Recommendation systems, which can effectively filter information based on user preferences, has been applied widely. The problem of cold start and data sparsity becomes more and more serious with the explosive growth of the number of users. Multi-source data fusion, which can effectively alleviate the recommendation accuracy under the conditions of data sparsity and the cold start problem, is favored by researchers. Its main idea is to fuse auxiliary information of users in other aspects for missing values filling to optimize the accuracy of target recommendation service. Nevertheless, more serious risk of privacy disclosure is introduced due to the relations between data. To solve the above problems, this study proposes a deep cross-domain recommendation model with privacy protection. In detail, a deep learning collaborative recommendation method is designed featuring multi-source data fusion and differential privacy protection. On the one hand, this method fuses auxiliary domain information to improve the accuracy of recommendation and corrects the deviation of abnormal points to improve the performance of the recommender system; on the other hand, this method adds noise in the collaborative training process based on differential privacy model to solve the data security problem in data fusion. In order to evaluate the long tail effect in the recommendation system, this study proposes a new metric—discovery degree for the first time, which is used to measure the ability of the recommendation algorithm to find users’ invisible requirements. Based on the performance comparison and analysis of the existing algorithms, the results show that the proposed method has better recommendation accuracy and diversity than the existing methods on the premise of ensuring privacy security, and can effectively discover the hidden needs of users.