HTML5 Webook
184/194

data can be transmitted with low latency. In addition, embedding model inference and training facilities in the network’s edge (Articial Intelligence Edge) [2] can also help solve the challenge of data-driven communication and enable marginal contribution to the entire infrastructure. Bringing training and inference close to hand also protects privacy and secrecy. It provides high security, thereby re-ducing network trac congestion and energy consumption and reducing computational load at the server. To enable knowledge sharing across the edges to generalize AI mod-els, the FL model [3][4] provides a framework for collab-oratively training global statistical models that demand multiple edge participants without accessing private raw data locally.e edge AI enabled in the federated learning frame-work can be depicted in Fig. 1. At the edge, devices can collect user data and pass it to the Edge AI server for local storage and model generation. Sometimes devices can also train and infer directly without edge AI server support, but users oen have multiple device types. We will address these issues in the section related to Edge AI. On top of the advanced AI environment above, the federated learning framework can support collaborative training by collecting locally updated models from the edges and combining them with the global model. In this study, we aim to inte-grate the collaboration between public and private data training to overcome the problems of FL of heterogeneous data distribution for the users to join and limit local data.is paper will provide a comprehensive picture of design FL systems that enable collaboratively training among edge AI and the association between models trained on private and public data. e contributions are in the following items. •We convince the readers that combining edge AI and federated training paradigms distill knowledge from edge to central server. us, better performance in general while keeping the privacy of personal infor-mation on edge. •We provide a system design that works with the IoT environment, processes on edge, and can share among participants. Furthermore, master models can be initially trained and updated through time using public data. ey can be combined with local models trained on private data. •We illustrate applications including environmental, healthcare, and trac monitoring systems in many congurations of federated learning paradigm and edge AI suitable capacity. •We address future research and development in general and direct to NICT innovation promotion which drives economic growth and society auent, safe, and secure. e paper is organized as follows: Following the rst section, the second section will revise the background of federated learning and edge AI within the context of B5G promotion and Society 5.0. Section 3 subsequently intro-duces our current achievement with demonstration and ongoing research in the eld. We will address challenges in Federated learning and edge AI research in section 4 before the conclusion in the last section.Research Background In this context, we describe the perspective of AI on cloud and edge environments, the collaboration, and chal-lenges. As shown in Fig. 2, the dierences and intricacies of AI modeling applications and availability across mobile 2ig. F1 Edge AI collaboration on Federated Learning Paradigmsig. F2 Emerging Federated Learning and Edge AI in Cyber-Physical Spaces Infrastructure 178   情報通信研究機構研究報告 Vol.68 No.2 (2022)4 スマートデータ利活用基盤技術

元のページ  ../index.html#184

このブックを見る