Machine Learning Based Misbehaviour Detection in VANET Using Consecutive BSM Approach

Abstract

Vehicular ad-hoc network (VANET) is an emerging technology for vehicle-to-vehicle communication vital for reducing road accidents and traffic congestion in an Intelligent Transportation System (ITS). VANET communication is vulnerable to various attacks and cryptographic techniques are commonly used for message integrity and authentication of vehicles. However, cryptograhpic techniques alone may not be sufficient to protect against insider attacks. Many VANET safety applications rely on periodic broadcast of basic safety messages (BSMs) from surrounding vehicles that contain important status information about a vehicle such as its position, speed, and heading. If an attacker (misbehaving vehicle) injects false position information in a BSM, it can lead to serious consequences including traffic congestion or even accidents. Therefore, it is imperative to accurately detect and identify such attackers to ensure safety in the network. This paper presents a novel data-centric approach to detect position falsification attacks, using machine learning (ML) algorithms. Unlike existing techniques, the proposed approach combines information from 2 consecutive BSMs for training and testing. Simulations using the Vehicular Reference Misbehavior (VeReMi) dataset demonstrate that the proposed model clearly outperforms existing approaches for identifying a range of different attack types.

Reliable Detection of Location Spoofing and Variation Attacks

“…We chose VeReMi since it provides location spoofing attack instances together with genuine samples with unmanipulated coordinates. Several previous studies [16], [17], [18], [19], [20] employed the VeReMi dataset to measure detection performance against spoofing attacks, and it was reported that conventional machine learning (ML) methods are able to yield highly accurate detection rates, e.g., >95% with a random forest classifier against five attack types [17]. However, our preliminary study using the same methodology and replicating the previous research reveals somewhat lower performance than what has been reported in previous studies, signaling the need for a thorough investigation into the problem of location spoofing attacks.…”

A Comprehensive Survey of V2X Cybersecurity Mechanisms and Future Research Paths

“…Existing V2X threat and attack detection approaches mainly depend on reactive mechanisms in order to balance communication cost and security overhead. The majority of those techniques employ supervised ML for known attacks using labelled V2X datasets [320], [322], [323], [324], [325]. The work in [320] employs supervised learning algorithms selected in the WEKA toolset [338] to detect and classify misbehaving vehicles (e.g., position and identity spoofing and replay attacks) based on physical properties and message content.…”

“…Supervised ML techniques (e.g., KNN, decision trees and LR) are also considered in [132] and [328] to quickly detect vehicles transmitting false alerts and position falsification. Using an augmented feature set which combines information from successive BSMs in VeReMi, the authors in [325] achieve improved detection performance for position falsification attacks compared to existing ML-based approaches. Position-related features are also used in [327] to extend VeReMi for a decentralized detection of position falsification attacks.…”

Roadside unit‐based pseudonym authentication in vehicular ad hoc network

“…Additionally, we note that the proposed work is not designed to detect “rogue” nodes, that is, nodes with valid credentials that intentionally insert false information in the BSMs. Detecting such rogue nodes is an important problem and in a separate related work 50 we have addressed how such misbehavior can be detected using different machine learning (ML) algorithms. The trained ML models can be deployed at the RSUs and run alongside the proposed authentication mechanism in order to detect rogue nodes.…”