Federated Learning Based Enhanced FedBA with MobileNet Convolutional Neural Network for the Identification of Columnar Cactus

Abstract

With the advancement in artificial intelligence (AI) technology and Internet of Things (IoT) growing popularity, the use of unmanned aerial vehicles (UAVs) as IoT devices is also being studied. However, privacy issues and limited communication resources restrict the use of unmanned aerial vehicles. Federated Learning (FL) has emerged as a promising approach for training machine learning models on decentralized devices while maintaining data privacy and reducing the communication costs. To aggregate the model on the server-side FL server uses various aggregation algorithms. One such promising aggregation algorithm is FedBA for non-identically and independently distributed (non-IID) data via UAVs. However, despite its advantages, FedBA may face challenges related to convergence speed, robustness to malicious clients, and overall efficiency. To solve these challenges, this study presents a FL model aggregation technique in which clients and servers communicate parameters as opposed to data, thereby enhancing privacy, and reducing communication costs UAVs images. This research proposes enhancements to the FedBA algorithm aimed at addressing these challenges and improving its performance. The method proposes federated learning based enhanced FedBA with MobileNet Convolutional Neural Network (CNN) for the identification of columnar cactus in the Tehuacán-Cuicatlán Valley of Mexico. Using a public dataset of over 20,000 remote sensing images, the suggested model is evaluated and found to be superior to InceptionV3 and modified MobileNet CNN. The key contributions of this work include the introduction of momentum for faster convergence, adaptive learning rates for better optimization and model aggregation clipping to prevent extreme updates. The proposed FL framework mitigates the issue of slow convergence and communication cost for non-IID data from UAVs. Enhanced FedBA more effective than typical FL algorithms. The classification accuracy before aggregation is 95% and improved to 97% after aggregation.