The identification of abnormal traffic is essential to network security and service quality. A big-step convolutional neural network traffic detection model based on the attention mechanism is provided as a solution to the significant challenges in abnormal traffic identification caused by feature similarity and the detection model's single dimension. First, the raw traffic is preprocessed and mapped into a two-dimensional grayscale picture after the network traffic characteristics are examined. After that, histogram equalization is used to create multi-channel grayscale pictures. An attention mechanism is then added to give traffic characteristics varying weights in order to improve local features. In order to improve the flaws in convolutional neural networks, including local feature omission and overfitting, pooling-free convolutional neural networks are finally integrated to extract traffic characteristics of various depths. Both a real data collection and a balanced public data set were used for the simulation experiment. The suggested model is contrasted with ANN, CNN, RF, Bayes, and the two most recent models using the widely used method SVM as a baseline. 99.5% accuracy percentage with several classes is achieved experimentally. The best anomaly detection is found in the suggested model. Additionally, the suggested technique performs better in F1, recall, and accuracy than existing models. It is shown that the model is not only effective in detecting things, but also resilient to a variety of complicated contexts.