Federated Learning (FL) is an emerging machine learning technique enabling collaborative model training across decentralised devices without central data aggregation. This study applies FL to image classification using the diverse Pedestrian Attribute Recognition at a Distance (PETA) dataset, aiming to create an effective FL framework for training image classification models while focusing on data privacy and communication efficiency. The motivations behind the adoption of FL in image classification, emphasising the importance of privacy preservation. The proposed FL pipeline involves decentralised edge devices, a central server orchestrating training, and a collaborative learning protocol. Each edge device processes a local subset of the PETA dataset and communicates only model updates to the central server, reducing raw data exchange. The FL approach is evaluated on the PETA dataset, comparing its performance with various pre-trained machine learning models. Different FL strategies, that are approaches used to combine the weights of different models, were explored, analysing their impact on model convergence and performance. The study confirms FedProx as the most effective Federated Learning (FL) strategy for accuracy and precision. It also highlights that FL-trained models achieve better performance in terms of accuracy and precision than centrally trained models. Moreover, potential limitations of FL in image classification is analysed, including the demand for substantial computational resources, particularly GPU RAM. Finally, this paper showcases FL’s applicability and presents the advantages in image classification tasks, emphasising collaborative model training on pedestrian image data while preserving data privacy and without requiring the exchange of confidential information between entities.