Federated Learning based Gender Classification in Heterogeneous and Distributed Data having Concept Drift

Abstract

Federated Learning is gaining significant traction in recent times, due to its ability to protect privacy while giving comparable results across various use cases. However, it often falls short in delivering good results in scenarios where the client devices at different places store huge and heterogeneous data. The studies explore the effectiveness of federated learning, when utilized for model training on heterogeneous data having concept drift nature and being placed on different client devices. The paper proposes a novel combination of EfficientNet-B0 model fine-tuned with adaptive and weighted normalization layers in a federated learning setup for gender classification in heterogeneous Fairface dataset having human images from different races. Several other Convolutional Neural Network (CNN) models namely, ResNet50, DenseNet121, GoogLeNet, AlexNet etc. have also been trained in different traditional and federated setups. Experiments revealed that EfficientNet-B0 outperformed the other models trained in all the scenarios with an outperforming accuracy of 88.18% in the traditional federated setup and 90.56% when fine-tuned using normalization methods and evaluated in a federated setup with four client devices. These findings highlight the potential of normalization techniques to improve the effectiveness of federated learning, especially when dealing with concept drift data.