Anti-stealing electricity is an indispensable component of electricity enterprise management. In view of the current problems such as the large number of users, the wide distribution area, the increasing year-on-year power stealing, and the lack of supervision personnel, this paper analyzes and handles the data of power stealing behavior, and proposes a stealing behavior recognition method which can identify the stealing users. First, based on the collected samples, this method adopts a filtering algorithm and a regular threshold to implement feature extraction, so as to improve the effectiveness of the collected data. Then, the user’s stealing behavior diagnosis model is constructed based on the logistic regression algorithm to realize the determination of suspected users. In addition, this paper uses the closed-loop working mechanism which continues to update data by the way of pushing, shooting, processing and feedback to continuously optimize the model. According to the collected data provided by the power consumption information collection system and marketing business application system of State Grid Shandong Electric Power Company, the experimental results prove the feasibility and applicability of the method.