ENHANCING IMAGE CLASSIFICATION WITH FEDERATED LEARNING: A COMPARATIVE STUDY OF VGG16 AND MOBILENET ON CIFAR-10

Abstract

In this project, I explored the application of federated learning (FL) algorithms in enhancing image classification tasks using the CIFAR-10 dataset, with a focus on the VGG16 and MobileNet architectures. The project compared the efficacy of various FL algorithms against non-federated baseline models using the same architectures. The non-federated VGG16 and MobileNet models served as baselines to evaluate the relative performance enhancements brought about by federated learning. Remarkably, all explored federated learning algorithms, with the exception of Federated Averaging (FedAvg), demonstrated superior accuracy over the baselines. Although FedAvg did not surpass the baseline models in terms of accuracy, it significantly enhanced the security aspect of model training, thereby reinforcing the trade-off between model performance and data privacy inherent in federated learning setups. This detailed comparison not only underscores the potential of federated learning in practical applications but also highlights the specific strengths and limitations of each algorithm within a federated framework, presenting a comprehensive view of their impacts in a controlled experimental setup.