An Adversarial Machine Learning Approach on Securing Large Language Model with Vigil, an Open-Source Initiative

Kushal Pokhrel; Cesar Sanín; Md Rafiqul Islam; Edward Szczerbicki

doi:10.1016/j.procs.2024.09.486

An Adversarial Machine Learning Approach on Securing Large Language Model with Vigil, an Open-Source Initiative

Abstract

Several security concerns and efforts to breach system security and prompt safety concerns have been brought to light as a result of the expanding use of LLMs. These vulnerabilities are evident and LLM models have been showing many signs of hallucination, repetitive content generation, and biases, which makes them vulnerable to malicious prompts that raise substantial concerns in regard to the dependability and efficiency of such models. It is vital to have a complete grasp of the complex behaviours of malicious attackers in order to build effective strategies for protecting modern artificial intelligence (AI) systems through the development of effective tactics. The purpose of this study is to look into some of these aspects and propose a method for preventing devastating possibilities and protecting LLMs from potential threats that attackers may pose. Vigil is an open-source LLM prompt security scanner, that is accessible as a Python library and REST API, specifically to solve these problems by employing a sophisticated adversarial machine-learning algorithm. The entire objective of this study is to make use of Vigil as a security scanner. and asses its efficiency. In this case study, we shed some light on Vigil, which effectively recognises and helps LLM prompts by identifying two varieties of threats: malicious and benign.

Citations

0

CrossRef
0

Web of Science
0

Scopus

Authors (4)

Kushal Pokhrel
- IHG Sydney, Australia
Cesar Sanín
- IHG Sydney, Australia
Md Rafiqul Islam
- IHG Sydney, Australia
Edward Szczerbicki prof. dr hab. inż.
- Faculty of Management and Economics
- Gdańsk University of Technology

Cite as

Full text

download paper

downloaded 8 times

Publication version: Accepted or Published Version
DOI:: Digital Object Identifier (open in new tab) 10.1016/j.procs.2024.09.486
License: open in new tab

Keywords

ADVERSARIAL MACHINE LEARNING, LARGE LANGUAGE MODEL, INFORMATION SECURITY, PROMPT INJECTION

Details

Category:

Conference activity

Type:

publikacja w wydawnictwie zbiorowym recenzowanym (także w materiałach konferencyjnych)

Published in:

Procedia Computer Science no. 246, pages 686 - 695,
ISSN: 1877-0509

Language:

English

Publication year:

2024

Bibliographic description:

Pokhrel K., Sanín C., Islam M. R., Szczerbicki E.: An Adversarial Machine Learning Approach on Securing Large Language Model with Vigil, an Open-Source Initiative// / : , 2024,

DOI: