Federated learning is an up-and-coming technique in dealing with the limitations related to handling personal information with cloud-based machine learning.
We are looking to adopt federated learning to assist various features in LINE especially in regards to the handling of private user information. Our usage requires simultaneous operation of multiple federated learning instances. In particular, the client platform of federated learning should support user data collection, model management, and training result uploading. In limited mobile resource environments, we also need training schedule management and on-device training capabilities. Last but not least, we also need specialized task management to avoid excessive battery drain and a compromised user experience.
In this presentation, I will share how the LINE federated learning client platform meets these requirements, and what lessons we learned along the way.