PDF redaction algorithm disclosure
Rules and the terms you provide identify candidate content. You review and choose what to remove, then generate a new PDF.
Updated: 2026-10-07
Algorithm and service provider
- Algorithm name
- 云屿文档敏感信息识别与内容过滤算法
- Type
- Retrieval and filtering → Content filtering
- Feature
- Yunyu Chengye PDF redaction; other tools’ layout, compression and verification algorithms are outside this disclosure.
- Rule version
- YYT-PRIVACY-RULES/2
- Provider
- 慈溪云屿来文化科技有限公司
- Unified social credit code
- 91330282MAKE1WHC5J
- Responsible person
- 赖海波
The feature filters document content through deterministic rules and user confirmation. It does not use model training, generative models, user profiles or personalized recommendations. This disclosure does not assert filing approval, security certification or regulatory endorsement.
Built-in rules and custom terms
Automatic detection applies only to extractable PDF text. Names, addresses and internal terms can be added as custom terms.
| Category | Method | Limitations |
|---|---|---|
| 18-character resident ID numbers | Number structure, province prefix, calendar date, sequence and weighted check digit; nearby ID labels may produce review candidates. | No identity database lookup or verification of a real person; other document types are outside this rule. |
| Mainland mobile numbers | An 11-digit format, optional country prefix, common separators and adjacent letter or digit boundaries. | No check of activation, validity or ownership; international formats are not comprehensively covered. |
| Email addresses | ASCII address structure, local part, domain labels and length boundaries. | No delivery verification; complex, obfuscated or some internationalized addresses may be missed. |
| Possible bank card numbers | 13–19 digits, nearby card or account labels, order-context exclusions and a Luhn check. | No account existence or ownership verification; long numbers without account context may not match. |
| Custom private terms | User-entered terms are deduplicated and matched literally, with limited normalization of fullwidth characters, spaces and hyphens. | Case sensitive, with no inferred synonyms, abbreviations or matches joined across pages. |
Meeting rule conditions indicates a format or checksum match, not that a number, document or account is real or valid. Review candidates require your judgment. Matches follow document pages and source positions, not commercial relationships, clicks or personal preferences.
From detection to export
- Choose or drop one PDF. The browser checks its format, size, page count and local resource limits.
- A local worker extracts readable text and its positions, then applies the built-in rules and custom terms.
- Review the original matches and their pages, select or deselect items, and add manual boxes for images, scans or missed content.
- Compare with the original page, review each rectangle and confirm. Selected text and supported content in marked regions are removed from a separate copy.
- Remove old metadata, annotations, attachments, signatures and unused resources. Check page count, dimensions and appearance outside masks; stop export if reliable processing is not possible.
- Add a new YunyuTrace identity to the processed file, register digests and required proof, then start the download and show signing results.
Automatic masks locate selected characters; a manual box touching text removes the whole affected character. Visual rectangles may include padding, so review content you want to keep. Solid and rainbow masks are opaque. OCR and mosaic modes are not provided.
Local data and registration data
Local browser processing
The source PDF, original filename, extracted text, original matches, custom terms, previews, mask coordinates and edit history are used in the current browser task. They are not sent to the server for detection, stored as cloud search history or pooled into a cross-user term library. Cancelling, changing the file or leaving the page ends the session; there is no automatic 15-minute idle expiry.
Signing and proof archives
Export registration sends digests of the processed file, its content fingerprint, size, page count, verification code, identity and signing-node metadata. The fingerprint is based on redacted output. Archives use a fixed product filename, not the source filename. Full page proof includes order, dimensions, random salts and page commitments. It is encrypted on the server with required signing proof, without page images or original body text.
The production application and registration support services are deployed in Tencent Cloud Shanghai Zone 2. Verification evidence is archived through the Tencent Cloud COS Shanghai regional endpoint. Website and registration access still involve necessary network information such as IP addresses. See the privacy policy for processing, retention and privacy requests.
Read the full privacy policyLimitations and your review
- Up to 50 MB, 200 pages and two million extracted characters. Custom terms have no fixed count or per-term length limit; device memory, page dimensions and complex structures can impose practical limits.
- Scans, image text and handwriting are not automatically recognized. Pages without text prompt box selection. Obfuscated numbers, cross-page content and semantic confidential information may be missed.
- Complex fonts, transparency, visible annotations, forms or content that cannot be reliably rebuilt can cause export refusal. An overlay-only file retaining source content is not substituted for a redacted result.
- Export removes old attachments, annotations, metadata and signatures. New YunyuTrace registration does not restore the source document’s original signature relationship.
- A local 72-DPI appearance check detects visible changes outside masks; it does not prove that all private information is gone. Zero matches, successful export or a valid signature is not a security certification.
Review every page, match and manual region before sharing. The feature does not comprehensively moderate all illegal or harmful content. Report suspected misuse, privacy violations or product issues through the complaints channel.
Your choices and feedback
You can deselect individual or all matches, edit custom terms, add or remove manual masks, undo changes, compare the original page or cancel processing. There is no personalized recommendation or search switch. You keep the original file and confirm the scope before export.
Contact the responsible person, Lai Haibo, for incorrect results, privacy concerns or requests to access, correct or delete related information held by the service. Do not send the original PDF, complete sensitive numbers, verification codes or access credentials. Start with minimal information such as the date and page or region reference.