privacy-redact

One-click privacy shield for screenshots & documents

Drop an image or PDF → sensitive info is auto-blurred → click Save.

Cards, emails, phones, addresses, faces, barcodes — gone. Hidden metadata — gone. Nothing ever leaves your device.

Why

We paste screenshots into ChatGPT, Claude, Slack, Discord, GitHub issues and social media all day. Every one of them can quietly contain a home address, a card number, a face that isn't yours, or a QR code that logs someone in. And the "fix" — drawing black boxes by hand in Paint or Photoshop — is tedious, easy to get wrong, and usually leaves the EXIF/GPS/camera-serial metadata untouched.

privacy-redact makes the safe thing the easy thing:

1. Drop any image or PDF (or paste a screenshot with Ctrl/⌘+V).
2. Auto-blur — sensitive items are found and hidden for you. Click any box to un-hide it, drag to hide something we missed.
3. Save (or Copy to clipboard) — a clean file with zero metadata, verified before you share it.

This repository is the library: detection, redaction, and document handling, plus browser components for OCR, faces, barcodes, and PDFs. The page in this repo is a harness for those components, not a brand. A small host that shows each call is in examples/.

Screenshots

| Landing | Workspace | | --- | --- | | | |

Before & after gallery

Three redaction styles, seven document types — all produced by the real app. See the full gallery →

| Document | Blur · Pixelate · Blackout | | --- | --- | | 🧾 Pay stub (photo, QR, barcode) | | | 💬 Dark-mode Slack chat (keys, password) | | | 🪪 Driver's license | | | 🏥 Medical intake form | | | 💳 Checkout page | | | 👨‍👩‍👧 Group photo (faces) | | | 📄 Lease PDF | |

Features

🔍 Detects

| Category | Examples | How | | --- | --- | --- | | 💳 Credit cards | 4242 4242 4242 4242, Amex 3782 822463 10005, masked **** 4242 | Regex + Luhn checksum (tolerates OCR slips on well-formed 4-4-4-4 groups) | | 🪪 SSN / national ID | 123-45-6789, SSN: 123456789, TIN, EIN | Pattern + invalid-range filtering (000, 666, 9xx) | | ✉️ Emails | [email protected] | RFC-ish regex | | 📞 Phone numbers | (415) 555-2671, +1 415 555 2671, +44 20 7946 0958 | US + international patterns | | 🏠 Home addresses | 742 Evergreen Terrace, Apt 4B, Springfield, IL 62704, PO Box 12, Canadian & UK postcodes | Street/suffix, City-State-ZIP, PO box rules | | 🏦 Account numbers | Account No: 00123456789, Routing # 021000021, MRN, Policy, IBANs | Label-aware: redacts the value, not the label | | 🆔 Aadhaar numbers | 2341 2341 2346, Aadhaar No: 2341-2341-2346 | 12 digits, first digit 2-9, Verhoeff checksum when unlabelled | | 🧾 PAN (India tax ID) | ABCPE1234F, PAN: ABCPE1234F | AAAAA9999A shape with a valid holder-type letter; labelled values tolerate OCR look-alike swaps | | 🎂 Dates of birth | DOB: 04/12/1988, Date of Birth March 3, 1990 | Only when labelled – ordinary dates are left alone | | 🔑 API keys & tokens | sk-…, AKIA…, ghp_…, xoxb-…, JWTs, password: … | Vendor prefixes + key=value | | 🙂 Faces | Photos, ID pictures, group shots | SSD-MobileNet via TensorFlow.js (tiled scan for small faces) | | ▦ Barcodes & QR | QR, Code 128/39, EAN/UPC, PDF417, Data Matrix, Aztec | Native BarcodeDetector if present, else ZXing with multi-scale retry |

🧼 Strips hidden data

Before you save, privacy-redact shows exactly what's hiding in your file — and what's about to disappear:

• 📍 GPS coordinates & altitude
• 📷 Camera / phone model, 🔢 body & lens serial numbers
• 🕒 Original capture & edit dates
• 👤 Owner / author / copyright, 🛠️ editing software
• 📄 PDF title, author, producer, keywords, dates
• XMP / IPTC blocks, embedded thumbnails, ICC profiles

After export it re-opens the file it just wrote and checks that no metadata is left ("re-checked: 0 metadata fields in the output").

🎨 Aesthetic, adjustable redaction

• Blur (default) – soft, pretty, and irreversible (see How redaction stays irreversible)
• Pixelate – classic mosaic
• Blackout – maximum-paranoia solid fill
• Strength slider, per-category switches, click-to-toggle on any box, drag-to-add manual boxes
• Hold to compare with the original · multi-page PDFs with page navigation

🔒 Private by design

• Everything runs in your browser tab — OCR (Tesseract), face detection (TensorFlow.js), barcode decoding (ZXing), PDF rendering (PDF.js).
• Models, WASM and fonts are bundled – no CDN calls at runtime. Works fully offline after first load.
• Production builds ship a strict Content-Security-Policy (connect-src 'self', form-action 'none', object-src 'none') so the page is technically unable to send your file anywhere.
• The end-to-end test fails if the app makes even one request to another origin.
• No analytics, no cookies, no accounts, no telemetry.

Quick start

Run it locally

npm install          # also copies the OCR / face / PDF assets into public/vendor
npm run dev          # → http://localhost:5173

Production build (static files, host anywhere — S3, Netlify, python -m http.server, a USB stick):

npm run build        # → dist/
npm run preview      # serve dist/ at http://localhost:4173

Requires Node 18+ and a modern browser (Chrome/Edge 119+, Firefox 121+, Safari 17.4+).

Desktop app (experimental)

A tiny Electron shell is included. It serves the built app from a loopback-only server and blocks every non-local network request.

npm run desktop        # build + launch
npm run desktop:pack   # installers via electron-builder (dmg / nsis / AppImage) → release/

The desktop wrapper hasn't had the same test coverage as the web app yet — PRs and bug reports welcome.

How it works

flowchart LR
    A[Drop image / PDF] --> B[Decode to canvas<br/>EXIF-orientation applied<br/>PDF pages rasterised]
    B --> C1[OCR<br/>Tesseract.js]
    B --> C2[Faces<br/>SSD-MobileNet]
    B --> C3[Barcodes / QR<br/>ZXing]
    C1 --> D[Pattern engine<br/>regex + Luhn + labels]
    D --> E[Word boxes → pixel rects]
    C2 --> F[Regions]
    C3 --> F
    E --> F
    F --> G[Live preview<br/>blur / pixelate / blackout]
    G --> H[Export from raw pixels<br/>PNG · JPG · flattened PDF]
    H --> I[Re-open & verify<br/>0 metadata fields]

1. Decode. Images become a canvas (transparency flattened, EXIF orientation baked in). PDFs are rendered page-by-page at 144 dpi.
2. Detect. OCR yields words with bounding boxes. Screenshots are upscaled and dark themes inverted first to help the OCR. Each line is scanned with the rule set in src/detect/patterns.js; matches are mapped back to pixels — even partial words like Email:[email protected] are trimmed to just the address.
3. Review. Every detection is a box you can toggle. Categories can be switched off wholesale.
4. Export. The redacted canvas is re-encoded from scratch and re-checked.

💥 How redaction stays irreversible

Bad redaction is worse than none, so:

• Blur and pixelate average pixels into blocks first (only ~1–2.4 cells across a line of text, ~5–11 across a face), add random per-block noise, and only then smooth. The original detail is discarded, not merely obscured — there is nothing to "un-blur". Text cells are deliberately coarser than a glyph because fine mosaics of a known font can be brute-forced back into text (Depix-style attacks); the jitter stops exact colour matching.
• Nothing is layered. The output image is a new set of pixels. There is no hidden original underneath a box, unlike PDF "highlight" annotations that can be deleted.
• PDFs are flattened. Each page is rasterised, redacted and re-embedded as an image, so the text layer beneath a redaction no longer exists. (The trade-off: exported PDFs aren't text-selectable.)
• Nothing is copied byte-for-byte from your source file, so EXIF, XMP, IPTC, thumbnails and PDF info dictionaries can't tag along. PDFs are written with metadata auto-stamping disabled (no Producer, no dates).
• The e2e suite OCRs the exported file and asserts that the SSN, email, phone, card, address and account numbers of the sample document are no longer readable.

Want zero doubt for something extremely sensitive? Pick Blackout and drag the strength up.

Limitations — please read

privacy-redact is a safety net, not a guarantee. Always glance at the preview before sharing.

• Names aren't detected. Names are context-dependent and would cause a lot of false positives with regex-only detection. Use the drag-to-hide tool (a local NER model is on the roadmap).
• OCR isn't perfect. Tiny fonts, low contrast, unusual fonts, handwriting, rotated or curved text can be missed. A digit misread by OCR can still leave a card pattern undetected if it's badly broken up.
• English-first. Address rules are US/Canada/UK oriented; phone rules include international formats; OCR uses the English model.
• Faces: heavily rotated, occluded or tiny faces can be missed; non-faces occasionally trigger. Wide images are scanned in overlapping tiles so small faces are found, at the cost of speed. Toggle boxes to correct.
• Barcodes: without the native BarcodeDetector API (Chrome on Windows/Linux, Firefox), 1D codes are located via scan-line heuristics and a structure check that rejects text mistaken for bars. Damaged or very small codes can be missed.
• Not detected by design: license plates, signatures, logos, company names, free-form medical text.
• HEIC photos aren't decodable by browsers – convert to JPG/PNG first.
• PDFs are limited to 60 pages and become image-only.
• Performance: first scan loads ~15 MB of models. Face detection uses WebGL when available; on machines without GPU acceleration it falls back to CPU and can take a while on big images.
• Filenames aren't scrubbed beyond adding -redacted. If your filename contains your name, rename it.

Project structure

privacy-redact/
├── index.html                # app shell
├── src/
│   ├── main.js               # UI controller (state, scanning pipeline, export)
│   ├── style.css
│   ├── detect/
│   │   ├── patterns.js       # 🧠 the rule engine (pure JS, unit-tested)
│   │   ├── textRegions.js    # OCR words ↔ character ranges ↔ pixel rects, wrapped-line + neighbour merging (pure, unit-tested)
│   │   ├── ocr.js            # Tesseract.js worker wrapper (+ pre-processing)
│   │   ├── faces.js          # TensorFlow.js face detection with tiling + NMS
│   │   └── barcodes.js       # BarcodeDetector / ZXing multi-scale fallback
│   ├── render/effects.js     # blur · pixelate · blackout
│   └── io/
│       ├── loader.js         # image + PDF → canvases
│       ├── metadata.js       # what's hidden in the file + post-export verification
│       └── exporter.js       # PNG / JPG / flattened PDF writers
├── scripts/
│   ├── copy-assets.mjs       # vendors OCR/face/PDF assets into public/vendor (postinstall)
│   ├── make-samples.mjs      # generates the fictional sample docs
│   └── make-before-after.mjs # README hero image
├── public/samples/           # fictional pay stub, chat, ID, medical form, checkout, group photo, GPS JPG, lease PDF
├── docs/EXAMPLES.md          # before/after gallery (docs/examples/*)
├── test/                    # node --test unit tests (*.test.js)
│   └── smoke.e2e.test.js     # Playwright end-to-end test incl. OCR-of-output check (npm run test:e2e)
├── desktop/                  # optional Electron wrapper
└── .github/workflows/        # CI + GitHub Pages deploy

Testing

npm test                                   # unit tests (rules, box mapping, metadata, redaction cell size) — under a second
npm run test:e2e:setup                     # one-time: download Chromium for Playwright
npm run build && npm run preview &         # then:
npm run test:e2e -- http://localhost:4173/ # full pipeline in headless Chromium
npm run test:e2e:ci                        # all of the above in one go: install Chromium, build, serve, test (used by the manual E2E workflow)

The e2e test drives the real UI against the production build (so the CSP is enforced) and asserts:

• every category is detected on the sample pay stub (email, phones, SSN, addresses, card, account/routing, DOB, face, QR + barcode)
• the exported PNG can no longer be OCR'd for any secret, while normal text remains readable
• a JPEG with GPS + camera model comes out with no EXIF segment
• a 2-page PDF with a real text layer and author metadata comes out with no selectable secret text and an empty info dictionary
• zero requests to external origins and zero console errors / CSP violations

All sample data is fictional (4242 4242 4242 4242 is a public test card number; the portrait is AI-generated). Regenerate with npm run samples.

Roadmap

• [ ] Optional local name / organisation detection (small on-device NER model)
• [ ] Batch mode – drop a folder, get a folder
• [ ] Watch-folder / clipboard-monitor mode for the desktop app
• [ ] More locales (EU addresses & national IDs, IBAN validation, non-Latin OCR packs)
• [ ] License-plate & signature detectors
• [ ] "Keep text selectable" PDF mode using true content-stream redaction
• [ ] Browser extension: right-click any image → privacy-redact
• [ ] Installable PWA (a web manifest ships already; offline service worker still to do)
• [ ] Custom rules (bring your own regex / keywords, e.g. an employee-ID format)

FAQ

Does anything get uploaded? No. There is no backend. Open your browser's Network tab and see for yourself — after the page loads, dropping a file causes zero requests.

Why not just blur it in Preview / Paint? You can, but you have to spot every item yourself, get the boxes right, and remember that the file's EXIF/GPS data is still there. privacy-redact does the spotting and the scrubbing.

Is PNG/JPG export lossless / does it change my image? PNG is lossless. JPG is re-encoded at quality 93. Colour profiles are dropped (part of metadata stripping), which may shift colours very slightly on wide-gamut photos.

Can I add my own patterns? Yes — see CONTRIBUTING.md. It's one object in patterns.js.

Why is the first scan slow? The OCR engine and face model load on first use, then get reused for the rest of the session.

Security

Found a way to recover redacted content or leak data? Please see SECURITY.md — use a private report, and only ever share fictional documents.

Credits

This repository started from Hushshot by SYasJ. That project is the offline browser tool that finds private content in a screenshot or PDF, redacts it, and strips metadata. Its license is MIT. This repository borrowed that first approach and aligned it as a library a host can call, including before a document goes to an AI.

This repository runs them through the JavaScript projects listed with each one. Those projects keep their own licenses.

| Project | What it does here | License | | --- | --- | --- | | Tesseract OCR | Recognizes text. The English traineddata vendored with the page is Tesseract language data. | Apache-2.0 | | Tesseract.js and tesseract.js-core | Runs that engine in the browser and in Node, from the vendored worker and WASM. | Apache-2.0 | | TensorFlow and TensorFlow.js | Runs the face model. TensorFlow.js is bundled inside face-api and is not loaded from the network. | Apache-2.0 | | face-api.js, continued as @vladmandic/face-api | Finds faces with the SSD MobileNet v1 model shipped in that package. | MIT | | ZXing and @zxing/library | Decodes barcodes and QR codes when the browser has no BarcodeDetector. | Apache-2.0, MIT for the JavaScript port | | PDF.js | Reads and rasterizes PDFs. | Apache-2.0 | | pdf-lib | Writes the flattened PDF. | MIT | | exifr | Reads image metadata so the export can be checked. | MIT |

The page is built with Vite (MIT). End-to-end tests use Playwright (Apache-2.0). Crimson Pro and IBM Plex Sans are used under the SIL Open Font License.

License

All rights reserved. See LICENSE. See CONTRIBUTING.md to get involved.