Evaluating the Communication Efficiency in Federated Learning Algorithms
Authors:
Muhammad Asad,
Ahmed Moustafa,
Takayuki Ito,
Muhammad Aslam
Abstract:
In the era of advanced technologies, mobile devices are equipped with computing and sensing capabilities that gather excessive amounts of data. These amounts of data are suitable for training different learning models. Cooperated with advancements in Deep Learning (DL), these learning models empower numerous useful applications, e.g., image processing, speech recognition, healthcare, vehicular net…
▽ More
In the era of advanced technologies, mobile devices are equipped with computing and sensing capabilities that gather excessive amounts of data. These amounts of data are suitable for training different learning models. Cooperated with advancements in Deep Learning (DL), these learning models empower numerous useful applications, e.g., image processing, speech recognition, healthcare, vehicular network and many more. Traditionally, Machine Learning (ML) approaches require data to be centralised in cloud-based data-centres. However, this data is often large in quantity and privacy-sensitive which prevents logging into these data-centres for training the learning models. In turn, this results in critical issues of high latency and communication inefficiency. Recently, in light of new privacy legislations in many countries, the concept of Federated Learning (FL) has been introduced. In FL, mobile users are empowered to learn a global model by aggregating their local models, without sharing the privacy-sensitive data. Usually, these mobile users have slow network connections to the data-centre where the global model is maintained. Moreover, in a complex and large scale network, heterogeneous devices that have various energy constraints are involved. This raises the challenge of communication cost when implementing FL at large scale. To this end, in this research, we begin with the fundamentals of FL, and then, we highlight the recent FL algorithms and evaluate their communication efficiency with detailed comparisons. Furthermore, we propose a set of solutions to alleviate the existing FL problems both from communication perspective and privacy perspective.
△ Less
Submitted 6 April, 2020;
originally announced April 2020.
