Deep Learning Based Intrusion Prevention System in Vehicular Network

Abstract

In the internet of vehicles, safety-based communication is carried out for prevention, mitigation, and alleviation of accidents through cooperative messages, position sharing, and the exchange of speed data between the vehicle (nodes) and corresponding roadside units. However, such networks are susceptible to false alarms and mispositioning of vehicles. It is therefore imperative to authenticate and identify normal messages from aggressive and incorrect messages. In this context, this paper has emphasized on a deep learning technique utilizing binary classification for segregating normal and malicious packets. The procedure is initiated by the preparation of training datasets from KDD99 and CICIDS 2018-type of open-source datasets having 1,20,223 packets and 41 features. An autoencoder is used in the preprocessing stage for the elimination of undesirable data right from the beginning. The 23 salient features are filtered out of 41. For training of the models, a structural deep neural network is utilized along with a Softmax classifier and rectified linear unit (ReLU) activation functions. The complete intrusion prevention (IP) mechanism is further trained & tested with Google co-labs as the open platform cloud service with open-source tensor flow. Furthermore, simulation data set developed in a network-simulating procedure is used for validation of the model. The results of the experimentation have established an accuracy 99.57% greater than existing recurrent neural network and convolution Neural Network models. For work in future, various datasets can be used for training to improve the accuracy and efficacy.