To improve the accuracy of mining results, this paper proposes a method of privacy preserving frequent pattern mining, based on sample learning and synchronized randomization of itemset (LS-PPFM). The method utilizes the data of individuals who do not care privacy. First, the data that does not need protecting are learned, and some strongly associated items are obtained. Then, when the data is randomized, the associated items are bound together and randomized synchronously to try to keep their potential associations. Experimental results show that compared with independent randomization, LS-PPFM can achieve significant improvements on accuracy, while losing a little privacy.