Graph Neural Networks Based Detection of Stealth False Data Injection Attacks in Smart Grids

Abstract

False data injection attacks (FDIAs) represent a major class of attacks that aim to break the integrity of measurements by injecting false data into the smart metering devices in power grid. To the best of authors' knowledge, no study has attempted to design a detector that automatically models the underlying graph topology and spatially correlated measurement data of the smart grids to better detect cyber attacks. The contributions of this paper to detect and mitigate FDIAs are twofold. First, we present a generic, localized, and stealth (unobservable) attack generation methodology and a publicly accessible dataset for researchers to develop and test their algorithms. Second, we propose a Graph Neural Network (GNN) based, scalable and real-time detector of FDIAs that efficiently combines model-driven and data-driven approaches by incorporating the inherent physical connections of modern AC power grids and exploiting the spatial correlations of the measurement data. It is experimentally verified by comparing the proposed GNN based detector with the currently available FDIA detectors in literature that our algorithm outperforms the best available solutions by 6.21%, 0.69%, and 2.73% in detection rate and by 3.65%, 0.34% and 1.38% in F1 score for standard IEEE testbeds with 14, 118, and 300 buses, respectively. NOMENCLATURE P i + jQ i Complex power injection at bus i. P ij + jQ ij Complex power flow between bus i and j. V i , θ i Voltage magnitude and phase angle of bus i. θ ij θ i − θ j . G ij + jB ij ijth elements of bus admittance matrix. g ij + jb ij Series branch admittance between bus i and j. g si + jb si Shunt branch admittance at bus i.Original state vector without an attack. x ∈ R n False data injected state vector. z ∈ R m A measurement vector. z o ∈ R m Original measurement vector. z a ∈ R m Attacked measurement vector. h(x)Nonlinear measurement function at x. H ∈ R m×n Jacobian matrix.

Graph Neural Networks in IoT: A Survey

“…Then they developed a context graph convolution block and a soft clustering graph convolution block to capture both local and global spatial dependencies between parking lots. Similar to this work, authors in the following studies [30,128,207,264,284,321] also created graph representations from geospatial sensors. They used the located cameras and satellite as nodes to create graph structures to perform traffic network prediction.…”

“…For example, in a multi-modal sensor network (e.g., lighting, environment), each sensor can be represented as a node in a graph, and their latent interconnections need to be learned by using data-driven approaches. Established works [19,25,30,40,42,56,58,71,160,161,175,201,207,208,239,258,284,286,290,318,320,321,334] illustrated the performance of applied GNNs in smart city applications that involved IoT sensor interconnections. Table 3 summarizes the sensor infrastructures, GNN models, and learning targets in the collected works.…”

“…Based on the node representation, adding edges between two related sensors can 1) clearly state the geospatial connection with natural setting between sensors and 2) help learn implicit interconnections through data-driven approaches. In our surveyed papers, authors in [30,128,207,264,284,321] use sensor location as node representation, and geospatial connection as edge connection to build the graph structure.In geospatial implication scenarios where cameras are deployed in the city and smart meters are placed in the electricity grid in fixed locations, it is much easier and more accurate to model the graph using sensor topology based on physical sensor information. As for the implicit interconnection scenarios, edge connections can be learned through data-driven approaches to connect two sensors.…”

“…There are two types of GNNs in our surveyed studies as shown in Table 3. The first type of GNNs for ISI is spatio-temporal GNNs, which are made of static structures and time-varying features, and such information in a graph requires a neural network that can deal with time-varying features as demonstrated in the these works [30,40,58,128,201,207,264,284,321]. The second type of GNNs for ISI is graph autoencoder, which maps graph data into a low-dimensional space for graph analytics [56,239,284,320].…”