Block ciphers are often protected against side-channel attacks by masking. When traces are available for each key hypothesis, the attacker usually resorts to templates attacks with a profiling phase. Lemke-Rust & Paar suggested at CHES2007 a way to profile templates for Gaussian mixture models, with the use of the well-known Expectation-Maximization (EM) algorithm. In this work, we present a new attack, “unprofiled-EM” (U-EM) that does not use the knowledge of the masks nor requires a profiling phase. This is done by “on-the-fly” regression of the coefficients of a stochastic model using the EM algorithm. Compared to previous methods, it is easy to implement, computa- tionally tractable and efficient in terms of success rate or guessing entropy. We discuss several variations of U-EM and compare their performances on simula- tions and on real DPA contest traces. The best attack scenario depends on the trade-off between measurement noise and epistemic noise.