Hybrid Deep Network Based Multi-Source Sensing Data Fusion for FDIA Detection in Smart Grid

Abstract

The false data injection attack (FDIA) can cause the unstable operation of the power grid by injecting false data into the power grid, which brings serious challenges to the modern new power system. However, diverse data redundancy and inconsistent temporal in the frequent interactive data of power sensors significantly improves the difficulty of detecting false data injection attacks. To solve this aforementioned problem, we propose a multi-sensing data fusion model based on hybrid deep learning network. Firstly, temporal alignment technique is employed to preprocess the original multi-source perceived data in time dimension. Subsequently, a long short-term memory based convolution neural network (CNN-LSTM) is designed to extract spatial and temporal feature from different sensor data, which can effectively represent the spatial and temporal distribution of multi-source data in the same temporal. Furthermore, through sequential convolution operation, an independent LSTM neural network is introduced to fuse multi-source features by further extracting deep temporal information, which can efficiently remove the redundancy in multi-sensor heterogeneous data and be used to FDIA attack detection in cyber physical systems. Extensive experiments demonstrate that our fusion method can improve data effectiveness with respect to original multi-source sensing data, and perform a superior detection performance for FDIA attacks in power systems.