user-model communication pipline的安全风险分类,重点关注prompt-based attack on LLM。根据目标和攻击类型分类,并使用具体示例对每种类型的影响进行展示。
Introduction

the user, the model, and a third party
Taxonomy
- black-box
- covering a broad spectrum of attacks that exploit prompts or misuse a given trained model
Attack Target
- user. 通过compromise用户和LLM的交互,扰乱用户workflow。
- model. 使模型输出unintended response。
- third party. 利用模型发动对第三方的攻击。
CIA Triad
Shield 提出使用integrity, source identification, attack detectability, and utility preservation几种属性进行防御的方法Shield。
- confidentiality. 只有authorized recipient才能与LLM交互。
- integrity. 与LLM交互的accuracy, validity, and completeness。
- availability.
User

- eavesdropping. 窃听。
- prompt bloating. 修改用户prompt导致模型response过长。