Securing VANETs: Multi-Objective Intrusion Detection With Variational Autoencoders

The rising prevalence of connected vehicles in Vehicular Ad-hoc Networks (VANETs) within Intelligent Transportation Systems (ITS) has introduced a heightened susceptibility to various cyber threats, particularly zero-day attacks. The magnitude of intercommunicating vehicles compounds the difficulty of identifying dynamic and spatially-distant anomalies that elude static rules. Existing solutions, often rule-bound and inflexible, face challenges in dealing with novel threats. Hence, the diverse and evolving nature of potential threats underscores the necessity for more adaptive and robust detection frameworks, especially when considering the end consumer's safety and privacy concerns.This paper introduces an innovative anomaly detection framework designed for VANETs. It employs a variational autoencoder (VAE) and optimizes multiple objectives: divergence, KL-divergence, and reconstruction error, using the AGE-MOEA and R-NSGA-III algorithms. Deployable within Roadside Units, it captures and analyzes broadcast vehicular data sequences, classifying messages as anomalous or normal to address not only the technological intricacies but also the paramount concern of consumer safety within the vehicular ecosystem. Experiments involve exploring various hyperparameters, with performance assessed using key metrics. The proposed framework undergoes comprehensive benchmarking against prior research, considering accuracy, precision, recall, and ROC-AUC. This underscores the efficacy of fully unsupervised learning and multi-objective optimization in enhancing VANETs security against emerging threats.