Federated learning study

• Main Author: Aratrika, Pal
• Other Authors: Jun Zhao
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166751
• Institution: Nanyang Technological University

Federated Learning is rapidly gaining traction as a machine learning technique in today’s world due to the prevalence of isolated islands of private data belonging to different organizations. It eliminates the need to collect data into a central location by taking the model to the client devices instead. With the metaverse’s growing popularity, it is essential that mobile augmented reality (MAR) devices of the metaverse can perform image classification. Federated Learning can be used here for collective training as each individual device will only have limited data. Thus it is important to benchmark the performance of federated learning systems on classification tasks for their use in the MAR devices. In our study, we explore Horizontal Federated Learning on classification tasks, primarily image classification, using neural networks and a centralized communication architecture. We implemented different federated learning strategies such as FedAvg, FedAvgM, FedProx, FedAdam, FedYogi and FedAdagrad either from scratch or using federated learning frameworks like Secretflow or Flower, to perform federated classification on the MNIST, FashionMNIST, Cifar10 and Criteo datasets. Implementing from scratch helps us develop a better intuition of federated learning concepts and using the frameworks helps us learn how experiments can be performed in a federated learning research setting. We compare the basic and most commonly used algorithm: FedAvg’s performance on image classification with Independent and Identically Distributed (IID) and Non-IID client data distributions and show that FedAvg is better suited for IID data distributions. We compare the other algorithms FedAvgM, FedProx, FedAdam, FedYogi and FedAdagrad with FedAvg on Non-IID client data distributions for image classification and report that all of them beat FedAvg on such data. Lastly, we also evaluate the federated classification system’s performance against varying input image resolutions and show that as resolution decreases, accuracy significantly decreases. We hope that the findings from the experiments conducted in our study will help make smart design decisions while developing and implementing federated classification systems for MAR devices of the metaverse.