Publication

Abstract : With the rapid growth of Internet of Vehicles technology, the attack surface has expanded significantly in recent years. Traditional intrusion detection systems, such as signature-based and rule-based approaches, are inadequate for IoV environments due to their limitations in detecting zero-day attacks. To address this challenge, machine learning techniques have been widely explored in the literature for zero-day attack detection. However, the highly imbalanced nature of IoV datasets often leads to biased performance, favoring majority attack classes and reducing detection accuracy for minority classes. This paper proposes a Zero-Day Intrusion Detection System built on a hybrid LSTM–BiLSTM architecture. To mitigate class imbalance, a Generative Adversarial Network is employed to generate synthetic samples. The proposed Z-IDS is evaluated using a cleaned version of the CICIoV2024 dataset under four experimental setups: 1 baseline testing; 2 training with GAN-augmented data; 3 leave-one-class-out testing for unseen attack detection i.e., zero-day attacks; and 4 cross-dataset validation using the Car-Hacking dataset. The model achieves an accuracy of 99.72% with GAN-augmented data and 99.99% in zero-day attack scenarios. These results demonstrate the effectiveness of the proposed LSTM–BiLSTM-based Z-IDS framework in detecting both known and previously unseen attacks within IoV environments.