Machine learning methods that protect privacy are now highly sought after in the big data and ubiquitous connection era. Federated Learning (FL) is a unique and revolutionary method that solves this problem by allowing machine learning models to be trained on several dispersed clients or devices without requiring the transfer of raw data between the devices and servers. Since then, it has emerged as a prospective paradigm for training various machine learning models while preserving data privacy. This paper offers a thorough analysis of different Federated Learning aggregation procedures and how they affect the accuracy and loss of the model. Within the Federated Learning framework, this paper explores and contrasts four different aggregating methods: Federated Averaging (FedAvg), Federated Proximal (FedProx), Federated Median (FedMedian) and q-FedAvg. We assess these strategies using the Modified National Institute of Standards and Technology (MNIST) dataset, a benchmark dataset for handwritten digit classification, using a Convolutional Neural Network (CNN) based model. Our test results show notable differences in model performance between various aggregation techniques. The paper unveils distinct convergence behaviors and stability characteristics of each algorithm under various data distribution scenarios. Notably, FedProx emerges as a promising choice, demonstrating consistent performance across diverse data distributions. The paper’s findings underscore the importance of careful algorithm selection based on specific application requirements and data characteristics.