In order to ensure the security of data transmission of classified terminals in power system, an automatic detection system for illegal external connection of classified terminals based on machine learning is designed. Constructing a detection system framework by adopting a machine learning technology, and filtering secret-related information of violation outreach according to a plurality of levels; dividing a main violation outreach mode, and unifying a violation outreach monitoring interface through a modularized loose coupling mode; acquiring a secret-related information clustering center by a clustering algorithm, and dividing a terminal violation outreach data type according to a nonlinear mapping relationship; The information gain rate is calculated based on machine learning, and the illegal outreach of the secret-related terminal is automatically detected based on the feature information entropy. Taking three groups of classified terminals as the test objects, the application effect of the detection system is verified in a variety of illegal outreach modes. The results show that the system can achieve rapid blocking, and has a high recall rate and precision rate to ensure the security of classified information.