Prompt-based Adversarial Text Generation using Large Language Models

Abstract

Large language models have demonstrated significant performance on Natural Language Processing (NLP) tasks, with prompt engineering playing a crucial role in enhancing their abilities. Prompt engineering involves creating natural language instructions called prompts to extract knowledge from LLMs in a structured manner, without requiring extensive parameter retraining or fine-tuning based on the given NLP task. As a rapidly evolving area of research, prompt engineering has garnered significant attention in recent years, leading to the development of various strategies aimed at improving the precision and reliability of information extracted from Large Language Models. Our paper investigates the ability of open source LLMs such as LLaMA-2-7B-Chat, OpenHermes-7B, and Mistral-7B-Instruct to generate smishing messages that evade detection by spam filters. We systematically evaluate ten prompting strategies, including Chain-of-Thought, Persona prompting, and synthetic prompt variants, measuring their success in bypassing a smishing classifier. We compute several linguistic realism metrics including contextual and semantic similarity, urgency, and trustworthiness to assess the plausibility of the generated messages. Our results reveal that certain prompting methods significantly enhance evasion success, highlighting the need for more robust defenses against LLM-generated adversarial messages.

Ähnliche Arbeiten