Detection to false data for smart grid

Abstract False data injection attack in smart grid might does not launch interference and attack behaviors, so that this attack is difficult found. To address this, this paper proposed a conformal neural network detection method being sensitive to false data. Firstly, using the conformity scores calculated by the mathematical probability to identify the false data. Then, the neural network learns a boundary separating false data and normal data on the conformal region yielded by the conformity scores. Finally, experiments on simulated and real datasets indicate that the proposed method obtains 0.9738 detected accuracy, and the sensitivity to false data reaches 0.9387, defeating against the competitors, moreover, the proposed method does not exhibit exponential detection time as data volume augments. We demonstrate that evaluating the consistency between the data does not rely on data distributions and the operation status in this real-time system like smart grids, since conformity scores can calculate the mathematical probability following the same data distribution. The boundaries learned from conformal regions are independent of data distribution and the information of operation status in smart grids. This evaluation manner of data consistency and that of boundary learning are equally applicable to the identification and separation of those false data injected into other real-time systems.