Personal Data Detection with Large Language Models

Amazon Web Services has introduced a new system for redacting and detecting personal data that can operate across different large language models and be configured with instructions.

Personal Data Detection and Risks

Training data contains sensitive information such as names, home addresses, email addresses, phone numbers, ID numbers, and bank accounts. Models trained on such data can memorize this information and leak it in undesired situations.

The Role of Large Language Models

While traditional tools are locked into specific entity types, the new approach leverages the instruction-processing capability of large language models to offer a more flexible structure. This allows new entity types to be added without the need for retraining.

Flexible Infrastructure and Operating Principle

The developed system keeps the language independent of the model, providing a flexible back-end interface. It can work seamlessly with models managed through Amazon Bedrock or open-source models hosted on your own infrastructure.