Modular defense represents a flexible and adaptable approach to protect complex systems such as federated learning environments, where security is crucial to preserve the integrity of global models. The main contribution of this study is the implementation of a server-side security module that integrates two advanced defense solutions, Xplain and FedArmor, exploring their applicability in a real federated context. The goal is to test the effectiveness of these defenses in a concrete scenario, such as the TRUSTroke project, which aims to develop a privacy-preserving AI platform for stroke management, evaluating their impact on performance and robustness against different attacks. Specifically, the proposed security module allows the alternative use of Xplain and FedArmor, allowing a direct comparison between the two solutions in the same experimental context. The results highlight that Xplain provides superior protection, maintaining an accuracy of more than 90% even with 70% of compromised clients. FedArmor instead is faster in computations but shows significant limitations in large-scale scenarios, with insufficient protection of the global model when the percentage of compromised clients is high. A key aspect of this work is the modularity of the defense system, which allows dynamic selection of the two defenses, thus providing greater flexibility in addressing attacks in a federated environment. The comparative analysis of the two defenses offers valuable insights into their performance, vulnerabilities, and potential applicability in real-world federated learning scenarios.
La difesa modulare rappresenta un approccio flessibile e adattabile per proteggere sistemi complessi, come gli ambienti di federated learning, dove la sicurezza è cruciale per preservare l’integrità dei modelli globali. Il principale contributo di questo studio è l’implementazione di un modulo di sicurezza lato server che integra due soluzioni avanzate di difesa, Xplain e FedArmor, esplorandone l’applicabilità in un contesto federato reale. L’obiettivo è testare l’efficacia di queste difese in uno scenario concreto, come il progetto TRUSTroke, che mira a sviluppare una piattaforma di intelligenza artificiale a protezione della privacy per la gestione dell’ictus, valutando il loro impatto sulle prestazioni e sulla robustezza contro diversi tipi di attacco. In particolare, il modulo di sicurezza proposto consente l’uso alternato di Xplain e FedArmor, permettendo un confronto diretto tra le due soluzioni nello stesso contesto sperimentale. I risultati evidenziano che Xplain offre una protezione superiore, mantenendo un’accuratezza superiore al 90% anche con il 70% di client compromessi. D’altra parte, mentre FedArmor è più veloce nei calcoli, mostra limitazioni significative in scenari su larga scala, con una protezione insufficiente del modello globale quando la percentuale di client compromessi è alta. Un aspetto chiave di questo lavoro è la modularità del sistema di difesa, che consente la selezione dinamica e l’integrazione di soluzioni in base al contesto, offrendo così una maggiore flessibilità nell’affrontare gli attacchi in un ambiente federato. L’analisi comparativa delle due difese offre preziose informazioni sulle loro prestazioni, vulnerabilità e potenziale applicabilità in scenari reali di federated learning.
Design and implementation of a server-side security module for real-world federated learning
FAIETTI, BENEDETTA
2024/2025
Abstract
Modular defense represents a flexible and adaptable approach to protect complex systems such as federated learning environments, where security is crucial to preserve the integrity of global models. The main contribution of this study is the implementation of a server-side security module that integrates two advanced defense solutions, Xplain and FedArmor, exploring their applicability in a real federated context. The goal is to test the effectiveness of these defenses in a concrete scenario, such as the TRUSTroke project, which aims to develop a privacy-preserving AI platform for stroke management, evaluating their impact on performance and robustness against different attacks. Specifically, the proposed security module allows the alternative use of Xplain and FedArmor, allowing a direct comparison between the two solutions in the same experimental context. The results highlight that Xplain provides superior protection, maintaining an accuracy of more than 90% even with 70% of compromised clients. FedArmor instead is faster in computations but shows significant limitations in large-scale scenarios, with insufficient protection of the global model when the percentage of compromised clients is high. A key aspect of this work is the modularity of the defense system, which allows dynamic selection of the two defenses, thus providing greater flexibility in addressing attacks in a federated environment. The comparative analysis of the two defenses offers valuable insights into their performance, vulnerabilities, and potential applicability in real-world federated learning scenarios.| File | Dimensione | Formato | |
|---|---|---|---|
|
tesi Benedetta Faietti.pdf
non accessibile
Dimensione
2.27 MB
Formato
Adobe PDF
|
2.27 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/246147