Adaptive false data injection attack detection in load frequency control using recurrent neural networks

This paper proposes an adaptive detection method for false data injection attacks based on the retraining of recurrent neural networks (RNNs). Machine learning-based detection methods heavily depend on historical pretraining data. Because of this dependency, the detection performance degrades when the power system largely fluctuates and becomes unexpected states. We propose retraining with new data obtained under unexpected states to avoid this serious degradation. To fasten the retraining, well-known RNNs such as long short-term memory (LSTM) and gated recurrent unit (GRU) should reduce the number of epochs, which prevents them from achieving high performance. To overcome this tradeoff, we utilize the echo state network (ESN), which is a representative model of Reservoir Computing. ESN significantly reduces training time compared with LSTM and GRU. We evaluate the proposed method with the three neural networks using the IEEE 68 bus system, confirming the performance of detection significantly degrades when the system enters an unexpected state. We also show the proposed retraining is effective in recovering the performance, and ESN outperforms the other RNNs in terms of retraining time.