Davide Di Marco
From Attack to Evidence: An Application-Forensic Approach to Membership Inference and Secret Extraction in LLMs.
Rel. Andrea Atzeni. Politecnico di Torino, Corso di laurea magistrale in Cybersecurity, 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) | Preview |
Abstract
The emergence and growth of artificial intelligence are among the most significant phenomena of our time. In particular, the adoption of Large Language Models (LLMs) has spread in a few years from research laboratories to business processes, economic services, and everyday life, becoming one of the fastest technological transitions of the decade. This pervasiveness opens new risk surfaces for cybersecurity, with particularly acute implications for confidentiality: a model trained on large corpora can memorize and, under certain conditions, return fragments of the data on which it was trained or fine-tuned, exposing information assumed to remain confined to the training phase. This thesis addresses two closely connected classes of attack: membership inference (MIA) and secret extraction.
Its main object of study is the traces an attack leaves at the observable level: request signatures, non-semantic characteristics of responses, and temporal and sequential patterns
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
