi-Supervised Multi-Label Adversarial Network (SMAN). To better adapt to harsh learning conditions with only a few labeled samples, this scheme combines semi-supervised mechanisms with state-of-the-art generative adversarial networks. The showdown between generator and discriminator aims to optimize the model for high detection accuracy in semi-supervised training. Among them, for the non-Euclidean structure of the power system, we propose a graph attention-based generation mechanism to improve the generator and enhance the authenticity of the generated samples. Furthermore, the classification network based on label correlation of FDIAs is proposed to capture the inconsistency and co-occurrence dependency in the measurements due to the potential attacks, which is applied to detect the exact injection locations with high dimension data. To this end, we perform label transformation and hierarchical training on the localization task to quickly localize data attacks based on multi-label classification. Extensive simulations and comparisons implemented on IEEE 14-bus and 118-bus power systems demonstrate the superiority of this scheme in localization detection. And it has been proved that it has high robustness and generalization ability under harsh conditions.