This report reviews and reproduces the foundational work of McMahan et al. (2017) on Federated Learning, specifically focusing on the Federated Averaging (FedAvg) algorithm. The report highlights how FL enables collaborative model training on decentralized, privacy-sensitive data while avoiding the need to transfer raw data to central servers. We summarize the methodology, key innovations, and empirical results of the original study, and complement it with our own implementation using PyTorch. We replicate experiments on the MNIST dataset using a Multi-Layer Perceptron (MLP), observe convergence behavior, and discuss performance metrics including accuracy and communication efficiency. Limitations and directions for future work, such as privacy guarantees and dynamic client participation, are also discussed.