Federated Learning on Cloud Platforms: Architectures, Challenges, and Opportunities
DOI:
https://doi.org/10.70179/wacm3x29Keywords:
Federated Learning Architecture,Cloud-Based Federated Learning,Privacy-Preserving Machine Learning,Edge-to-Cloud AI Training,Distributed Model Training,Federated Model Aggregation,Secure Multi-Party Computation,Cross-Device Federated Learning,Decentralized AI Systems,Federated Learning Scalability,Data Sovereignty in AI,Federated Learning Challenges,Hybrid Cloud AI Infrastructure,Federated Learning Opportunities,AI Governance in Federated Systems.Abstract
Federated Learning is a novel machine learning paradigm that trains a shared model on decentralized devices. It addresses data privacy and communication efficiency issues, but current federated learning methods mainly focus on developing device-based learning algorithms or improving communication efficiency. There is only limited discussion that explores federated learning in cloud-edge collaborative architecture at a higher level for its key technologies, challenges, applications, and future directions. This work focuses on discussing federated learning in a cloud-edge collaborative architecture and emphasizes system design instead of implementation strategies at the learning algorithm level. The first step is to introduce the federated learning problem, including its components and objectives. The second step introduces the cloud-edge collaborative architecture and discusses the key technologies that support this architecture. Then, discuss the applications of federated learning in the cloud-edge collaborative architecture. The last step summarizes the research challenges and future directions to promote its development.
Federated Learning on Cloud Platforms seems to be seen as a line of work by itself when it’s still unclear to what extent a single server available can be called a “cloud”. Federated learning has been primarily studied in cloud scenarios, where clients exchange their models with a central server in the cloud. Due to this kind of architecture, the cloud can be made low-cost, and therefore accessible to any resource-limited user. In general, taking advantage of a cloud server has been deemed as an architectural decision. Performant training systems can be deployed in the cloud. However, even there, this must be done carefully, in many little details, to strive to attain efficient model training with a high end-to-end throughput.
Decentralized systems bring their own architectural challenges that stem from sheer complexity and dynamics. These difficulties come not only from the incorporation of a higher number of devices but also from the desire and more urgent need for a more diverse set of devices. Decentralized FL seemed to have already been attained in terms of understanding what deep models can do, and federated averaging emerged as a baseline for future decentralized federated training systems. Still, most attention has not been directed so far towards a thorough understanding of the challenges posed by decentralized FL on a cloud platform, which may hinder its deployment.
