Intrusion Detection Systems (IDS) play a crucial role in ensuring the security of in-vehicle networks by detecting malicious activities in real time. However, traditional IDS heavily rely on large labeled datasets, which are often scarce, especially when making novel attacks. To address this limitation, we propose a few-shot learning-based IDS that leverages the Model-Agnostic Meta-Learning framework, Long Short-Term Memory (LSTM) networks, and Self-Attention mechanisms. Such an approach enables rapid adaptation to previously unseen attacks with minimal labeled data, significantly improving detection performance in real-world vehicular environments. To facilitate model evaluation, we introduce the FSIDS-IVN dataset, a curated collection derived from publicly available in-vehicle network datasets. Experimental results demonstrate that the proposed method outperforms conventional approaches in terms of accuracy, precision, detection rate, and false positive rate. Notably, when only 5 or 10 labeled samples of a new attack type are available, it achieves detection accuracies of 96.45% and 96.55%, respectively.