Abstract: To meet the rigid requirements of prior knowledge in template attack, a semi-supervised template attack method based on clustering is investigated, where the selection of power trace feature points is studied, and a feature selection method based on Pearson correlation coefficient and principal component analysis is put forword.In the process of clustering, the assumption of template attack is relaxed through clustering for unmarked data under the guidance of a certain marked information.The effect of factors such as feature selection is studied in a test based on LED encryption, and at the same time, the data dependence on power traces is analyzed.Compared to the traditional semi-supervised template attack method, the result shows that this feature selection method can effectively reduce the effect of the abnormal data and noise, and improve the utilization of the prior information and success rate of key recovery.