AI Content Watermark Robustness Lab

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 lab embeds a real spread-spectrum watermark into the DCT coefficients of a synthetic image (the same class of technique behind invisible content-authenticity marks), then lets you attack the result with JPEG re-compression, additive noise and border cropping. A correlation detector, computed with the exact matched-filter formula real private watermarking systems use, tracks live whether the mark still authenticates — and a frequency-domain view shows the watermarked coefficients literally sinking into the noise floor as the attack gets harsher.