Enhancing Image Classification with Federated Learning through a Flower Simulation Study on the CIFAR100 Dataset

The aim of this study is to enhance image classification using Federated Learning (FL) applied to the CIFAR100 dataset. We created a decentralized FL environment with multiple clients, leveraging the Flower framework and Ray for distributed computing. The methodology involved federated averaging for parameter updates, client-specific data handling, and performance evaluation using various metrics. Our results demonstrate a significant improvement in model performance, with loss decreasing from 0.060846 to 0.038253 and accuracy increasing from 41.64% to 57.96% over five training rounds. This study concludes that FL is effective in improving image classification while preserving data privacy and decentralization.