ADVENT: Attack/Anomaly Detection in VANETs

TL;DR

This paper presents ADVENT, a real-time, privacy-preserving system for detecting attacks and identifying malicious vehicles in VANETs, achieving high accuracy through integrated statistical, machine learning, and federated learning techniques.

Contribution

It introduces a novel system that simultaneously detects attack onset and identifies malicious nodes in VANETs, combining statistical, machine learning, and federated learning methods for improved performance.

Findings

F1-score of 99.66% for attack detection

F1-score of 97.85% for malicious node identification

Reduced false negatives with federated learning integration

Abstract

In the domain of Vehicular Ad hoc Networks (VANETs), where the imperative of having a real-world malicious detector capable of detecting attacks in real-time and unveiling their perpetrators is crucial, our study introduces a system with this goal. This system is designed for real-time detection of malicious behavior, addressing the critical need to first identify the onset of attacks and subsequently the responsible actors. Prior work in this area have never addressed both requirements, which we believe are necessary for real world deployment, simultaneously. By seamlessly integrating statistical and machine learning techniques, the proposed system prioritizes simplicity and efficiency. It excels in swiftly detecting attack onsets with a remarkable F1-score of 99.66%, subsequently identifying malicious vehicles with an average F1-score of approximately 97.85%. Incorporating federated learning in both stages enhances privacy and improves the efficiency of malicious node detection, effectively reducing the false negative rate.