StegaDNA 🧬

Adversarial Deep-Digital Watermarking with Mojo-Powered Error Correction.

StegaDNA is a high-performance steganography engine designed to "stamp" any medium (Images, Audio, Text) with indestructible digital signatures (DNA). It leverages a hybrid architecture combining the flexibility of PyTorch for neural embedding and the raw speed of Mojo for bit-level error correction.

RESEARCH AT WANDB

🚀 Key Features

• Indestructible DNA: Uses Reed-Solomon ECC implemented in Mojo with SIMD optimizations to recover signatures even from partially destroyed media.
• The Analog Hole (v4 Breakthrough): Specialized Multi-Scale Inception Kernels and a Dense Bit Projection Network allow DNA survival under real-world capture conditions (phone cameras, print-and-scan) while maintaining high visual fidelity (20+ dB PSNR).
• Adversarial Robustness: Neural networks trained against a differentiable Noise Layer (Geometric warping, color jitter, variable blur) to survive digital and physical attacks.
• Universal Gateway: A unified FastAPI controller that routes traffic to specialized modality engines.
• Mac Optimized: Native MPS (Metal Performance Shaders) support for lightning-fast training on Apple Silicon.
• Research-Driven: Fully documented architectural evolutions and training experiments available in research.md.

---

🏗️ Architecture (v4 Multi-Scale)

Media Input ───► [ Multi-Scale Encoder ] ───► [ Noise Layer V3 ] ───► [ High-Entropy Decoder ]
                          ▲                   (Analog Distortion)              │
                          │                                                    │
                 [ 128-bit DNA Bits ] <────────────────────────────────────────┘
                  (Dense Projection)

The v4 architecture utilizes a U-Net style Encoder with triple-path kernels (3x3, 5x5, 7x7) to capture and hide data across multiple spatial frequencies, suppressing the "foggy blob" artifacts common in earlier deep steganography models.

---

🛠️ Tech Stack

• Core Logic: Mojo (SIMD, GF(2^8) Arithmetic)
• Neural Engine: PyTorch (Modified HiDDeN Architecture)
• API Layer: FastAPI + Uvicorn
• Infrastructure: UV (Fast Python Package Manager)
• Monitoring: Weights & Biases + Loguru

---

🚦 Getting Started

1. Installation

Ensure you have uv and mojo installed on your Mac.

uv sync

2. Build the Mojo Core

Compile the high-performance Reed-Solomon module:

uv run mojo build mojo_core/ecc.mojo --emit shared-lib -o dna_ecc.dylib

3. Training (Robust V5)

Launch the robust StegaStamp adversarial training loop which uses InstanceNorm and STN for camera-angle survival:

bash train_stegastamp.sh

4. Serve the API & Interactive Dashboard

Run the universal gateway and access the premium web dashboard:

uv run python main.py

Then navigate to: http://localhost:8000

---

🗺️ Roadmap

• Phase 1: Mojo Reed-Solomon Core (SIMD Optimized).
• Phase 2: Universal FastAPI Gateway & Modality Routing.
• Phase 3: Adversarial Training Pipeline (MPS Enabled).
• Phase 4: Analog Hole Survival (Multi-Scale Pivot).
• Phase 5: STFT-Transformer Engine for Audio Watermarking.
• Phase 6: LLM Logit Processor for Text DNA.
• Phase 7: ONNX Export for Edge Deployment.

---

Developed as part of the Grably Data Engineering ecosystem. 🛡️