Protecting User Privacy in Cloud LLMs
A framework for pseudonymizing sensitive information in LLM prompts
This research introduces a general pseudonymization framework that preserves privacy when users interact with cloud-based LLMs like ChatGPT, without requiring model modifications.
- Replaces sensitive information in user prompts with pseudonyms before transmission to LLM providers
- Maintains context and semantic meaning while hiding personal identifiable information
- Automatically restores original information in the model's response
- Works with various cloud LLMs without requiring access to model internals
This approach addresses critical security concerns for organizations using third-party LLMs by creating a protective layer between users and model providers, helping businesses comply with data privacy regulations while still leveraging cloud AI capabilities.