The current network traffic anomaly detection technology based on traditional machine learning is highly dependent on much fully marked data, making it difficult to effectively address the problem of imbalanced traffic training data. In this paper, we propose STAD, a traffic anomaly detection technology based on Siamese Neural Network to address the scarcity of attack samples in traffic training sets. First, we design a K-medoids-based few-shot sampling (KFS) method to extract few representative samples from each class of traffic, hoping to improve the detection reliability of our few-shot model. We then design a traffic anomaly detection method based on a Siamese Multi-Layer Perceptron (MLP). This method utilizes two identical MLP models to train paired traffic samples, capturing non-linear relationships across traffic features and learning the similarities and differences between data. Meanwhile, we also design a cost function containing two specific losses to improve the detection performance. The experimental results based on the CICIDS2017 dataset show that STAD is more suitable for few-shot traffic anomaly detection. Under the condition that only a small amount of malicious traffic is trained, STAD still maintains over 95% detection accuracy and detection rate for attacks.