Due to the continuous growth in Internet data, cybersecurity practitioners have developed new defenses based on Machine Learning (ML). ML-based solutions offer numerous benefits, from learning patterns among large amounts of data to generalizing to unknown data. This dissertation covers three significant aspects derived from the interaction between machine learning and cybersecurity: (i) definition of novel Network Intrusion Detection Systems (NIDS), (ii) cybersecurity for web content monitoring, and (iii) Adversarial Machine Learning (AML). The first part of the dissertation presents two NIDS themes: XeNIDS, aiming to study and design cross-networking NIDS, and DETONAR, a NIDS for low-powered IoT networks. The second part covers cybersecurity for web content monitoring. In particular, as users interact in forums and Online Social Networks (OSN), their activity might threaten others (e.g., hate speech). The dissertation covers two themes: helpful review prediction, aiming to forecast whether a review from forums (e.g., Amazon, Yelp) will be considered helpful, and PRaNA, a heuristic that leverages videos’ Photo Response Non-Uniformity (PRNU) to spot real videos from their deepfake versions. The third - and last - part of the dissertation presents two evasion attacks: ZeW, an evasion attack on Natural Language Processing applications that leverages invisible UNICODE characters, and CAPA, which discusses real examples of threats created by OSN’s users that undermined Automatic Content Moderators.
A causa della continua crescita dei dati Internet, i professionisti della sicurezza informatica hanno sviluppato nuove difese basate sul Machine Learning (ML). Le soluzioni basate su ML offrono numerosi vantaggi, dalla loro capacità di apprendimento in grandi quantità di dati alla generalizzazione a dati sconosciuti. Questa tesi copre tre aspetti significativi derivati dall'interazione tra machine learning e cybersecurity: (i) definizione di nuovi Network Intrusion Detection Systems (NIDS), (ii) cybersecurity per il monitoraggio dei contenuti web e (iii) Adversarial Machine Learning (AML). La prima parte della tesi presenta due temi in ambito NIDS: XeNIDS, con l'obiettivo di studiare e progettare cross-network NIDS, e DETONAR, un NIDS per reti IoT a bassa potenza. La seconda parte riguarda la sicurezza informatica per il monitoraggio dei contenuti web. In particolare, poiché gli utenti interagiscono nei forum e nei social network online (OSN), la loro attività potrebbe minacciare gli altri (ad esempio, incitamento all'odio). La tesi copre due temi: previsione dell'utilità delle recensioni, con l'obiettivo di prevedere se una recensione scritta in forum (e.g., Amazon, Yelp) sarà considerata utile da futuri utenti, e PRaNA, un'euristica che sfrutta il Photo Response Non-Uniformity (PRNU) dei video per individuare video genuini dalle loro versioni deepfake. La terza - e ultima - parte della dissertazione presenta due attacchi di evasione: ZeW, un attacco di evasione alle applicazioni di elaborazione del linguaggio naturale che sfrutta caratteri UNICODE invisibili, e CAPA, che discute esempi reali delle minacce create dagli utenti di OSN che hanno minato i moderatori automatici dei contenuti.
Data-driven cybersecurity / Pajola, Luca. - (2023 Feb 27).
Data-driven cybersecurity
PAJOLA, LUCA
2023
Abstract
Due to the continuous growth in Internet data, cybersecurity practitioners have developed new defenses based on Machine Learning (ML). ML-based solutions offer numerous benefits, from learning patterns among large amounts of data to generalizing to unknown data. This dissertation covers three significant aspects derived from the interaction between machine learning and cybersecurity: (i) definition of novel Network Intrusion Detection Systems (NIDS), (ii) cybersecurity for web content monitoring, and (iii) Adversarial Machine Learning (AML). The first part of the dissertation presents two NIDS themes: XeNIDS, aiming to study and design cross-networking NIDS, and DETONAR, a NIDS for low-powered IoT networks. The second part covers cybersecurity for web content monitoring. In particular, as users interact in forums and Online Social Networks (OSN), their activity might threaten others (e.g., hate speech). The dissertation covers two themes: helpful review prediction, aiming to forecast whether a review from forums (e.g., Amazon, Yelp) will be considered helpful, and PRaNA, a heuristic that leverages videos’ Photo Response Non-Uniformity (PRNU) to spot real videos from their deepfake versions. The third - and last - part of the dissertation presents two evasion attacks: ZeW, an evasion attack on Natural Language Processing applications that leverages invisible UNICODE characters, and CAPA, which discusses real examples of threats created by OSN’s users that undermined Automatic Content Moderators.File | Dimensione | Formato | |
---|---|---|---|
phd_thesis_luca_pajola.pdf
accesso aperto
Descrizione: Data-driven cybersecurity
Tipologia:
Tesi di dottorato
Dimensione
7.47 MB
Formato
Adobe PDF
|
7.47 MB | Adobe PDF | Visualizza/Apri |
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.