The growing ubiquity of Internet of Things (IoT) devices has led to an exponential increase in network traffic, making the timely and accurate classification of traffic essential for ensuring network security. Traditional deep learning methods such as CNNs and LSTMs perform well with large-scale labeled data but struggle in few-sample or imbalanced-class scenarios common in real-world IoT environments. Addressing these challenges, this paper proposes a Meta-learning-based IoT Network Traffic classifier (MINT), a novel few-shot learning framework that integrates a neural feature extractor with a relation network and Dynamic Task Augmentation (DTA). MINT effectively learns from limited labeled data and generalizes across multiple classes by extracting high-level semantic features and comparing them in a low-dimensional embedding space. Evaluated on the CICDataset 2017, MINT achieved over 98.5% accuracy and significantly outperformed baseline models in minority class detection, particularly under imbalanced conditions. In the future, we will work to explore zero-shot learning and real-time deployment of MINT on resource-constrained edge devices.