Say Goodbye to Information Leaks! OpenAI Launches Privacy Filter Supporting Ultra-Long Context of 128,000 and 8 Types of Privacy Recognition

 OpenAI has recently announced the launch of a cutting-edge Personal Identifiable Information (PII) anonymization model called Privacy Filter. The model is now available under the Apache 2.0 license on Hugging Face and GitHub, aiming to provide developers with a local-run, highly customizable privacy protection tool.

Deep Semantic Understanding, Saying Goodbye to Mechanical Matching

Differing from traditional rule-matching tools, Privacy Filter has deep language comprehension capabilities. It can accurately identify sensitive information in unstructured text based on context. This means it can effectively obscure individual private data while preserving as much public useful information as possible in the text.

Lightweight MoE Architecture, Outstanding Performance

In terms of technical architecture, the model demonstrates high flexibility and efficiency:

• Mixture of Experts (MoE) Design: Although the total parameter scale reaches 1.5 billion, only about 50 million parameters are activated during each inference. This allows it to run smoothly on edge devices with limited resources, such as laptops or browsers.

• Extended Context Support: It has a 128,000 Token context window. Using a bidirectional token classification architecture and a constrained Viterbi algorithm, it ensures coherence and accuracy in processing long texts.

• High Accuracy Recognition: In the revised version of the PII-Masking-300k benchmark test, the model achieved an F1 score of 97.43%, with a recall rate as high as 98.08%.

A Comprehensive Privacy Classification System

Privacy Filter can accurately identify and label eight types of core sensitive information:

1. Basic Identity: Names, addresses, email addresses, phone numbers.

2. Online Assets: URL links.

3. Financial Security: Account information (including bank cards, credit cards, etc.).

4. Confidential Credentials: Passwords, API keys, etc.

5. Time-Sensitive: Date information.

Application Scenarios: "Local Firewall" for Cloud LLMs

OpenAI positions it as a pre-filter layer. Before sending text to cloud-based large models, data can be processed locally for PII detection and anonymization. This "data stays on device" approach effectively mitigates the risk of users accidentally pasting private information into AI tools.