Local Image and Text Desensitization Tool

Problem

Ops, data cleaning, and agent workflows often contain sensitive values such as phone numbers, national ID numbers, bank card numbers, and email addresses in logs, chat text, JSON payloads, and screenshots. Sharing or storing them without review can leak personal data, while manual redaction is slow and easily misses structured fields or text embedded in images.

How It Works

This skill runs local desensitization for text, JSON, and images. For text, it recognizes common types including phone numbers, ID numbers, bank cards, and emails, covering 9 listed categories. For JSON, it identifies sensitive fields and processes them. For images, it uses OCR to find sensitive text and blurs the corresponding regions. It also selects a processing mode based on image size, which helps balance small and large inputs.

The skill can initialize desensitization rules for a new agent. On first import, it creates ./基础设定/DESENSITIZE_RULES.md for global rules and updates ./MEMORY.md with desensitization settings. After image desensitization, the original image is deleted by default, while requests explicitly asking not to desensitize are skipped.

Boundaries

The core flow needs Python 3.6+ and standard-library modules such as re, json, base64, and os. Image OCR and face detection are optional extensions that may require easyocr or opencv-python-headless. Text and JSON desensitization can run more lightweight, but image desensitization depends on the OCR environment. Because original images are removed, use it only where irreversible cleanup is acceptable, and add custom rules for business-specific fields rather than assuming it recognizes all internal sensitive data.