Federated Learning for Edge Intelligence

This research investigates the comparative performance of federated and centralized learning models for bird image classification across multiple training rounds. We implement a complete federated learning system using TensorFlow and Flower framework, with a MobileNetV2-based architecture capable of classifying five categories (bluetit, jackdaw, robin, unknown_bird, unknown_object). Our system demonstrates that federated learning achieves 92.3% accuracy compared to 94.1% in centralized mode, with the added benefit of data privacy preservation. The implementation includes a web-based interface for real-time classification, model statistics visualization, and prediction history tracking. Key findings show that while centralized models maintain slightly higher accuracy, federated models exhibit competitive performance (within 2% accuracy difference) with significant privacy advantages, making them suitable for ecological monitoring applications where data sharing is restricted. Keywords:Federated Learning, Centralized Learning, Accuracy Comparison, Machine Learning, Data Privacy, Decentralized Systems.