Federated learning typically concentrates on how models are trained or aggregated across distributed clients. A new paper on arXiv argues that this overlooks a more fundamental question: what information should leave the client? The authors introduce FedPix, a method that learns what to share, with the goal of reducing leakage.
The paper notes that existing methods often share information, but do not explicitly decide what is necessary. FedPix instead focuses on learning which client information is essential for the federated task, aiming to minimize what leaves the device.
The significance of this approach is that it addresses privacy at the source, before data is transmitted. While the abstract does not provide technical details, the core idea suggests a shift from optimizing training to optimizing sharing policies.