Abstract: Collaborative filtering techniques require tremendous amount of personal data to provide personalized recommendation services, which has caused the rising concerns about the risk of privacy leakage. Most existed methods for implementing privacy protection in recommender systems are prone to introduce excessive noises, which significantly degrades the recommendation quality. To address this problem, a matrix factorization algorithm satisfying differential privacy is proposed. The method first converts the matrix factorization problem into two alternate optimization problems, in which user latent factors and item latent factors are optimized respectively. Then a genetic algorithm is introduced to solve these two optimization problems, in which the enhanced exponential mechanism is applied into the individual selection and a novel mutation operation is designed based on the idea of finding important latent factors. Theoretical analysis and experimental results show that the algorithm can not only provide strong differential privacy protection for user data, but also ensure the accuracy of recommendations. Therefore, it has good application value in recommender systems.