Prior-Free Double Auctions in Incentive Design for Federated Learning

Federated learning (FL) is a paradigm that allows distributed clients to learn a shared machine learning model without sharing their sensitive training data. While FL is successfully applied in environments where a central orchestrator has the primary interest in the shared model’s quality, incentive problems complicate FL in settings with self-interested, heterogeneous clients. In this talk, we model FL incentive design as a digital goods double auction with a submodular production function. Inspired by existing prior-free auction designs, we present a prior-free mechanism that constant-factor approximates optimal revenue. We introduce how “model leakage”, a measure of model sharing between clients beyond the control of the designer, impacts mechanism performance. To facilitate further study, we present code that allows testing incentive designs for federated learning using clients' training on data splits from the MNIST, Fashion-MNIST, and CIFAR-10 datasets. A potential broader impact of this research might be that, for standard learning tasks, incentivized FL might allow for privacy-preserving, sustainable, and scalable data transactions.

Joint work with V. Mugunthan.

Link to paper: https://arxiv.org/abs/2103.14375