This thesis examines the security challenges inherent in blockchain-based decentralized federated learning systems, with a particular focus on the threats posed by malicious trainers and aggregators. For this purpose, a custom-built simulator is employed to evaluate various system setups against diverse attack settings performed by adversarial trainers, including label flipping, targeted data poisoning, and additive noise attacks. The study examines how effective various consensus algorithms, validation methods, and aggregation methods are in countering these attacks. Furthermore, the thesis addresses the trade-offs among system security, computational complexity, and global model quality during training. It examines how various system parameters - e.g., the number of trainers, the number of validators, and the number of training rounds - affect the overall robustness of the system. The simulator provides a way of investigating a large variety of scenarios, thus allowing close examination of system performance under varying conditions. Furthermore, the research investigates how different data distributions among trainers affect the success of various attack strategies and defense mechanisms. Moreover, the thesis identifies vulnerabilities introduced by malicious aggregators and proposes a novel defensive mechanism using Centered Kernel Alignment (CKA) similarity scores. The proposed CKA-based defense shows promise in detecting manipulated global models provided by malicious aggregators. In addition to this, we showed how trust scores, slashing mechanisms, and CKA-based evaluations dynamically reduced the likelihood of malicious aggregators being elected over time, thus mitigating their influence and enhancing the overall resilience of the system. The research methodology includes a thorough literature review, extensive empirical evaluations to characterize malicious trainer behaviors, a theoretical analysis of the likelihood of electing malicious aggregators, and empirical assessments to evaluate the effectiveness of the novel defensive mechanism. Overall, this work makes a significant contribution to understanding and enhancing the security of decentralized federated learning systems, providing valuable insights for developing more robust and reliable solutions. The thesis also lays the groundwork for future research directions, including the development of more sophisticated defense mechanisms and the exploration of security issues in emerging decentralized federated learning paradigms.
Questa tesi esamina le sfide di sicurezza insite nei sistemi di apprendimento federato decentralizzato basati su tecnologia blockchain, con un focus particolare sulle minacce poste da formatori e aggregatori malevoli. A tal fine, viene impiegato un simulatore appositamente sviluppato per valutare diverse configurazioni di sistema in presenza di varie modalità di attacco condotte da trainer avversari, tra cui l'inversione delle etichette, l’avvelenamento mirato dei dati e gli attacchi mediante aggiunta di rumore. Lo studio esamina l’efficacia di vari algoritmi di consenso, metodi di validazione e tecniche di aggregazione nel contrastare tali attacchi. Inoltre, la tesi affronta i compromessi tra la sicurezza del sistema, la complessità computazionale e la qualità del modello globale durante l’addestramento. Viene analizzato come diversi parametri di sistema - ad esempio, il numero di trainer, il numero di validatori e il numero di round di addestramento - influenzino la robustezza complessiva del sistema. Il simulatore offre un metodo per investigare un’ampia varietà di scenari, consentendo così un’analisi approfondita delle prestazioni del sistema in condizioni variabili. Inoltre, la ricerca esamina come differenti distribuzioni dei dati tra i trainer incidano sul successo delle varie strategie di attacco e dei meccanismi di difesa. Inoltre, la tesi identifica le vulnerabilità introdotte dagli aggregatori malevoli e propone un nuovo meccanismo difensivo basato sui punteggi di similarità Centered Kernel Alignment (CKA). La difesa proposta basata su CKA mostra potenzialità nel rilevare modelli globali manipolati forniti da aggregatori malevoli. In aggiunta, abbiamo dimostrato come i punteggi di fiducia, i meccanismi di penalizzazione e le valutazioni basate su CKA abbiano ridotto dinamicamente la probabilità che aggregatori malevoli venissero eletti nel tempo, mitigando così la loro influenza e migliorando la resilienza complessiva del sistema. La metodologia di ricerca include una revisione approfondita della letteratura, valutazioni empiriche estensive per caratterizzare i comportamenti dei formatori malevoli, un'analisi teorica della probabilità di eleggere aggregatori malevoli e un'analisi empirica per valutare l'efficacia del nuovo meccanismo difensivo. Nel complesso, questo lavoro contribuisce in modo significativo alla comprensione e al miglioramento della sicurezza dei sistemi di apprendimento federato decentralizzato, fornendo preziose intuizioni per sviluppare soluzioni più robuste e affidabili. La tesi pone inoltre le basi per future direzioni di ricerca, inclusi lo sviluppo di meccanismi di difesa più sofisticati e l'esplorazione delle problematiche di sicurezza nei paradigmi emergenti di apprendimento federato decentralizzato.
Enhancing security in blockchain-based decentralized federated learning: a study of malicious trainers and aggregators
Caroli, Federico;Bleggi, Francesco
2023/2024
Abstract
This thesis examines the security challenges inherent in blockchain-based decentralized federated learning systems, with a particular focus on the threats posed by malicious trainers and aggregators. For this purpose, a custom-built simulator is employed to evaluate various system setups against diverse attack settings performed by adversarial trainers, including label flipping, targeted data poisoning, and additive noise attacks. The study examines how effective various consensus algorithms, validation methods, and aggregation methods are in countering these attacks. Furthermore, the thesis addresses the trade-offs among system security, computational complexity, and global model quality during training. It examines how various system parameters - e.g., the number of trainers, the number of validators, and the number of training rounds - affect the overall robustness of the system. The simulator provides a way of investigating a large variety of scenarios, thus allowing close examination of system performance under varying conditions. Furthermore, the research investigates how different data distributions among trainers affect the success of various attack strategies and defense mechanisms. Moreover, the thesis identifies vulnerabilities introduced by malicious aggregators and proposes a novel defensive mechanism using Centered Kernel Alignment (CKA) similarity scores. The proposed CKA-based defense shows promise in detecting manipulated global models provided by malicious aggregators. In addition to this, we showed how trust scores, slashing mechanisms, and CKA-based evaluations dynamically reduced the likelihood of malicious aggregators being elected over time, thus mitigating their influence and enhancing the overall resilience of the system. The research methodology includes a thorough literature review, extensive empirical evaluations to characterize malicious trainer behaviors, a theoretical analysis of the likelihood of electing malicious aggregators, and empirical assessments to evaluate the effectiveness of the novel defensive mechanism. Overall, this work makes a significant contribution to understanding and enhancing the security of decentralized federated learning systems, providing valuable insights for developing more robust and reliable solutions. The thesis also lays the groundwork for future research directions, including the development of more sophisticated defense mechanisms and the exploration of security issues in emerging decentralized federated learning paradigms.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Bleggi_Caroli_tesi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Testo tesi
Dimensione
2.98 MB
Formato
Adobe PDF
|
2.98 MB | Adobe PDF | Visualizza/Apri |
2025_04_Bleggi_Caroli_executive summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Testo executive summary
Dimensione
503.42 kB
Formato
Adobe PDF
|
503.42 kB | 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/10589/234154