Watermark Correlation Lab: Matched-Filter Detection

Provenance labels for AI-generated media only matter if they survive the trip through the real world — re-uploads, re-compression, screenshots and crops. This 2D lab embeds a real spread-spectrum watermark into the DCT coefficients of a synthetic image, then lets you attack the result with JPEG re-compression, additive noise and border cropping — the same math as the 3D bar-grid version, but read here as native 2D plots: the received image itself, a DCT magnitude heatmap, a matched-filter scatter of extracted-vs-key coefficient values (the statistic the detector actually computes), and a live curve sweeping JPEG quality end-to-end to reveal the whole robustness trade-off at once.