In this paper, machine learning algorithms are first utilized to extract features of campus network traffic, and then the multi-attention mechanism is introduced to fuse the massive features extracted at different scales. Unsupervised learning is used to propose a method for detecting network traffic anomalies, and simulation experiments are conducted to verify the model’s performance. The results show that the detection rates of machine learning algorithms are all above 80%, the false alarm rate basically stays below 10%. The machine algorithms have higher accuracy than other algorithms in network data flow anomaly detection. This study has important reference value for campus network security research and verifies the important role of machine learning algorithms in detecting anomalies in campus network traffic.