Adversarial attack and defense methods for neural network based state estimation in smart grid

Deep learning has been recently used in safety-critical cyber-physical systems (CPS) such as the smart grid. The security assessment of such learning-based methods within CPS algorithms, however, is still an open problem. Despite existing research on adversarial attacks against deep learning models, only few works are concerned about safety-critical energy CPS, especially the state estimation routine. This paper investigates security issues of neural network based state estimation in the smart grid. Specifically, the problem of adversarial attacks against neural network based state estimation is analysed and an efficient adversarial attack method is proposed. To thwart this attack, two defense methods based on protection and adversarial training, respectively, are proposed further. The experiments demonstrate that the proposed attack method poses a major threat to neural network based state estimation models. In addition, our results present that defense methods can improve the ability of neural network models to defend against such adversarial attacks.