ABSTRACT Vehicular ad hoc networks (VANETs) face critical security threats, including false position reports, replayed messages and denial of service (DoS) attacks, which can disrupt cooperative safety systems. We present a reproducible machine learning–based intrusion detection pipeline that leverages the vehicular reference misbehavior dataset (VeReMi) dataset to build a domain‐aware meta‐learning framework for few‐shot anomaly detection in vehicular ad hoc networks. The system constructs a spatiotemporal domain which covers over 22 million raw telemetry records and generates fixed‐length sequences with extracted feature statistics to enable scalable episodic sampling. A topology aware encoder, trained with a reptile‐style meta‐learning loop, captures relational and temporal patterns essential for identifying malicious behavior. To make the system practical for use on real vehicles, the encoder is combined with small adapter modules. These modules can be quickly fine‐tuned for new driving environments, allowing the system to respond to new types of attacks with very little delay. Tests show steady improvements when using different adapter sizes and small training sets, with an average area under the curve (AUC) score of 0.70 and an F1‐score (F1) of 0.62. The system can adapt in under 1 s. By focusing on the detection of misbehavior in VANETs, this framework have handled the problem of changing conditions, reduces the need for large labeled datasets and offers a reliable and scalable solution for real‐time intrusion detection.
Khan et al. (Thu,) studied this question.