Document tampering detection: what a vision model can and can't catch

Detecting an altered ID card or edited document is a real, hard vision problem. What signals are checkable today, and what's still beyond a straightforward pipeline.

Written by

Pranjul Rathour

GenAI Engineer at SCULT INDIA, based in Kanpur, India. Builds production RAG systems, fine-tunes open LLMs and ships agentic AI products; has mentored 200+ college developers. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks.

pranjulrathour41@gmail.com LinkedIn GitHub X Instagram Bluesky Dev.to

At an Integral Startup Foundation hackathon

Detecting whether a document — an ID card, a certificate — has been digitally altered is a genuinely hard vision problem, and it's worth being honest about what's checkable today versus what requires more sophisticated forensics than a straightforward pipeline provides.

What's checkable with a reasonable pipeline

Font and layout consistency against a known template for that document type — misaligned text or an inconsistent font is a common tell.
Compression artifacts around a suspected edit region, which basic error-level analysis can sometimes surface.
Metadata inconsistencies (creation date, editing software) when the original file, not just a photo of it, is available.

What's genuinely hard

A well-executed edit on a photographed (not digital) document, using competent editing tools, can defeat most straightforward checks. Be honest with users and stakeholders about this limitation rather than implying a pipeline catches all tampering.

See document AI for invoices: an extraction schema.

— Pranjul Rathour, GenAI Engineer from Kanpur, India. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at any campus: pranjulrathour41@gmail.com.