Sekvencijalne strategije napada na velike jezične modele

undergraduate thesis

• Title (english): Sequential attack strategies on large language models
• Author: Filip Borić
• Mentor: Mihaela Vranić (mentor)
• Committee member: Mihaela Vranić (predsjednik povjerenstva)
• Committee member: Damir Pintar (član povjerenstva)
• Committee member: Ana Šarčević (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-10, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Kako se veliki jezični modeli sve više koriste u različitim industrijskim sektorima i u svakodnevnom životu, njihova sigurnost postaje sve važnije pitanje. Ovo područje istraživanja zahtijeva stalnu pažnju kako bi se osiguralo da AI tehnologije budu pouzdane i korisne za društvo. Razvijena je originalna strategija napada koja koristi strukturirani i sekvencijalni pristup te dokazuje slabosti u velikim jezičnim modelima koje napadači mogu iskoristiti. Primjenom standardiziranih evaluacijskih metoda, dokazana je efikasnost osmišljene strategije napada u otkrivanju i iskorištavanju sigurnosnih slabosti velikih jezičnih modela, uz predložene metode za daljnje unaprjeđenje strategije.

Abstract (english)

As large language models are increasingly used in various industrial sectors and in everyday life, their security becomes an increasingly important issue. This research area requires constant attention to ensure that AI technologies are reliable and useful for society. An original attack strategy is developed that uses a structured and sequential approach and demonstrated weaknesses in large language models that attackers can exploit. By applying standardized evaluation methods, the effectiveness of the invented attack strategy was proven in detecting and exploiting security weaknesses of large language models, with proposed methods for further improvement of the strategy.

• Keywords: veliki jezični modeli; sekvencijalne strategije napada; evaluacija sigurnosnih slabosti; etička pitanja u AI; ChatGPT; LLaMA; Mistral
• Keywords (english): large language models; sequential attack strategies; evaluation of security weaknesses; ethical issues in AI; ChatGPT; LLaMA; Mistral
• Language: croatian
• URN:NBN: urn:nbn:hr:168:209329
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 20:08:09