Citadel — AI-Proof Infrastructure for Creators

A platform that protects artists from unauthorized AI training by embedding cryptographic signals into their images and recording tamper-proof commitments on the Base Sepolia blockchain.

---

Files

File	Description
app.py	Streamlit demo app (3 tabs: Image Protection, AI Detection, Dashboard)
citadel_signal.py	Radioactive Signal injection via CLIP ViT-B/32
citadel_detector.py	Black-box statistical detection + court-ready report generation
blockchain.py	Base Sepolia on-chain commitment (commit-reveal scheme)
precomputed_detection.json	Precomputed detection statistics (Mann-Whitney U, Cohen's d)

---

Setup

pip install -r requirements.txt

torch and open_clip_torch are required for real signal injection. Without them, the app runs in simulation mode automatically.

Environment Variables

Create a .env file in the project root:

CITADEL_PRIVATE_KEY=0xyour_base_sepolia_private_key

Required for real Base Sepolia transactions. Without this, blockchain commits will fail.

---

Run

streamlit run app.py

---

Demo Walkthrough

Tab 1 — Image Protection Upload an image → Citadel injects a Radioactive Signal → SHA-256 commitment is recorded on Base Sepolia → download protected image.

Tab 2 — AI Detection Load precomputed detection stats → view Mann-Whitney U test results (p < 10⁻⁶³, Cohen's d = 3.01) → generate court-ready evidence report with on-chain commit-reveal verification.

⚠️ Detection statistics are simulated values based on Radioactive Data (Sablayrolles et al., ICML 2020). Not derived from an actually trained model.

Tab 3 — Dashboard Platform metrics and signal version history.

⚠️ All dashboard metrics are mock data for demonstration purposes.