Analisis Kerentanan Dan Mitigasi Large Language Model Terhadap Common Attack Vectors

Sukmana, Adhitya Raufarhan (2026) Analisis Kerentanan Dan Mitigasi Large Language Model Terhadap Common Attack Vectors. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024221067-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (14MB) | Request a copy

Abstract

Large Language Model (LLM) merupakan teknologi kecerdasan buatan yang banyak digunakan dalam sistem percakapan, analisis teks, dan aplikasi pendukung pengambilan keputusan. Meskipun memiliki kemampuan pemrosesan bahasa alami yang tinggi, LLM tetap rentan terhadap berbagai vektor serangan yang dapat memengaruhi integritas, keandalan, dan keamanan sistem. Penelitian ini berfokus pada analisis tiga vektor serangan umum terhadap LLM, yaitu prompt injection, data poisoning, dan adversarial attack. Prompt injection memungkinkan penyerang memanipulasi instruksi model melalui masukan teks, sedangkan data poisoning dapat mengubah perilaku model melalui penyisipan data berbahaya pada tahap pelatihan atau pembaruan data. Sementara itu, adversarial attack memanfaatkan perubahan kecil pada input untuk menghasilkan keluaran yang menyimpang dari perilaku yang diharapkan. Penelitian ini bertujuan untuk mengidentifikasi karakteristik, dampak, dan potensi risiko dari masing-masing vektor serangan terhadap performa serta keamanan LLM. Metode penelitian meliputi studi literatur, penyusunan skenario serangan, dan pengujian eksperimental pada lingkungan lokal untuk mengamati perubahan keluaran model akibat serangan yang dilakukan. Selain itu, penelitian ini mengkaji strategi mitigasi yang relevan, seperti validasi dan sanitasi input, pengendalian kualitas data pelatihan, penguatan prompt, serta mekanisme deteksi anomali terhadap masukan adversarial. Hasil penelitian ini diharapkan dapat menjadi referensi dalam pengembangan dan penerapan sistem berbasis LLM yang lebih aman serta berkontribusi pada peningkatan keamanan sistem kecerdasan buatan secara menyeluruh.
=======================================================================================================================================
Large Language Models (LLMs) are artificial intelligence technologies widely used in conversational systems, text analysis, and decision-support applications. Despite their advanced natural language processing capabilities, LLMs remain vulnerable to various attack vectors that may affect system integrity, reliability, and security. This research focuses on the analysis of three common attack vectors against LLMs: prompt injection, data poisoning, and adversarial attack. Prompt injection enables attackers to manipulate model instructions through crafted textual inputs, while data poisoning can alter model behavior by inserting malicious data during training or data updates. Meanwhile, adversarial attack exploits subtle modifications to input data to produce outputs that deviate from the expected behavior. This study aims to identify the characteristics, impacts, and potential risks of each attack vector on the performance and security of LLMs. The research methods include a literature review, attack scenario design, and experimental testing in a local environment to observe changes in model outputs caused by the applied attacks. Furthermore, this research examines relevant mitigation strategies, such as input validation and sanitization, training data quality control, prompt hardening, and anomaly detection mechanisms for adversarial inputs. The results of this study are expected to serve as a reference for the development and deployment of more secure LLM-based systems and to contribute to the overall improvement of artificial intelligence security.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Large Language Model, Prompt Injection, Data Poisoning, Adversarial Attack, Keamanan AI, AI Security.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Adhitya Raufarhan Sukmana
Date Deposited:	30 Jul 2026 03:12
Last Modified:	30 Jul 2026 03:12
URI:	http://repository.its.ac.id/id/eprint/137296

Actions (login required)

View Item