Malicious traffic detection models (MTDMs) prevent cyber-attacks by monitoring the network traffic and detecting threats in network throughput. However, due to the imperceptible difference between the features of normal traffic and malicious traffic, the existed detection methods achieve low detection rate targeting small-scale imbalanced dataset. This paper proposes a feature enhancement-based malicious traffic detection model, namely FE-MTDM. The feature enhancement method groups the features of raw traffic according to the value of Gaussian characteristics, and generates the cluster features with k-means algorithm. The raw features and generated features are feed into a dual classification model that constructed with shallow neural networks and random forests algorithm for detecting the network traffics. In experiments, we use two classic datasets (CIC-IDS-2017 and IOT-23) and simulate the malicious traffic detection targeting small-scale imbalanced dataset. The results show that FE-MTDM has stronger detection ability and reliability than deep learning network (DNN) based models. Regarding to IOT-23, when the percentage of attack samples is 30%, the detection rate of FE-MTDM for DDoS and PortScan is improved 42% and 33% compared with DNN.