A Comparative Study of Federated and Centralised Learning for Waste Classification with Non-IID Data

Abstract

Accurate waste classification is essential for sustainable environmental management, as traditional manual approaches are time-consuming, labour-intensive, and prone to human error. Deep Learning (DL) has achieved remarkable progress in image- based classification, but its dependence on large, labelled datasets and centralised training raises concerns about data privacy and scalability. Federated Learning (FL) provides a privacy-preserving alternative by enabling model training across decentralised devices without sharing raw data. However, applying FL to waste image classification remains challenging due to the non-independent and identically distributed (non-IID) nature of client data, caused by variations in environment, lighting, and user habits.To address this, we propose a privacy-preserving and adaptive FL framework tailored for waste image classification under heterogeneous data distributions. Five Convolutional Neural Network (CNN) architectures—ResNet-18, ResNet-50, GoogLeNet, DenseNet-121, and VGG- 19—were systematically compared under both centralised DL and FL settings. Experimental results show that GoogLeNet achieves the highest accuracy, reaching 80.45% under non-IID FL conditions, outperforming centralised DL by up to 9.8% in specific configurations. These findings demonstrate the effectiveness of FL in improving generalisation and robustness while preserving privacy, providing practical insights for developing scalable, intelligent waste management systems in real-world, diverse environments.