Research Article
Design and Implementation of Distributed Image Recognition App with Federal Learning Techniques
@INPROCEEDINGS{10.1007/978-3-031-20398-5_8, author={Yu-Wei Chan and Bo-You Wu and Yi-Ming Huang and Chao-Tung Yang}, title={Design and Implementation of Distributed Image Recognition App with Federal Learning Techniques}, proceedings={Smart Grid and Internet of Things. 5th EAI International Conference, SGIoT 2021, Virtual Event, December 18-19, 2021, Proceedings}, proceedings_a={SGIOT}, year={2022}, month={11}, keywords={Federated learning Convolutional neural network Image recognition Decentralized deep learning Machine learning}, doi={10.1007/978-3-031-20398-5_8} }- Yu-Wei Chan
Bo-You Wu
Yi-Ming Huang
Chao-Tung Yang
Year: 2022
Design and Implementation of Distributed Image Recognition App with Federal Learning Techniques
SGIOT
Springer
DOI: 10.1007/978-3-031-20398-5_8
Abstract
In recent years, machine learning technology has been widely used in many fields, such as smart transportation, smart healthcare, smart finance and smart cities. Although machine learning technology has brought people a lot of convenience, the privacy problem of user data has also emerged [1]. Considering that users are not necessarily willing to upload personal privacy data to the cloud for deep learning training, therefore, instead of consuming a lot of bandwidth to upload data to the cloud, it is better to train on the local device and then use the model parameters obtained after training. (For example: weights and bias, etc.) upload to the server for aggregation. This emerging machine learning technology is called federated learning. In this way, the privacy and security of data can be guaranteed, and the purpose of decentralized learning can be achieved through aggregation. This study uses the architecture of federated learning technology and convolutional neural network algorithms to implement distributed image recognition mobile applications. This application allows users to use their mobile devices and the central servers for repeated training. After multiple rounds of repeated training, the convergence will be stabilized, and the accuracy will be significantly improved. At the same time, it can take into account privacy and achieve the machine the purpose of learning.
