Enhancing cybersecurity awareness and automation through domain specific LLM

Abstract

The increasing complexity of modern networks and the persistence of human error in cybersecurity underscore the need for intelligent and adaptive defence mechanisms. This thesis investigates the application of large language models in generating secure network configurations to mitigate cyber threats, focusing on reducing human-induced vulnerabilities. By integrating the Mistral 7B model this work fine-tunes a domain-specific LLM using a curated cybersecurity dataset comprising over 15,000 structured entries. Additionally, a hybrid approach combining fine-tuning and Retrieval Augmented Generation is implemented to enhance contextual response accuracy.

The thesis presents a pipeline from data collection and preprocessing to model deployment and performance evaluation. Key implementation techniques include Lora-based fine-tuning, performance tracking using Weights & Biases and chatbot deployment with quantised inference. Results indicate the model effectiveness in generating secure and context-aware network structures. A survey conducted with domain experts further validates the practical usability of the system and its potential in minimising human error. The findings suggest that LLMs when trained and deployed effectively, can serve as powerful tools in automating cybersecurity tasks and improving network resilience.