Under the assumption that power leakages at different interesting points follow multivariate normal distribution, maximum likelihood principle (MLP) can be used as an efficient distinguisher for template attack (TA). Therefore, in key‐recovery, one uses MLP to recover the correct key. In pattern recognition, Mahalanobis distance similarity measure (MDSM) is usually used to measure the similarity of two vectors in terms of their distance. A merit of MDSM is that, when measuring the similarity of two vectors, one takes the cross correlation between different variables into consideration. In this paper, we investigate the application of MDSM as a distinguisher in TA. We will show that there exists a certain relationship between MLP‐based TA and MDSM‐based TA under the assumption that the covariance matrices of different templates are identical. However, in MDSM‐based TA, power leakages at different interesting points are not required to follow multivariate normal distribution. We perform practical experiments to evaluate the key‐recovery efficiency of MDSM‐based TA. Experimental results verify that, in the same attack scenario, the key‐recovery efficiency of MDSM‐based TA can be higher than that of MLP‐based TA. Copyright © 2014 John Wiley & Sons, Ltd.