DocAuth
- metadata analysis
- Fourier analysis
- Error Level Analysis
- template matching
- vector database
Problem
Photoshop edits, splices and copy-move forgeries in scanned documents are hard to catch by eye, and manual verification does not scale when a team has to check authenticity across many documents.
What I built
A document-analysis system that combines four signals — metadata analysis, Fourier-domain inspection, Error Level Analysis and template matching — to decide whether a document has been altered and to highlight where. Reference templates are encoded as vectored feature-maps and stored in a vector database, so matching a new document against a large template set is a nearest-neighbour lookup rather than a pixel-by-pixel scan.
Key decision or trade-off
Matching on vectored feature-maps in a vector database, instead of direct pixel-level template comparison. It keeps lookup fast as the template set grows, at the cost of an upfront step to embed every template.
Result
Detects and highlights manipulations with 97% accuracy. The project won first place at Smart India Hackathon 2023.