Design, Implementation, and Analysis of a Federated Learning Framework for Music Genre Classification

In the contemporary era of pervasive digital media, the volume of audio data generated by consumer devices has reached unprecedented levels, serving as the foundation for Music Information Retrieval (MIR) systems. However, traditional centralized architectures introduce heavy bandwidth costs, high latency, and severe user privacy vulnerabilities by requiring raw audio data transmission. This research report presents the design, implementation, and analysis of a decentralized Federated Learning (FL) framework for Music Genre Classification. Utilizing the Flower framework and TensorFlow, Deep Neural Networks (MLP architecture) are trained locally on edge devices using the GTZAN dataset. By keeping raw audio data on the local device and sharing only abstract mathematical parameter weights with a central server, the pipeline inherently enforces data minimization and preserves privacy. The report comprehensively evaluates the end-to-end pipeline, including audio feature extraction (MFCCs, Spectral Centroid, Zero-Crossing Rate via Librosa), the mathematical behavior of the Federated Averaging (FedAvg) algorithm under Non-IID data environments, and real-time telemetry orchestration via a Next.js frontend web dashboard. Finally, commercial and strategic use cases are explored alongside a future roadmap incorporating FedProx and Differential Privacy.