Unique Security and Privacy Threats of Large Language Model: A Comprehensive Survey

With the rapid development of artificial intelligence, large language models (LLMs) have made remarkable progress in natural language processing. These models are trained on large amounts of data to demonstrate powerful language understanding and generation capabilities for various applications, from machine translation and chatbots to agents. However, LLMs have exposed a variety of privacy and security issues during their life cycle, which have become the focus of academic and industrial attention. Moreover, these risks LLMs face are pretty different from previous traditional language models. Since current surveys lack a clear taxonomy of unique threat models based on diverse scenarios, we highlight unique privacy and security issues based on five scenarios: pre-training, fine-tuning, RAG system, deploying, and LLM-based agent. Concerning the characteristics of each risk, this survey provides potential threats and countermeasures. The research on attack and defense situations LLMs face can provide feasible research directions, making more areas reap LLMs' benefits.