OpenAI隐私过滤器发布三款Web应用支持128k上下文PII识别
OpenAI基于1.5B参数隐私过滤器模型推出三款Web应用,支持文档、图像和文本的敏感信息自动识别与脱敏。模型采用Apache 2.0许可,在PII-Masking-300k基准测试中达到最优性能,可识别8类个人身份信息(PII)。
三款应用均基于gradio.Server构建,实现前后端解耦与资源调度。Document Privacy Explorer支持PDF/DOCX全文档PII高亮显示;Image Anonymizer提供图像敏感信息遮盖与画布编辑功能;SmartRedact Paste允许生成公开脱敏链接与私有还原链接。模型单次处理128k token上下文,无需分块即可精准定位PII边界。
该方案推动企业级隐私保护工具向低代码、高可用方向发展,为敏感数据共享提供标准化技术路径。
模型支持128k上下文单次处理
三款应用基于gradio.Server统一架构
PII识别涵盖8类敏感信息
Title:
OpenAI Launches Privacy Filter 1.5B Model 50M Active Parameters 128k Context PII Detection
Summary:
OpenAI has introduced Privacy Filter, a 1.5-billion-parameter AI model with 50 million active parameters, designed for scalable privacy-preserving web applications. The model, licensed under Apache 2.0, supports 128,000-token context and detects nine PII categories including private_person, private_email, and account_number. It achieves state-of-the-art performance on the PII-Masking-300k benchmark, with full evaluation details published in the official release blog.
The model powers three Gradio-based applications: Document Privacy Explorer highlights PII in uploaded PDFs or DOCX files; Image Anonymizer redacts sensitive visual data with editable overlays; and SmartRedact Paste enables secure sharing of redacted text via public URLs with private reveal links. All apps use gradio.Server for backend consistency, leveraging its queueing, ZeroGPU allocation, and gradio_client SDK integration.
This release reflects a growing trend toward modular, privacy-first AI tools built on open frameworks. By combining high-performance PII detection with flexible deployment via Gradio, OpenAI enables developers to create compliant, user-centric applications without sacrificing functionality or scalability.
Key Takeaways:
Privacy Filter processes 128k-token documents in single forward pass without chunking
Three Gradio apps demonstrate real-world PII redaction across text, image, and shared content
Model uses 50M active parameters for efficient inference with Apache 2.0 licensing
Source: Original Article
查看原文 →
View Original →