Artificial Intelligence has become a transformative force in modern technology, impacting various sectors such as healthcare, finance, and manufacturing, driven by data analysis, pattern recognition, and prediction. The availability of large volumes of highquality data is crucial for the effectiveness of artificial intelligence. However, data collection faces challenges like sharing restrictions, variability, and privacy concerns due to centralized storage. Federated Learning addresses these issues by enabling collaborative artificial intelligence model training without exchanging raw data, keeping data local while improving model generalizability and security. This thesis analyzes federated learning systems applied to image and time series classification tasks, with a specific focus on learning performance. It discusses advancements and open problems in federated learning, while showcasing potential solutions that have been successfully applied. The main contributions of this dissertation are built on a subset of the most problematic aspects of current federated learning approaches: 1) the reliance on a central node as an aggregator, which represents a single point of failure and raises trustworthiness issues, 2) the aggregation mechanism strictly dependent on model parameters, 3) the requirement for clients to train a shared model of fixed dimension and with the same architecture, 4) the need to aggregate only weights associated with the same type of feature or task. This research explores innovative solutions for scenarios in which the typical federated learning conditions, such as the presence of a central server, uniform model architectures across clients, or the availability of labeled data, do not apply. Each research problem is addressed from both a methodological and practical perspective, and five open-source applications are made freely available and discussed to support the research results: FROCKS, DROCKS, FedER, MERGE, and FedRec. All experiments in the case studies have been tested on various computational facilities, including clusters, cloud environments, and high-performance computing infrastructures. FROCKS is a method for federated time series binary classification that exchanges the best-performing features, while DROCKS extends this to the multiclass scenario, adopting a decentralized communication schema. FedER introduces a decentralized federated learning approach using continual learning and generative AI, eliminating the need for a central server and enabling clients to overcome model-sharing limitations. 3 MERGE enhances model performance by handling diverse data types like images and tabular records. FedRec leverages unlabeled data to improve supervised model training. Each tool has been validated on relevant datasets, demonstrating their adaptability and effectiveness.

Advancing Federated Learning. Towards Decentralization and Personalized Models

CASELLA, BRUNO
2025

Abstract

Artificial Intelligence has become a transformative force in modern technology, impacting various sectors such as healthcare, finance, and manufacturing, driven by data analysis, pattern recognition, and prediction. The availability of large volumes of highquality data is crucial for the effectiveness of artificial intelligence. However, data collection faces challenges like sharing restrictions, variability, and privacy concerns due to centralized storage. Federated Learning addresses these issues by enabling collaborative artificial intelligence model training without exchanging raw data, keeping data local while improving model generalizability and security. This thesis analyzes federated learning systems applied to image and time series classification tasks, with a specific focus on learning performance. It discusses advancements and open problems in federated learning, while showcasing potential solutions that have been successfully applied. The main contributions of this dissertation are built on a subset of the most problematic aspects of current federated learning approaches: 1) the reliance on a central node as an aggregator, which represents a single point of failure and raises trustworthiness issues, 2) the aggregation mechanism strictly dependent on model parameters, 3) the requirement for clients to train a shared model of fixed dimension and with the same architecture, 4) the need to aggregate only weights associated with the same type of feature or task. This research explores innovative solutions for scenarios in which the typical federated learning conditions, such as the presence of a central server, uniform model architectures across clients, or the availability of labeled data, do not apply. Each research problem is addressed from both a methodological and practical perspective, and five open-source applications are made freely available and discussed to support the research results: FROCKS, DROCKS, FedER, MERGE, and FedRec. All experiments in the case studies have been tested on various computational facilities, including clusters, cloud environments, and high-performance computing infrastructures. FROCKS is a method for federated time series binary classification that exchanges the best-performing features, while DROCKS extends this to the multiclass scenario, adopting a decentralized communication schema. FedER introduces a decentralized federated learning approach using continual learning and generative AI, eliminating the need for a central server and enabling clients to overcome model-sharing limitations. 3 MERGE enhances model performance by handling diverse data types like images and tabular records. FedRec leverages unlabeled data to improve supervised model training. Each tool has been validated on relevant datasets, demonstrating their adaptability and effectiveness.
26-giu-2025
Inglese
ALDINUCCI, Marco
Università degli Studi di Torino
File in questo prodotto:
File Dimensione Formato  
thesis.pdf

accesso aperto

Dimensione 12.73 MB
Formato Adobe PDF
12.73 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/215142
Il codice NBN di questa tesi è URN:NBN:IT:UNITO-215142