Fuzzing is an automatic test generation technique. It consists in repeatedly executing a program with automatically generated inputs, in order to trigger crashes, symptoms of underlying bugs in the code, which can then be fixed. A major challenge in this area is moving from indiscriminate exploration of how programs work to artful guidance towards the discovery of important bugs. Taking advantage of the expertise of our team and our previous work, the objective of the thesis is to propose, implement and evaluate means to meet this challenge, taking advantage of finer-grained guidance mechanisms.