Prompt Engineering-Based Network Intrusion Detection System

Recent advancements in generative AI and evolution of Large Language Models have opened up new paths to explore them for various applications including network attack detection. LLMs’ capability of learning the distribution of training data and subsequently generating new content based on this distribution on the basis of prompts supplied as input has paved the way for predicting anomalies in network features represented as text. In this paper, we propose an LLM prompt-engineering based network intrusion detection method for detection of attacks from raw network packet features represented in text form. The proposed method uses multiple extracted views based on features derived from raw network packets and uses various prompt formats for LLM inferencing. We perform various experiments using packet-level and flow-level information on a recent dataset i.e. CICIDS2018 for FTP Bruteforce attack, and the results show that our proposed prompt engineering based method performs better than current state-of-the-art techniques.