A Lightweight Time-Adaptive Data-Driven Method for Detecting Multi-False Data Injection in Smart Meters

False data injection attacks (FDIA) in smart meters are growing concerns. Conventional detection methods are difficult to be deployed on resource-limited edge devices and insufficiently adaptive to different attack scenarios and random attack commencement times. To address these issues, this paper proposes a lightweight time-adaptive data-driven FDIA detection method. First, a lightweight hybrid FDIA detector is developed, integrating the MobileNet and a statistical indicator computing block for features extraction, and a deep learning process for classification. The load data is dimensionally augmented by recurrence plots (RPs) and fed into a series of detectors operating at different time scales. Then, a time-adaptive structure to accommodate diverse attack types and timing is designed, comprising multiple FDIA detectors and a credibility evaluation process. The detection decision can be made at an appropriately early time and with high accuracy. To optimize the overall performance, NSGA-II is used to search trade-off solutions between detection speed and accuracy. Simulation results on a dataset including 20 residents and eight types of FDIA demonstrate that the proposed method outperforms benchmarks in terms of detection accuracy and speed. Moreover, the proposed method can adapt to different detection scenarios by offering different solutions with balanced speed-accuracy.