Pozivni injekcijski napadi na velike jezikovne modele

magistrsko delo

Matic Bobnar (Author), Sašo Karakatič (Mentor)

Abstract

V magistrskem delu raziskujemo vlogo velikih jezikovnih modelov v vzponu generativne umetne inteligence. Predstavimo osnovne koncepte, kot so transformerji, žetoni in vektorske reprezentacije, ter opisujemo ključne prednosti, slabosti in izzive z generativnimi modeli. Posebno pozornost namenjamo izzivom varnosti, kot so pozivni injekcijski napadi. Podrobno analiziramo delovanje teh napadov, njihove vrste in predstavimo možne pristope za obrambo pred njimi. V okviru eksperimenta prikazujemo izdelavo spletne ankete, ki implementira različne jezikovne modele. S pomočjo pridobljenih podatkov iz ankete nato analiziramo občutljivost posameznih modelov na različne intenzitete injekcijskih napadov ter preučujemo njihove vplive na uporabniške dimenzije, kot so uporabnost, točnost, razumljivost in relevantnost. Na koncu ugotavljamo, kateri modeli se najbolje odzivajo na napade in predstavljajo najvarnejšo uporabo.

Keywords

generativna umetna inteligenca;generativni modeli;veliki jezikovni modeli;pozivni injekcijski napadi;inženering pozivov;magistrske naloge;

Data

Language:	Slovenian
Year of publishing:	2025
Typology:	2.09 - Master's Thesis
Organization:	UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher:	[M. Bobnar]
UDC:	004.8.056(043.2)
COBISS:	226895619
Views:	0
Downloads:	12
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Prompt injection attacks on large language models
Secondary abstract:	In the master's thesis, we explore the role of large language models in the rise of generative artificial intelligence. We present fundamental concepts such as transformers, tokens, and vector representations, and describe the key advantages, disadvantages, and challenges of generative models. Special attention is given to security challenges, such as prompt injection attacks. We analyze the functioning of these attacks in detail, their types, and propose possible defense approaches. As part of the experiment, we develop an online survey application that implements various language models. Using the data collected from the survey, we analyze the sensitivity of individual models to different intensities of injection attacks and examine their impacts on user dimensions such as usability, accuracy, understandability, and relevance. Finally, we identify which models respond best to the attacks and represent the most secure usage.
Secondary keywords:	generative artificial intelligence;generative models;large language models;prompt injection attacks;prompt engeneering;
Type (COBISS):	Master's thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Informatika in podatkovne tehnologije
Pages:	1 spletni vir (1 datoteka PDF (XVII, 72 str.))
ID:	25750865