Federated learning is gaining popularity as data are massively generated in a distributed manner. One of the major benefits is to mitigate the privacy risks as the learning of algorithms can be achieved without collecting or sharing data. While federated learning has shown great promise mainly based on the stochastic gradient based optimization, there are still many challenging problems in protecting privacy, especially during the process of gradients update and exchange for a federated optimization. This paper presents the first gradient-free federated learning framework called GRAFFL for learning many distributed models simultaneously to learn a population distribution of partitioned data without moving of it by leveraging deep generative models. Unlike conventional federated learning algorithms based on exchanging parameters or gradients generated as a byproduct of local updating of the shared model, our framework does not require to disassemble a model (i.e., to linear components) or to perturb data (or encryption of data for aggregation) for preserving privacy leakage. Instead, this framework uses implicit information derived from each participating client to generate sufficient summary statistics of its paired samples. They are derived from NSDR network that is a neural network developed in this study to create reduced representations reduced in dimension and containing statistically sufficient information, thereby protecting sensitive information from leakage. By introducing squared-loss mutual information term as an objective and enforcing parameters to be on the Stiefel manifold, this is proved to provide sufficient summary statistics. Generator model in the central server acts as a distributed sufficient summary statistics aggregator without explicit move of it. Using several datasets, the feasibility and usefulness of proposed framework in terms of privacy protection and prediction performance is demonstrated.
Publisher
Ulsan National Institute of Science and Technology (UNIST)