This thesis investigates key challenges in enhancing the security, privacy, and compliance of Internet of Things (IoT) systems, focusing on vulnerabilities and proposing solutions for privacy-preserving operations across various scenarios. As IoT systems grow with countless interconnected devices generating vast streams of data, they become increasingly prone to anomalies and malicious attacks. To address these challenges, this thesis is composed of three parts. In Part A, the focus is on ensuring that IoT devices adhere to Security and Privacy Service Level Agreements (SLAs) during service acquisition, while also verifying the reliability of devices through advanced anomaly detection. A Deep Reinforcement Learning (DRL)-based solution is developed to teach IoT devices to autonomously comply with SLAs, ensuring privacy and security in a more autonomous manner. AI-driven anomaly detection models and blockchain technology are employed to verify device reliability, allowing IoT systems to maintain their "liveness" by detecting and responding to suspicious behavior. This ensures that devices providing services are trustworthy and functioning correctly. A privacy-preserving Federated Learning (FL) approach is also proposed for collaborative anomaly detection across IoT devices, leveraging blockchain to coordinate the learning process without centralized data sharing. By integrating SLA compliance and anomaly detection, the framework enables IoT devices to autonomously secure services and verify the reliability of connected devices, maintaining both security and operational efficiency. Moreover, in Part B, given the central role of FL in the proposed solution for Anomaly Detection, this thesis addresses key unresolved security challenges in the field, specifically focusing on model poisoning, backdoor attacks, and inference attacks. Model poisoning involves malicious participants corrupting the global model, while backdoor attacks introduce hidden behaviors activated under specific conditions. Inference attacks, instead, aim to extract private information from models, thus compromising data privacy. This part of the thesis systematically analyzes these threats across Horizontal FL, Vertical FL, and Federated Transfer Learning approaches, identifying vulnerabilities and proposing countermeasures to mitigate them, with the goal of improving security and privacy in decentralized learning environments. In its final remarks, this thesis examines, in Part C the significance of Cyber Threat Intelligence (CTI) in cybersecurity, particularly in light of the immense data generated through human interactions with IoT devices. Given that valuable insights regarding emerging threats may be concealed within this data, advanced Natural Language Processing (NLP) techniques are employed to analyze textual information derived from interactions with natural language model-based assistants. This section underscores the critical role of CTI in distributed systems and emphasizes the need for future research in this area. Overall, this research contributes significant advancements in IoT security, Federated Learning, and gives some insights on the importance of CTI, proposing novel, privacy-preserving solutions to contemporary challenges. It aims to make IoT systems more autonomous and intelligent while maintaining robust security and privacy mechanisms throughout the anomaly detection and service acquisition processes.
Questa tesi indaga le sfide chiave nel miglioramento della sicurezza, della privacy e della conformità dei sistemi dell'Internet of Things (IoT), concentrandosi sulle vulnerabilità e proponendo soluzioni per operazioni che preservano la privacy in diversi scenari. Con la crescita dei sistemi IoT, con innumerevoli dispositivi interconnessi che generano vasti flussi di dati, aumentano anche le probabilità di anomalie e attacchi malevoli. Per affrontare queste sfide, la tesi è composta da tre parti. Nella Part A, l’attenzione è focalizzata sull'assicurare che i dispositivi IoT rispettino gli Accordi sui Livelli di Servizio per la Sicurezza e la Privacy (SLA) durante l'acquisizione di servizi, verificando al contempo l'affidabilità dei dispositivi attraverso l'individuazione avanzata delle anomalie. Viene sviluppata una soluzione basata sul Deep Reinforcement Learning (DRL) per insegnare ai dispositivi IoT a conformarsi autonomamente agli SLA, garantendo privacy e sicurezza in modo più autonomo. Modelli di rilevamento delle anomalie guidati dall'AI e la tecnologia blockchain sono impiegati per verificare l'affidabilità dei dispositivi, consentendo ai sistemi IoT di mantenere la loro "liveness" rilevando e rispondendo a comportamenti sospetti. Ciò garantisce che i dispositivi che forniscono servizi siano affidabili e funzionino correttamente. Viene inoltre proposta un'approccio di Federated Learning (FL) che preserva la privacy per il rilevamento collaborativo delle anomalie nei dispositivi IoT, sfruttando la blockchain per coordinare il processo di apprendimento senza condivisione centralizzata dei dati. Integrando la conformità agli SLA e il rilevamento delle anomalie, il framework consente ai dispositivi IoT di proteggere autonomamente i servizi e verificare l'affidabilità dei dispositivi collegati, mantenendo sia la sicurezza che l'efficienza operativa. Inoltre, nella Part B, dato il ruolo centrale del FL nella soluzione proposta per il rilevamento delle anomalie, questa tesi affronta le principali sfide di sicurezza non risolte in questo campo, concentrandosi in particolare su avvelenamento del modello, attacchi backdoor e attacchi di inferenza. L’avvelenamento del modello coinvolge partecipanti malevoli che corrompono il modello globale, mentre gli attacchi backdoor introducono comportamenti nascosti attivati in condizioni specifiche. Gli attacchi di inferenza, invece, mirano a estrarre informazioni private dai modelli, compromettendo così la privacy dei dati. Questa parte della tesi analizza sistematicamente queste minacce in vari approcci di FL, tra cui Federated Learning Orizzontale, Verticale e Federated Transfer Learning, identificando vulnerabilità e proponendo contromisure per mitigarle, con l'obiettivo di migliorare la sicurezza e la privacy negli ambienti di apprendimento decentralizzati. Nelle considerazioni finali, la tesi esamina, nella Part C, l'importanza della Cyber Threat Intelligence (CTI) nella cybersicurezza, soprattutto alla luce dell'enorme quantità di dati generati dalle interazioni umane con i dispositivi IoT. Poiché informazioni preziose riguardanti minacce emergenti potrebbero essere nascoste in questi dati, vengono utilizzate tecniche avanzate di Natural Language Processing (NLP) per analizzare informazioni testuali derivate dalle interazioni con assistenti basati su modelli di linguaggio naturale. Questa sezione sottolinea il ruolo critico del CTI nei sistemi distribuiti e la necessità di ulteriori ricerche in questo ambito. Nel complesso, questa ricerca contribuisce con significativi avanzamenti nella sicurezza dell’IoT, nel Federated Learning, e fornisce alcune riflessioni sull'importanza della CTI, proponendo soluzioni orientate alla privacy per le sfide contemporanee. L’obiettivo è rendere i sistemi IoT più autonomi e intelligenti, mantenendo solidi meccanismi di sicurezza e privacy durante i processi di rilevamento delle anomalie e di acquisizione dei servizi.
Sicurezza basata sull'AI per l'Internet of Things e Federated Learning
Arazzi, Marco
2025
Abstract
This thesis investigates key challenges in enhancing the security, privacy, and compliance of Internet of Things (IoT) systems, focusing on vulnerabilities and proposing solutions for privacy-preserving operations across various scenarios. As IoT systems grow with countless interconnected devices generating vast streams of data, they become increasingly prone to anomalies and malicious attacks. To address these challenges, this thesis is composed of three parts. In Part A, the focus is on ensuring that IoT devices adhere to Security and Privacy Service Level Agreements (SLAs) during service acquisition, while also verifying the reliability of devices through advanced anomaly detection. A Deep Reinforcement Learning (DRL)-based solution is developed to teach IoT devices to autonomously comply with SLAs, ensuring privacy and security in a more autonomous manner. AI-driven anomaly detection models and blockchain technology are employed to verify device reliability, allowing IoT systems to maintain their "liveness" by detecting and responding to suspicious behavior. This ensures that devices providing services are trustworthy and functioning correctly. A privacy-preserving Federated Learning (FL) approach is also proposed for collaborative anomaly detection across IoT devices, leveraging blockchain to coordinate the learning process without centralized data sharing. By integrating SLA compliance and anomaly detection, the framework enables IoT devices to autonomously secure services and verify the reliability of connected devices, maintaining both security and operational efficiency. Moreover, in Part B, given the central role of FL in the proposed solution for Anomaly Detection, this thesis addresses key unresolved security challenges in the field, specifically focusing on model poisoning, backdoor attacks, and inference attacks. Model poisoning involves malicious participants corrupting the global model, while backdoor attacks introduce hidden behaviors activated under specific conditions. Inference attacks, instead, aim to extract private information from models, thus compromising data privacy. This part of the thesis systematically analyzes these threats across Horizontal FL, Vertical FL, and Federated Transfer Learning approaches, identifying vulnerabilities and proposing countermeasures to mitigate them, with the goal of improving security and privacy in decentralized learning environments. In its final remarks, this thesis examines, in Part C the significance of Cyber Threat Intelligence (CTI) in cybersecurity, particularly in light of the immense data generated through human interactions with IoT devices. Given that valuable insights regarding emerging threats may be concealed within this data, advanced Natural Language Processing (NLP) techniques are employed to analyze textual information derived from interactions with natural language model-based assistants. This section underscores the critical role of CTI in distributed systems and emphasizes the need for future research in this area. Overall, this research contributes significant advancements in IoT security, Federated Learning, and gives some insights on the importance of CTI, proposing novel, privacy-preserving solutions to contemporary challenges. It aims to make IoT systems more autonomous and intelligent while maintaining robust security and privacy mechanisms throughout the anomaly detection and service acquisition processes.File | Dimensione | Formato | |
---|---|---|---|
Marco_Arazzi_PhD_Manuscript_final.pdf
accesso aperto
Dimensione
16.35 MB
Formato
Adobe PDF
|
16.35 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/193062
URN:NBN:IT:UNIPV-193062