Design, implementation and analysis of a cloud federated learning architecture

This is a B-mode project carried out under an employment contract with the company Capgemini. In the recent years, it has been observed that clients are increasingly drawn to use cloud resources due to their scalability, cost-effectiveness and accessibility. Most of them have migrated their environm...

Descripción completa

Detalles Bibliográficos
Autor: Pita Argemi, Albert
Tipo de recurso: tesis de maestría
Fecha de publicación:2023
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/401815
Acceso en línea:https://hdl.handle.net/2117/401815
Access Level:acceso abierto
Palabra clave:Machine learning
Cloud computing
núvol
aprenentatge automàtic
ml
aprenentatge federat
fl
preservació privadesa
models
arquitectura
marc
distribuït
cloud
machine learning
federated learning
privacy-preservation
architecture
framework
distributed
Aprenentatge automàtic
Computació en núvol
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Descripción
Sumario:This is a B-mode project carried out under an employment contract with the company Capgemini. In the recent years, it has been observed that clients are increasingly drawn to use cloud resources due to their scalability, cost-effectiveness and accessibility. Most of them have migrated their environments there, or are in the process of doing so. Simultaneously, the field of machine learning (ML) has expanded rapidly, becoming a standard for many customer processes and applications. Although having huge advantages, it is a method that requires a large amount of data to work, which is not always available in all areas or from the customers themselves, and also one that can arise some privacy concerns, as traditionally the data is centralized in a certain location. In order to circumvent this drawbacks, a privacy-preserved method to train ML models on decentralized data that can be used is federated learning (FL). After a thorough analysis of the situation of the company's current customers and the possible new services that could be offered, it was considered that it would be very beneficial to have the knowledge of how to develop a system where FL can be applied in a productive environment in the cloud. Therefore, the objective of this project is to design and implement a federated architecture in the cloud where models can be trained in a federated way, determining its feasibility, and additionally, also create a package that will be the basis from which coworkers can learn about this field and when facing a similar project, take into account all the options to adapt as much as possible to the customer's situation. For the first, a multi-cloud federated system based on VMs and the framework NVFlare has been deployed. Multiple experiments have been conducted in it, such as validating that federated models achieve similar results as central ones or how a federated model achieves better results than models trained with the isolate data. Results suggesting a similar or even superior performance than with the centralized design. For the later, in the thesis a solid FL background is provided, with two additionally studies on, first, all the possible cloud services from the three most popular cloud providers that could be used to deploy a federated architecture and how they may interconnect, and second, all the existing federated frameworks, evaluating them on its possible productive use.