Data is the most important part of machine learning. I believe that data should, in most cases, stay decentralized. Firstly, it avoids consolidating monopolies that thrive on early data collection. Secondly, avoiding raw data leakage is still the most important privacy feature. Thirdly, keeping data in its local context is the best way to explore better data exploitation and empower users with their data. My PhD work explored the consequences of decentralization in terms of privacy guarantees, showing that privacy attacks are still possible in this setting, but that differential privacy guarantees implemented at the scale of the node are amplified by decentralization, depending on the graph and algorithm used. More generally, I am interested in the point of view of the user in the federated and decentralized setting: what are the consequences of participating in a training? Under which settings and parameters?