User privacy is getting more and more important. We want to improve user experience while receiving less privacy-sensitive information. Federated Learning (FL) is one of the promising technologies that enable this scenario. FL has user devices participate into the model training process. User devices send locally trained models on behalf of raw data. The collected local models are used to update "a" global ML model, which is again distributed to individual users. FL can also be combined with Differential Privacy (DP), to make exploiting user privacy more difficult.
We apply FL in LINE messenger's keyboard area to make users sticker selection easier and more personalized while preserving user privacy. In this talk, we'll present how we design a whole system in practice.
