Type: dissertation
Abstract: The pervasivity of artificial intelligence is structurally changing how societies envision their development.The perception of what is valuable is rapidly shifting, and data is in the eye of the storm.This process affects every level of organisation in the modern world: from single individuals to multinational companies, from public administration to international organisations, everybody is developing the consciousness that data is a new, powerful social and economic device.Data is precious, and the conflict between who tries to harvest it and who instead tries to keep it private and protect its owners' rights rages.This friction constitutes the natural environment for developing new ideas and concepts to mitigate this contention.Federated learning arose in 2016 as one possible solution to the abovementioned struggle.The fundamental concept of this paradigm is to eliminate the need to exchange data between entities in favour of exchanging the knowledge extracted from it.This approach bypasses the privacy limit imposed by data owners, effectively enabling large-scale cooperative machine learning model training and eliminating the need for invasive data harvesting.Starting from the real-world problem of creating a machine learning-based risk score for cardiological pathologies, the unsustainability of creating data lakes for such sensible tasks is described and discussed.Federated learning is identified as a solution to such an issue, and its main positive and negative aspects are investigated and analysed.Three assumptions of current-days federated learning software constitute the starting points for the main contributions of this dissertation: 1) the centralised structure currently implemented by many commercial frameworks, 2) their inner workings being strictly tied to deep learning models, and 3) their assumption of being deployed on private, specialised computing infrastructures.The proposed research expands the federated learning paradigm to handle scenarios in which these three conditions do not hold.Such research problems are addressed methodologically and practically during the dissertation, and three open-source, proof-of-concept software is made freely available as tangible research results: FastFL, OpenFL-x, and xFFL.All the software discussed is experimentally tested on many computational infrastructures (cluster, cloud, and high-performance computing facilities) and different micro-architectures (x86-64, ARM-v8, RISC-V), and the obtained learning and computational performance is collected, presented, and discussed.A PhD is not a mere career choice or a one-more title to add to one's CV.A PhD is a long, complex, and life-modeling journey that shapes one person's mind and world vision into something else: it turns a student into a researcher.Such a transformation requires time, dedication, and hard academic work, but not just that: it requires traveling, seeing different research realities, developing social skills, discussing with foreign people, and getting involved in new and unforeseen situations.Simply put, a PhD means growth, both academic and personal.Such a growth process requires the dedication of precious time and resources by a PhD supervisor.In this context, I express my deepest gratitude to my supervisor, Prof. Marco Aldinucci.Despite his packed and busy academic life, he always did his best to answer my questions, clear my doubts, and guide me through the intricacies of the academic world.He never made me lack any resources to experience my PhD path fully, and he always gave me the right advice in every situation.Thank you sincerely for all of that.However, a PhD is a team effort, not a one-person effort.The team that supported me during these intense years is numerous and varied, and I could not have found better people.Thank you, Iacopo, for all the evenings and nights spent in video call during the pandemic period of my PhD and for all the subsequent beer consumed in person after that; your technical and personal suggestions will not be forgotten.Thank you,
Author: MITTONE, GIANLUCA
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7217265159