Federated Learning for Intrusion Detection in Internet of Vehicles

Abstract

The Internet of Vehicles (IoV) is rapidly evolving, bringing vehicles into a unified network and revolutionizing connectivity and convenience. By enabling seamless communication between vehicles, infrastructure, and cloud services, IoV holds the potential to improve traffic management, road safety, and passenger experiences. However, this increased connectivity also exposes the IoV to new security challenges. Protecting against cyber-attacks on Vehicle-to-Everything (V2X) network communications is now more crucial than ever before. The objective of this senior project is to create an Intrusion Detection System (IDS) tailored for charging stations for electric vehicles, using federated learning (FL) and machine/deep learning methods to improve the detection of abnormal activities and cyber threats. To address IoV challenges, we require complex network traffic analysis and efficient machine learning algorithms for real-time cyber attack prediction. Scaling out a distributed IDS adds further complexity. Our project proposes an FL approach, that decentralizes training and preserves privacy. IoV components can update the intrusion detection model using local data without sharing sensitive information. The project aims to achieve two outcomes: first, developing an IDS capable of detecting and mitigating common vehicular network attacks effectively, with results including performance metrics; second, contributing significantly to advancing cybersecurity in the IoV domain by demonstrating the feasibility of FL for intrusion detection. The privacy-preserving nature of our approach aligns with emerging trends in decentralized cybersecurity solutions, ensuring data security and integrity in IoV environments. The project evaluated 20 machine learning algorithms that led to the selection of ExtraTreesClassifier as the optimal ML algorithm for our global model, which achieved an accuracy of 83.38%. While the specialized scenario-specific submodels attained very high accuracies ranging from 99.91% to 99.95%, the FL approach allowed for a comprehensive and adaptable solution capable of handling the complex and evolving IoV environment.