Abstract : Recent works have shown that the mutual information is a generic side-channel distinguisher, since it detects any kind of statistical dependency between leakage observations and hypotheses on the secret. In this study the mutual information analysis (MIA) is tested in a noisy real world design. It indeed appears to be a powerful approach to break unprotected implementations. However, we observe that the MIA fails when applied on a DES cryptoprocessor with masked substitution boxes (Sboxes) in ROM. Nevertheless, this masking implementation remains sensitive to Higher-Order Differential Power Analysis (HO-DPA). For instance, an attack based on a variance analysis clearly shows the vulnerabilities of a first order masking countermeasure. We propose a novel approach to information-theoretic HO attacks, called the Entropy-based Power Analysis (EPA). This new attack gives a greatest importance to highly informative partitions and in the meantime better distinguishes between the key hypotheses. A thorough empirical evaluation of the proposed attack confirms the overwhelming advantage of this new approach when compared with MIA.