Vehicular Ad hoc Networks (VANETs) are vulnerable to various types of attacks. Intrusion Detection System (IDS) based on machine learning can effectively detect malicious network attacks in VANETs. However, machine learning training necessitates ample data which contain significant ample private information, increasing the risk of privacy disclosure. The privacy protection of training data for machine learning used in the IDS of VANETs is rarely investigated. Meanwhile, Differential Privacy (DP) is one of the most secure privacy protection methods based on perturbations. Therefore, we propose a lightweight hybrid IDS (LH-IDS) based on machine learning and DP. It uses algorithms based on unsupervised learning to detect anomalous network behaviour with high performance, especially unknown attacks in VANETs, while protecting data privacy. The DP is used to secure the privacy of the training data. Noise from different privacy budgets is added to datasets to obtain DP datasets. Subsequently, LH-IDS is used to verify the utility of the DP datasets. Extensive experiments confirm LH-IDS can not only detect anomalous and normal traffic with excellent performance but can also protect the private information of the training data. Additionally, the proposed model incurs only minimal CPU and memory overhead, making it a lightweight solution.