Ensuring confidentiality and personal data protection in large language models
Διασφάλιση της εμπιστευτικότητας και προστασίας προσωπικών δεδομένων στα μεγάλα γλωσσικά μοντέλα
View/
Abstract
This thesis deals with how Large Language Models can protect confidentiality and personal data. It starts from the observation that existing solutions, meaning legislation and infrastructure security measures, are not enough since the model itself can reveal sensitive information through conversation. Protection therefore has to become part of the model’s own behavior. After covering the legal framework, the risks and existing techniques, th thesis tests two training approaches: first, training a model in the role of the defender, acting as a customer service representative, with reinforcement learning through a PPO trainer. Then, a new framework proposed in this work, the LLM-GAN, where an attacker model and the defender are trained against each other. The main conclusion is that PPO training on its own has weaknesses, whereas with the LLM-GAN, where the defender constantly faces an opponent that adapts in order to extract information from it and tries out many strategies, the model receives more meaningful training and is more secure across a wider range of attacks and situations.
Postgraduate Studies Programme
Τεχνητή Νοημοσύνη - Artificial IntelligenceDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνCorporate Department
National Center of Scientific Research "Demokritos"Number of pages
135Language
EnglishCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα