bynn-document-fraud

Document fraud detection

One submission runs the full forensic pipeline: forgery template matching, tampering and manipulation detection, font and layout consistency, signature validation, deepfake and AI-generation scoring, MRZ and barcode parsing, EXIF analysis, classification, and an overall risk score.

This is asynchronous. Submitting returns a receipt, not a verdict.

Base: https://api.bynn.com/v1. Private key on both calls.

1. Submit

Multipart, which is what you want for anything of real size:

curl -X POST https://api.bynn.com/v1/documents \
  -H "Authorization: Bearer <YOUR_PRIVATE_KEY>" \
  -F "file=@/path/to/document.pdf" \
  -F "reference_id=order-8123"

Or JSON with strict base64, no line breaks and no data: prefix:

curl -X POST https://api.bynn.com/v1/documents \
  -H "Authorization: Bearer <YOUR_PRIVATE_KEY>" \
  -H "Content-Type: application/json" \
  -d '{ "document_base64_strict": "<...>", "reference_id": "order-8123" }'

Send file or document_base64_strict, not both.

Response:

{ "submission_id": "document_...", "document_id": "...", "status": "received" }

submission_id is what you poll. Store it.

Submit the original file, never a derivative

Do not resize, crop, re-encode, re-compress, or otherwise process the document before submitting it. Send the bytes exactly as you received them.

Tampering detection, AI-generation scoring, and deepfake analysis all work on traces the manipulation left in the file: compression and quantization artifacts, resampling patterns, sensor noise, font and edge inconsistencies, and EXIF. Those traces live in the exact bytes. Resizing resamples them away, re-encoding overwrites them with your own encoder's artifacts, and stripping metadata removes evidence outright.

The failure mode is quiet and dangerous: a processed forgery comes back looking clean.

• Store the original upload and submit that, not a thumbnail or a normalized copy.
• Do not flatten a PDF to an image, and do not re-render it. Send the PDF.
• Do not run the file through an image optimizer, a screenshot, or a messaging app.
• If a file exceeds the size limit, reject or escalate it. Do not shrink it and submit anyway. A downscaled submission can return a false clean verdict.