StegaDNA-V5-StegaStamp 🧬
StegaDNA-V5-StegaStamp is a high-performance deep steganography model mathematically designed to embed an indestructible "DNA" digital signature into images, making them robust to print-and-scan and screen-capture distortions (the "Analog Hole").
🚀 Key Achievements
- Near-Perfect Visual Fidelity: Attains ~48 dB PSNR, rendering the embedded data invisible to the human eye.
- Reed-Solomon Error Correction: Encodes the message with Reed-Solomon (via Mojo) mapping 10 signature bytes and 6 ECC repair bytes into 128 Bits, meaning full text recovery is achieved even if up to 20% of bits flip during camera scanning.
- Camera-Angle Localization: Built natively with a Spatial Transformer Network (STN) enabling the Decoder to automatically predict affine homography adjustments on captured imagery before reading the bits.
- Domain-Shift Immune: Eliminates traditional BatchNorm failures by utilizing
InstanceNorm2d, guaranteeing pure mathematical stability during isolated physical deployments.
🏗️ Architecture Stack
This V5 architecture relies on a specialized workflow:
- U-Net Encoder: A multi-scale residual generator that disperses the 128-bit payload into redundant, variable-frequency spatial features without forming visible artifacts.
- Kornia Adversarial Engine: During training, simulated 3D perspectives, zoom, brightness fluctuations, and lens blurs force the encoder into securing extreme physical robustness.
- STN-Empowered Decoder: An internal affine tracker aligns the distorted image before high-entropy extraction.
🛠️ Usage
This model requires the StegaDNAUniversalGateway to operate out of the box.
Install the project infrastructure from the source repository:
git clone https://github.com/antonvice/StegaDNA.git
cd StegaDNA
Install using uv:
uv sync
With the server running, you can hit the local endpoint to automatically encode or decode via the browser or Python requests.
import requests
# Example of stamping an image with DNA text
files = {'file': open('asset.jpg', 'rb')}
data = {'user_id': 'Secret Provenance Data'}
resp = requests.post('http://localhost:8000/stamp/image', files=files, data=data)
with open('stamped.png', 'wb') as f:
f.write(resp.content)
📈 Evaluation Matrix
Current isolated network performance metrics per epoch on hold-out standard clean-image splits under heavy v3 noise:
- BER (Bit Error Rate): Converging toward < 0.15 (Sufficient for Reed-Solomon guaranteed string recovery)
- PSNR (Peak Signal to Noise): > 48.0 dB
Developed as part of the Grably Data Engineering ecosystem. 🛡️