The advent of mass communication has led to the ubiquitous presence of information. Secure treatment of these data has become a main concern not only for governments but also for companies and individuals. In this context, cryptography studies the mathematical techniques related to the security of information. In this thesis, we mainly focus on the way data confidentiality can be reached through the use of encryption algorithms such as block ciphers. Encryption refers to the process of transforming some piece of information, using an algorithm, to make it «unreadable» to anyone except those possessing the secret key. The goal of cryptanalysis is to scrutinize encryption algorithms in order to detect possible weaknesses. This thesis investigates one major class of statistical attacks against block ciphers, denoted as linear cryptanalysis, and some of its extensions. Our contribution is threefold: In the first part of the thesis, we present two tools developed to facilitate the application of linear cryptanalysis and other techniques. The first tool is a heuristic algorithm for generating good linear approximations for block ciphers. The second tool allows to decrease the time complexity of linear cryptanalysis. In the second part, we investigate the cryptanalysis of PRESENT, a 64-bit block cipher designed for constrained environments. We introduce a dedicated attack denoted Statistical Saturation Attack and provide an analysis of its expected complexity. Finally, we apply the tools developed in the first part to different versions of a small block cipher. On the basis of those practical experiments, we assess various fundamental assumptions commonly used in linear cryptanalysis and its extensions. This last part opens new research directions for the design and analysis of block ciphers. (FSA 3) -- UCL, 2011