FedUNA: A Federated Learning Approach for Robust and Privacy-Preserving Pothole Classification using EfficientNet

Pothole detection is an essential aspect of road maintenance and safety. This paper introduces a novel method for pothole classification using the EfficientNet architecture within a Federated Learning (FL) framework, implemented through our proposed Federated Unified Network Aggregation (FedUNA) algorithm. EfficientNet-B3 serves as the base model for balancing the accuracy and computational efficiency. FedUNA operates by training the model across distributed datasets without compromising data privacy, utilizing random client selection and weighted aggregation to enhance generalization. Various data augmentation techniques, including rotation, flipping, and distortion, have been employed to improve model robustness. Our federated model achieves a high accuracy of 94%. This approach facilitates collaborative learning across multiple road management authorities while safeguarding data confidentiality, contributing to the growing field of privacy-preserving machine learning in infrastructure maintenance.