Sigurnosni izazovi velikih jezičnih modela

master's thesis

• Title (english): Security challenges in large language models
• Author: Marko Lihter
• Mentor: Marin Vuković (mentor)
• Committee member: Marin Vuković (predsjednik povjerenstva)
• Committee member: Miljenko Mikuc (član povjerenstva)
• Committee member: Dragan Jevtić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-10, Croatia

Abstract

Diplomski rad se bavi sigurnosnim izazovima velikih jezičnih modela. Rad istražuje rizike poput netočnih informacija, pristranosti, manipulacije ulazima i izlazima te curenja podataka. Autor analizira slučajeve kao što su manipulacija podacima, napadi socijalnog inženjeringa, generiranje zlonamjernog softvera i širenje dezinformacija. Također se razmatraju protumjere za ublažavanje tih rizika, uključujući tehničke zahtjeve za sigurnost i certifikaciju. Cilj rada je pružiti sveobuhvatan pregled sigurnosnih problema i rješenja kako bi se osigurala sigurna i etična primjena LLM-ova u različitim područjima.

Abstract (english)

Thesis deals with the security challenges of large language models (LLMs). The paper explores risks such as incorrect information, bias, manipulation of inputs and outputs, and data leakage. The author analyzes cases such as data manipulation, social engineering attacks, the generation of malicious software and the spread of disinformation. Countermeasures to mitigate these risks are also considered, including technical requirements for security and certification. The aim of the paper is to provide a comprehensive overview of security issues and solutions to ensure the safe and ethical application of LLMs in various fields.

• Keywords: kibernetička sigurnost; umjetna inteligencija; veliki jezični modeli; halucinacije; ubrizgavanje upita; curenje podataka; socijalni inženjering; etika; sigurnost podataka; pristranost
• Keywords (english): cybersecurity; artificial intelligence; AI; large language models; LLM; hallucinations; prompt injection; data leakage; social engineering; ethics; data security; bias
• Language: croatian
• URN:NBN: urn:nbn:hr:168:820244
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 21:19:23