Use of federated learning for personal smart media cloud solutions

Abstract

The privacy of user data is a critical concern when it comes to media cloud platforms. Public media cloud services do not guarantee the privacy of their users’ data while offering smart features like facial recognition. Private cloud platforms, while ensuring privacy for stored content, cannot often deliver these smart features with continuously improving accuracy. This paper proposes PicsSmart, a novel approach for personal smart media cloud architecture that addresses these limitations. PicsSmart prioritizes user data privacy by keeping data on-premise within the personal cloud. It leverages federated learning to collaboratively train machine learning models across user devices, enabling continuous improvement of the accuracy of smart features it offers. Unlike traditional cloud platforms, PicsSmart allows for the attachment of various storage solutions, to overcome the storage limitations in cloud platforms. The results of the work demonstrate that PicsSmart effectively delivers smart features with increasing accuracy over time with the use of federated learning. It achieves high performance on diverse and heterogeneous user data while maintaining their privacy. PicsSmart offers a promising solution for users who are concerned about both data privacy and the benefits of smart media cloud functionalities. The implementation of PicsSmart is available at https://github.com/PicsSmart.