Machine learning is gaining more and more application fields. Different methods exist that allow model constructions for decision support purposes. Nevertheless, machine learning models are vulnerable and exposed to different types of security attacks during the model learning process and after their deployment. Therefore, these threats must be first identified, defined and classified in order to propose defensive measures in the aim to have robust models. In this paper, we study various threats that can affect a machine learning process. We present a classification of threats divided into three parts, the objective, the knowledge and the capability of the attacker. Then, we show some examples of attacks on applications using machine learning.