Defending Against Prompt Reverse Attacks: An Information Theory-Based Privacy Protection Framework for LLM Collaborative Inference

Introduction / Main Floor: Defending Against Prompt Reverse Attacks: An Information Theory-Based Privacy Protection Framework for LLM Collaborative Inference

This paper proposes an information theory-based defense framework that minimizes the mutual information between intermediate activations and input prompts. It preserves user privacy while maintaining model inference utility, providing theoretical guarantees and practical solutions for edge-cloud collaborative inference scenarios.