OpenAI: The Privacy Filter Redefines Data Protection

OpenAI Introduces the Privacy Filter for Better PII Management

OpenAI has recently introduced the Privacy Filter, an open-source tool on its Hub designed to detect personally identifiable information (PII) in texts. This filter categorizes data into eight different types in a single pass over a 128k context. This technology enables developers to create web applications that respect user privacy. We explored three applications that demonstrate the varied capabilities of this filter.

Document Privacy Explorer: A Secure Document Reading Experience

The Document Privacy Explorer application, available at ysharma/OPF-Document-PII-Explorer, addresses the need to read documents containing PII, such as contracts or resumes, while highlighting this sensitive information. The Privacy Filter processes an entire document at once, without splitting or stitching, thanks to its 128k context. The BIOES decoding ensures accuracy in marking spans, even in long and complex sequences.

To provide a smooth user experience, the application features a summary dashboard at the top, allowing an overview of detected PII. Users can toggle CSS classes on the client side without restarting the model, enhancing the fluidity of navigation within the document.

Gradio.Server allows serving the reader view as a single HTML file and exposes the model behind a queued endpoint. This enables switching between PII categories without reloading the page, thus offering a continuous and natural reading experience.

Image Anonymizer: Protecting Shared Images

The Image Anonymizer application, available at ysharma/OPF-Image-Anonymizer, is designed for those who wish to share images or screenshots while masking PII. By integrating Tesseract for OCR, the application identifies words in an image and uses the Privacy Filter to detect PII. Users can then add, move, or remove black bars over sensitive information before exporting the modified image.

The application's workflow allows for the activation of all bars in a category at once, and the export to PNG occurs on the client side at the natural resolution, without server round trips. Gradio.Server plays a key role by enabling layered annotation via gr.ImageEditor. This facilitates the management of category metadata and the export of images without requiring additional communication with the server, thanks to a custom <canvas> frontend.

SmartRedact Paste: A Secure Pastebin

SmartRedact Paste, accessible at ysharma/OPF-SmartRedact-Paste, offers a solution for securely sharing sensitive texts. Users can paste lines from logs or emails and receive two URLs: a public one with PII replaced by placeholders like <PRIVATE_PERSON>, <PRIVATE_EMAIL>, <ACCOUNT_NUMBER>, and a private one, protected by a token, showing the original text.

The Privacy Filter replaces PII with placeholders, and gr.Server manages two distinct GET routes for each paste, ensuring the security and confidentiality of shared data. Multilingual text, including Spanish, French, Chinese, Hindi, and other languages, is processed in the same manner without requiring additional modifications. The format of the URL is crucial, as the reveal URL is the one the user must keep to access the uncensored content.

The Contribution of gradio.Server in These Applications

Gradio.Server simplifies the integration of the Privacy Filter into these applications by using FastAPI routes to handle requests. Each application uses @server.[api](/glossaire/api) for queued computation, while simple routes @server.get are reserved for HTML pages and file searches. This architecture allows for efficient request management and a consistent user experience, even with varied interfaces.

In summary, OpenAI's Privacy Filter, combined with gradio.Server, opens new perspectives for personal data protection in web applications. These innovations are essential to meet the growing privacy demands in the digital age.