Abstract : Content distribution systems are developping quickly. In these systems, in addition to the traditionnal information corruptions, arises the problem of packet losses. The need of reliability in this context has lead to the emergence of erasure correcting codes. Thanks to the addition of redundant information, they are able to recover the lost information. In this thesis we address the problem of designing erasure codes having good correction capabilities while having a complexity enabling to reach high throughput. For this, we have chosen to work jointly on the codes and there implementations inside a software codec, and more specificaly on the decoding algorithms. The first part of our work shows that solutions based on Low Density Parity Check (LDPC) codes enable us to get excellent results. In particular, when these codes are decoded with an hybrid Iterative (IT)/Maximum Likelihood (ML) decoder, that enables to obtain correction capabilities close to the optimal whilst keeping an acceptable complexity. Furthemore, we show that thanks to the use of structured LDPC codes, the complexity of the ML decoding can be largely reduced. We then sudy the design of schemes combining both erasure correction and cryptographic features. These systems allow to reduce the global complexity while keeping a high security level. Finaly, we present a fault tolerant scheme for matrix multiplication, based on error correcting codes. This approach allows to build a distributed system on P2P platforms that can tolerate efficiently fail stop and malicious errors.