Personalized Federated Learning via Domain Adaptation with an Application to Distributed 3D Printing

AbstractAbstractOver the years, Internet of Things (IoT) devices have become more powerful. This sets forth a unique opportunity to exploit local computing resources to distribute model learning and circumvent the need to share raw data. The underlying distributed and privacy-preserving data analytics approach is often termed federated learning (FL). A key challenge in FL is the heterogeneity across local datasets. In this article, we propose a new personalized FL model, PFL-DA, by adopting the philosophy of domain adaptation. PFL-DA tackles two sources of data heterogeneity at the same time: a covariate and concept shift across local devices. We show, both theoretically and empirically, that PFL-DA overcomes intrinsic shortcomings in state of the art FL approaches and is able to borrow strength across devices while allowing them to retain their own personalized model. As a case study, we apply PFL-DA to distributed desktop 3D printing where we obtain more accurate predictions of printing speed, which can help improve the efficiency of the printers.KEYWORDS: Concept shiftCovariate shiftFederated inferenceInternet of Things3D Printing Supplementary MaterialsThe code for numerical experiments in this article are available in the GitHub repository. In the supplementary material, we will show the proof of Theorem 3.1. We will also present additional experiments and illustrations, mathematical derivations of the proof of concept experiment, and details of our case study implementations.