Distributed denial of service (DDoS) attacks are increasing at a rapid pace and have emerged as a major threat to the security and stability of the Internet. Most current mainstream DDoS traffic identification methods rely on deep learning techniques, but still suffer from the same issue as traditional machine learning methods, i.e., over-reliance on statistical features. This paper proposes a DDoS attack identification method based on raw temporal features. The method only requires three basic timing features: direction, time, and length of traffic, with a simple preprocessing process. A deep neural network model was also designed to extract deeper traffic features from the temporal sequence of DDoS traffic and classify them. The method is experimentally evaluated on several real DDoS datasets, and the results show that it achieves more than 98% recognition accuracy across different scenarios.