FSProtoTransfer: Synergizing Few-Shot, Prototypical Networks, and Transfer Learning for Intrusion Detection in VANETs

Abstract

The Internet of Vehicles (IoV) integrates Vehicular Ad Hoc Networks (VANETs) to enhance the safety and efficiency of vehicular transportation. However, the distinct characteristics of IoV networks make them vulnerable to a range of attacks that compromise their confidentiality and functionality. To address these threats, Intrusion Detection Systems (IDSs) have been widely proposed as countermeasures. Recently, Machine Learning (ML) and Deep Learning (DL)-based IDSs have gained significant attention for their ability to enhance the security of IoV and related vehicular networks. However, traditional ML/DL models rely heavily on large, labeled datasets for training, which are both costly and time-consuming to obtain. The challenge is further compounded by class imbalance within datasets, which skews the model's learning toward majority classes, leading to poor detection of attacks from minority classes. This challenge is particularly significant in VANETs environments, where the unique topology and diverse communication methods increase susceptibility to various cyberattacks. In this paper, we propose a novel framework FSProtoTransfer, to detect intrusions in VANETs environment. FSProtoTransfer consists of three phases. The first phase utilizes self-supervised learning to extract latent patterns and robust representations from unlabeled data. In the second phase, few-shot learning with a prototypical network is applied to train the pre-trained model, enabling it to learn effectively from a small number of labeled examples and reducing the reliance on large labeled datasets. The third phase leverages transfer learning with a Multi-Layer Perceptron (MLP) to perform the final classification. We conducted extensive experiments on two different VANETs datasets. Our experimental results demonstrate that FSProtoTransfer, utilizing only 20% of labeled data, outperforms fully supervised state-of-the-art models by 41%, 78% and 70% in terms of precision, recall, and F1 score, respectively.