Abstract : This PhD is composed of two main parts. The first one focuses on Internet traffic modelling. From the analysis of many traffic traces, we have proposed a parsimonious model (Gamma-Farima) adapted to aggregated throughput traces and valid for wide range of aggregation levels. In order to produce synthetic traffic from this model, we have also studied the generation of sample path of non-gaussian and long memory stochastic processes. We have then used the Gamma-Farima model in order to build an anomaly detection method. To this end we have introduced a multiresolution model that can differentiate a regular traffic from a malicious one (including a DDoS attack). This method was evaluated both on real traces and simulations. Finally, we have studied the production of long range dependent traffic in a network simulator (NS-2). The second part of this PhD deals with the analysis and synthesis of on-chip traffic, i.e. the traffic occurring in a system on chip. In such systems, the introduction of networks on chip (NOC) has brought the interconnection system on top of the design flow. In order to prototype these NOC rapidly, fast simulations need to be done, and replacing the components by traffic generators is a good way to achieve this purpose. So, we have set up and developed a complete and flexible on-chip traffic generation environment that is able to replay a previously recorded trace, to generate a random load on the network, to produce a stochastic traffic fitted to a reference trace and to take into account traffic phases. Indeed most of the traffic traces we have obtained were non-stationary, we therefore need to split them into reasonably stationary parts in order to perform a meaningful stochastic fit. We have performed many experiments in the SOCLIB simulation environment that demonstrate that i) our traffic generation procedure is correct, ii) our segmentation algorithm provides promising results and iii) multiphase stochastic traffic generation is a good tradeoff between replay and simple random traffic generation. Finally, we have investigated the presence of long memory in the trace as well as the impact of long memory on the NoC performance.