ReferNet · reference-assisted document forgery detection

Pick a design and a document. The paper's deployment model runs in your browser, with or without the design reference.

1Design reference

Design reference

Median of genuine documents · hover to see the personal-data zone

2Presented document

Presented document

Ground truth hidden until the run

3ReferNet

Shared encoder, early stagessame weights for both inputs Document+ data zone Deeper stagesjudge alone score Design referencemedian of genuine Comparevs the reference correction + Refer / Clear

4Decision

Run ReferNet to see its decision.

How it works

A design reference is the pixel-wise median of genuine documents of one design, registered to each other. Personal data differ from document to document and fade in the median, while the printed design stays sharp. The personal-data zone marks where a document is allowed to differ from its design.

ReferNet judges the presented document on its own, through a single-image path, and, with the reference, compares it with the design reference through a second path that shares the same encoder. The comparison produces a correction that can raise or lower the single-image decision. Without the reference, ReferNet is a single-image classifier like the published baselines.

In this demo, each document has already been registered to its design reference with the paper's code, and its compression equalized as in the paper. The network itself runs live in your browser. Inference times are measured on your device, whereas the paper's 280 ms is end to end on a laptop processor.

About the difference map. The map in step 4 is not part of the paper. It is shown for demonstration only. It is computed in your browser from the same run, from the comparison features weighted by the correction layer, to indicate which region drove the reference's correction, and it has not been evaluated as a localization method.