Comparison of Federated Learning Strategies Using the Flower Framework on Embedded Devices

Abstract

Federated Learning (FL) is a decentralized machine learning approach that provides a compelling solution to the constraints of centralized model training by allowing collaborative learning across distributed devices while protecting data privacy. This thesis performs a thorough analysis of FL strategies, concentrating on their efficacy and performance when deployed on resource-constrained embedded devices like Raspberry Pis. The study introduces the Flower framework as a versatile interface for FL implementation, particularly suited for deployment on embedded systems like Raspberry Pi. Through systematic experimentation and empirical analysis, various FL strategies are evaluated, including Federated Averaging (FedAvg), Federated Proximal (FedProx), and others, leveraging the Flower framework. The experimental setup involves training an image classification model on a cluster of Raspberry Pis, simulating real-world scenarios with diverse data distributions, client populations, and privacy constraints. Performance metrics such as model accuracy, convergence speed, and computational and communication overhead are meticulously analyzed to discern the strengths and weaknesses of each FL strategy. The experimental results illustrate the delicate balance between preserving privacy and maintaining model performance across various Federated Learning (FL) strategies. Among these, FedAvg demonstrates robustness and efficiency in typical FL scenarios, while FedProx exhibits competitive performance, especially in scenarios with non-identically distributed (non-IID) data. FedOpt demonstrates promising performance across all conditions, while FedMedian and FedYogi, although slightly slower in convergence speed and accuracy, demonstrate resilience to non-IID data and stragglers. Notably, FedAvg, FedMedian, and FedProx maintain relatively higher accuracies even under stringent differential privacy constraints. These findings emphasize the importance of carefully selecting FL techniques that are tailored to the specific objectives and restrictions of the application environment.