Intelligent in-car operating system integrates various web platforms including browsers, therefore facing serious issues of malicious web content. Current detection models perform poorly in Internet of Vehicles (IoV) environment due to simple model structures and lack of data samples, which make models difficult to learn and extract multi-level and fine-grained sample features. In this paper, we propose a novel few-shot model, CarMaNet, to detect and classify malicious websites in IoV environment based on website images. We import inception module into Relation networks, which enhance the ability of CarMaNet to learn multi-level features of images. At the same time, we design a feature pipeline featuring adaptive recalibration between the embedding module and the relation module based on adaptive attention mechanism (FCAA), enhancing the model's ability to learn fine-grained features. In few-shot experiments on three dataset of pornography, gambling and counterfeit websites, CarMaNet demonstrates greater performance than current models.