Federated Learning of Generative Adversarial Networks with Resource Constraints and Unreliable Communication
Abstract
Generative Adversarial Networks (GANs) are designed to generate synthetic examples in a target problem space, such as imagery, sound, video, and so on. Traditional approaches for learning GANs are entirely centralized, and assume a central data repository for GAN training. Such centralization, however, is often impractical due to concerns about communication overhead, sensitivity of the training data, distributional heterogeneity, and the dynamic nature of the data streams. Federated learning is a promising alternative paradigm, but has been explored predominantly in supervised learning settings and tends to assume synchrony and relative homogeneityof clients, as well as full client participation. In practice, heterogeneity among clients in their computational and communication capabilities, individual incentives, as well as unreliable communication, make typical existing federation schemes in the context of GANs inadequate. In this project, we will develop novel federated GAN learning schemes that leverage algorithm and system co-design to maximize learning efficacy, accounting for the incentives of clients to fully participate in the scheme. Specifically, we will first explore the system design space for federated GAN learning through the lens of decentralized topology and aggregation algorithms, in order to provide adaptivity to heterogeneous data streams and manage performance tradeoffs. Next, we will tackle unreliable communication through effective asynchronous GAN federated training. Finally, we will analyze and design incentives for heterogeneous clients to maximally participate in the federated scheme.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412663
Entities
People
- Yevgeniy Vorobeychik
Organizations
- Office of Naval Research
- United States Navy
- Washington University in St. Louis