False Data Injection Attack Detection in Smart Grid Based on Learnable Unified Neighborhood-Based Anomaly Ranking

Abstract

To address the detection of stealthy False Data Injection Attacks (FDIA) that evade traditional detection mechanisms in smart grids, this paper proposes an unsupervised learning framework named SHAP-LUNAR (SHapley Additive ExPlanations-Learnable Unified Neighborhood-based Anomaly Ranking). This framework overcomes the limitations of existing methods, including parameter sensitivity, inefficiency in high-dimensional spaces, dependency on labeled data, and poor interpretability. Key contributions include (1) constructing a lightweight k-nearest neighbor graph through learnable graph aggregation to unify local anomaly detection, significantly reducing sensitivity to core parameters; (2) generating negative samples via boundary uniform sampling to eliminate dependency on real attack labels; (3) integrating SHAP for quantifying feature contributions to achieve feature-level model interpretation. Experimental results on IEEE 14-bus and IEEE 118-bus systems demonstrate F1 scores of 99.40% and 96.79%, respectively, outperforming state-of-the-art baselines. The method combines high precision, strong robustness, and interpretability.