Abstract : With the advent of mainstream computing, the Internet and digital photography, many natural images (acquired by a camera) circulate around the world. The images are sometimes altered by a legitimate or illegal information in order to transmit confidential or secret information. In this context, steganography is a method of choice to transmit and to hide information. Therefore, it is necessary to detect the presence of hidden information in natural images. The objective of this thesis is to develop a new statistical approach to perform this detection with the highest reliability possible. As part of this work, the main challenge is to control the probability of detection error. For this purpose, a parametric model locally non-linear of a natural image is developed. This model is built from the physics of optical acquisition system and from the imaged scene. When the parameters of this model are known, a statistical test is proposed and its theoretical optimality properties are established. The main difficulty in the construction of this test is based on the fact that image pixels are always quantified. When any information on the image is not available, it is proposed to linearize the model while respecting the constraint on the probability of false alarm and guaranteeing a loss of optimality bounded. Many experiments on real images have confirmed the relevance of this new approach.