CAREER: Foundations of Federated Multi-Task Learning

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Mobile phones, wearable devices, and smart homes form just a few of the modern distributed networks generating a wealth of data each day. Due to the growing computational power of edge devices, coupled with concerns over transmitting private data, it is increasingly attractive to store data locally and push network computation to the edge. Federated learning explores training machine learning models at the edge in distributed networks. While federated learning has shown tremendous promise for enabling edge applications, practical deployment is currently stymied by a number of competing constraints. In addition to being accurate, federated learning methods must scale to potentially massive networks of devices, and must exhibit trustworthy behavior---addressing pragmatic concerns related to issues such as user privacy, accuracy, and robustness. In this project, we explore multi-task learning, a technique that learns separate but related models for each device in the network, as a unified approach to address the competing constraints of federated learning. The objective of the project is to develop scalable multi-task learning methods that are suitable for practical federated networks, and to rigorously study the foundational properties of federated multi-task learning in terms of the goals of accuracy, scalability, and trustworthiness. In doing so, the research will unlock a new generation of federated learning systems that can holistically address the constraints of realistic federated networks. The goal of this project is to establish and rigorously study the use of federated multi-task learning. While the accuracy benefits of federated multi-task learning are well-known, the work charts two new directions. First, the project develops methods to realize multi-task learning at scale in massive federated networks. Secondly, the project shows that multi-task learning, by improving privacy, accuracy, and robustness, is in fact key for trustworthy federated learning. The technical aims of the project work are divided into three thrusts. First, by approximating standard notions of multi-task learning, the project will develop and rigorously study a family of highly scalable federated multi-task learning objectives. Second, the privacy implications of multi-task learning will be analyzed and evaluated in order to understand trade-offs between privacy and utility in federated networks. Finally, this project will explore tensions between accuracy (in terms of performance disparities across devices) and robustness (to data and model poisoning attacks) in federated learning. Although these goals may be at odds, this project aims to show that multi-task learning can inherently improve both fairness and robustness, helping both to be achieved jointly. Taken together, this work has the potential to cause a paradigm-shift in the way federated learning systems are designed, implemented, and analyzed. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.