AI Art Provenance & Protection Workbench v2.4 Live

Client-side steganographic, perceptual hash & adversarial defense laboratory

1. Protection Layers

Active
Spatial
Bit Redundancy: 8x Block Copy Capacity 100%
pHash / dHash

Generates 64-bit spatial gradient & DCT low-frequency fingerprints resilient against minor transformations.

12%

Injects imperceptible high-frequency pixel perturbations to break unconsented img2img VAE feature extraction.

JUMBF Standard
Assertion: st.ai.generated

Protection Profile

Watermark Resilience High (Redundant)
pHash Bit-Length 64-bit Hex
C2PA Hash Match SHA-256 Valid
512x512 px
↔
LEFT: Protected Art | RIGHT: Ripped Attack
Visual Inspection Split Position: 50%
Split View Comparison Real-time Shader Map

Drag split slider to inspect pixel integrity between protected master and simulated rip artifact.

2. Rip Attack Simulator

65%
30%

3. Detection Audit

Live Compute
Signature Retrieval Confidence 94.2%
Threshold: 80.0% VERIFIED MATCH
pHash Hamming Dist 3 / 64 bits
SSIM Similarity Index 0.914
Detected Crop Bounds: [x:0, y:0, w:358, h:358]
C2PA Manifest Status: Stripped (Recoverable)
Attribution Verdict: OWNERSHIP PROVEN

What the watermark demo computes and what its verdict cannot prove

Read the explanation

The source copies the image data and writes repeated signature bits into the blue channel. Clearing its last bit with two hundred fifty four and inserting a one turns a channel value one hundred seventy two into one hundred seventy three. Before other transforms, that changes the value by at most one. The implementation repeats signature bits across consecutive pixels, despite its comments describing spatial blocks. It then adds random noise to all three channels, which can itself alter previously written least significant bits. The audit counts same position blue bit agreements between protected and transformed images. But it does not display that raw match rate. At thirty percent crop and quality sixty five, it assigns ninety four point two. Elsewhere it uses one hundred minus crop and quality penalties, clipped to a fixed range. Even disabling least significant bit protection does not make that formula a recovered signature. The OWNERSHIP PROVEN verdict is only a threshold label on this assigned score, not authenticated provenance or proof of authorship. A thirty percent crop keeps the top left three hundred fifty eight by three hundred fifty eight region from a five hundred twelve square image, about forty nine percent of its area, then stretches it back to the original dimensions. The quality control applies block averaging and random jitter. It does not encode and decode a real JPEG. Attack mode names do not select distinct transform algorithms in this function. The canvas shows an illustrative crop and compression proxy. The perceptual comparison samples nine by eight grayscale values and tests neighboring brightness, yielding sixty four difference hash bits. The DCT spectrum view is a radial sine illustration, not a Fourier transform. The SSIM label is a clipped PSNR derived approximation, not full structural similarity. The manifest includes a named issuer and timestamp but no actual signature verification or certificate chain. Exports save local demonstration outputs; they do not establish legal ownership, real watermark robustness, or compliance. The original canvas and controls remain intact.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.