A Novel Reputation-aware Client Selection Scheme for Federated Learning within Mobile Environments

This paper studies the problem of training federated deep learning models over a mobile environment. Stemming from the federated learning (FL) concept, deep learning models on mobile devices can be trained for various use cases including but not limited to image sorting and prediction of upcoming words. Mobile devices have access to rich data sets through embedded sensors and as well as installed software, and these feature rich data can facilitate solid training models, including personal images and other behaviometric features. However, utilizing the data through conventional approaches can potentially lead to privacy leakages. In this paper, we propose an alternate strategy that builds on the Federated Learning (FL) concept, to keep the training data on distributed mobile devices, and train a shared model by aggregating updated local models. The contribution of this study is an optimal user selection method for the federated learning environment based on reputation scores. Through extensive validation experiments considering two different model architectures and three datasets, our experiments show that the proposed approach is stable over data that is not independent nor identically distributed (i.e., non-IID) and under imbalanced distribution. Experimental results show that the proposed reputation-aware FL scheme can achieve improvements in the test accuracy up to 9.30% under different data sets.