With the widespread adoption of intelligent connected vehicles (ICV) and the rapid development of the internet of vehicles (IoV), the security of in-vehicle network (IVN) has become increasingly critical. In particular, vulnerabilities in in-vehicle bus systems have become more exposed, with networks facing a growing range of cyber threats, including denial-of-service (DoS) attacks, spoofing attacks and fuzzing attacks. This paper proposes a spatial-temporal collaborative intrusion detection method for IVN based on federated learning (FL), aiming to address the limitations of traditional intrusion detection methods in data privacy protection, temporal modeling, and computational efficiency. The method employs an autoencoder (AE) to achieve feature compression, reducing data dimensionality and extracting core spatial features. Additionally, a temporal embedding transformer (TET) model is used to capture dynamic temporal dependencies, overcoming the shortcomings of traditional positional encodings in processing dynamic signals in vehicular networks. The model is trained and tested in FL framework to ensure data privacy security. Experimental results on the Car-Hacking dataset demonstrate that FL-TET achieves 100% accuracy, precision, recall, and F1 score, outperforming existing centralized and FL methods. Notably, the model parameters are reduced to just 0.306M, representing an 86.3% reduction compared to current FL methods. By enabling privacy-preserving and efficient intrusion detection, FL-TET has the potential to improve connected-vehicle security, enhancing the safety and reliability of autonomous and networked transportation systems.