A Meta-Learning Enabled Method for False Data Injection Attack Detection in Smart Grid

The deep coupling of the power system network layer and the physical layer makes the risk of the power system being subjected to cyber attack constantly rise. Effective cyber attack detection plays an important role in the safe and stable operation of power system. However, due to the limited data available, the problem of cyber attack diagnosis in power system has a weak generalization. To this end, this paper proposes a model-agnostic meta-learning (MAML)-based false data injection attack (FDIA) diagnosis method with limited samples for power systems. More specifically, a basic-learner is first trained to learn the attributes of a series of related FDIA diagnostic tasks. In this training stage, the proposed model can obtain the meta-knowledge from the learning experience of these priori tasks. This technique makes the model have fast adaptation ability to unseen tasks by utilizing only limited data. Then, a meta-learner with fast learning ability is obtained. In addition, two learnable learning rates are applied in basic and meta-learner, which makes the model to converge faster compared with the fixed learning rate. The performance of the proposed FDIA detection model is evaluated on the New England 10-machine 39-bus test system. Experimental results show that the proposed can achieve promising performance with limited data under different scenarios, which can well prove the effectiveness of the proposed model.