Rampart: 14.7MB model for privacy by redacting personal information (ndstudio.gov)

Rampart has been launched as an open-source, on-device personal information filtering system designed to enhance privacy when using chatbots. The model, only 14.7MB in size, efficiently redacts personally identifiable information (PII) by integrating a deterministic layer that identifies structured data like Social Security numbers, with a small language model, MiniLM, that understands contextual nuances in unstructured data. This approach ensures user data remains on their device, alleviating concerns about the vulnerabilities and privacy risks associated with sending sensitive information to remote servers. The significance of Rampart lies in its ability to provide a robust first line of defense in PII removal while being accessible to a broader user base, as many existing models are excessively large and impractical for everyday users. Supporting seven languages, Rampart operates in real time within a browser, making it a pivotal tool in safeguarding privacy in AI chat applications. As an alpha product, it aims to initiate a comprehensive approach to managing personal data in AI interactions, emphasizing user agency and security in increasingly connected digital environments.