Abstract: Personalized recommendation system can satisfy the users' demand with pertinent and personalized information, but it is easy to be attacked maliciously by the description file, which will influence the recommendation result. The attribute, model, and classification method of the description files are analyzed and studied. The rough set theory is used to design an algorithm of data pretreatment discretization, decision table reduction, and personalized recommendation treatment. The method of description file classification and attack detection is proposed. The safety of the recommendation system is improved to decrease the influence of the attack on the recommendation results. The frame of personalized recommendation model with dynamic interaction is considered. The example verification proves that the model, the attribute classification, and the detection method of the description files are effective with high accuracy and can effectively improve the safety of the personalized recommendation system.