FedEVCP: Federated Learning-Based Anomalies Detection for Electric Vehicle Charging Pile

Abstract

AbstractVehicle-to-Grid (V2G) is a technology that enables electric vehicles to use smart charging methods to harness low-cost and renewable energy when it is available, and obtain income by feeding energy back into the grid. With the rise of V2G technology, the use of electric vehicles has begun to increase dramatically, which relies on the reliable Electric Vehicle Charging Pile (EVCP). However, most EVCPs are online and networked, introducing many potential network threats, such as Electricity Theft, Identity Theft and False Data Injection etc. Prior work has mostly focused on machine learning, which is not able to effectively capture the relationships and structures in network traffic, making it difficult to deal with the propagation and infection of the novel network attacks. Moreover, most neural network models collect and transfer data from EVCPs to the central server for training, which makes the central server attractive to attackers. It poses a serious threat to user privacy. To address these issues, propose an anomaly detection model that incorporates Federated Learning and Deep Autoencoder, which can increase the amount and diversity of data used to train deep learning models without compromising privacy. The proposed model forms a layer-by-layer unsupervised representation learning algorithm by autoencoder stacking, while batch normalization of hidden layers accelerates the convergence of the model to avoid overfitting and local optima, and introduces an attention mechanism to enhance key features of sequences composed of data vectors to improve the accuracy rate. To prevent the risk of user privacy leakage on the central server, EVCP is allowed to retain local data for model training and send model parameters to the central server for constructing new global models. Experimental results show that the proposed scheme achieves improved detection accuracy with superior performance than other similar models.