The explosive growth of GAN and diffusion-based image generators has turned synthetic-face detection into a critical task for forensics’ community. Detectors trained in a centralised setting often falter when confronted with images from unseen generators. We tackle this challenge with a Federated Learning (FL) framework in which several independent clients–each holding its own mix of real faces (from public datasets) and fakes (from different generative models) in a non-identically distributed (non-IID) setting–train locally and share only their model updates. A central server aggregates these updates, refreshes the global model, and measures performance on a large, held-out test set. Experiments show that the FL-trained detector matches the accuracy of a fully centralised baseline while reducing data sharing to zero, and more importantly, generalises more reliably to images created by new synthesis engines. These results highlight FL as a practical, privacy-preserving path toward robust synthetic-face forensics in the fast-evolving deep-fake landscape.