Since side-channel attacks pose a very serious threat to cryptographic products, leakage assessment plays a key role in the evaluation of the physical security of cryptographic products before putting into the market. However, the widely used Common Criteria (CC) method is both costly and difficult to apply, while the Test Vector Leakage Assessment (TVLA) method fails to accurately detect key leakage points. In this paper, a feature selection technique in machine learning is introduced, with which the leakage points detected by TVLA are treated as features for supervised learning. According to their contribution to the classification, irrelevant features are removed, while the optimal features remained, which identify more important leakage points. We further extend the assessment from two groups to multiple groups, which achieves much better performance. Experimental results show that, compared with existing methods such as TVLA, \(\chi ^2\) -test, TVLA-Bonferroni and ANOVA test, our methods detect leakage points more accurately and stably, and greatly improve the evaluation performance in term of accuracy, leakage hit rate and false positive rate. Besides, our methods require much fewer power traces and shorter time for evaluation than existing methods.