RONOVA is an end-to-end, air-gapped AI assurance, security vetting, and cryptographic provenance platform designed for safety-critical machine learning deployments (Defense, Healthcare, Finance, and Enterprise Infrastructure).
Modern machine learning supply chains are vulnerable to data poisoning, Trojan backdoors, adversarial perturbations, data leakage, and covert weight tampering.
RONOVA provides automated, deterministic pre-deployment validation, runtime anomaly detection, and tamper-evident audit trails with zero reliance on cloud verification.
┌────────────────────────────────────────────────────────┐
│ RONOVA ASSURANCE PIPELINE │
└───────────────────────────┬────────────────────────────┘
│
┌─────────────────────────┬─────────────┴────────────┬────────────────────────┐
▼ ▼ ▼ ▼
┌──────────────┐ ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Dataset │ │ STRIP │ │ Image Sentinel │ │ Distribution │
│ Forensics │ │ Trojan Detector │ │ Robustness │ │ Drift Analysis │
└──────┬───────┘ └────────┬────────┘ └────────┬────────┘ └────────┬────────┘
│ │ │ │
└────────────────────────┼─────────────────────────┴─────────────────────────┘
▼
┌─────────────────────────────────┐
│ Deterministic Safety Gate │ ◄── Multi-Profile Governance
│ (Defense / Health / Corp) │
└────────────────┬────────────────┘
▼
┌─────────────────────────────────┐
│ Tamper-Evident Ledger (Ed25519) │ ──► Signed QR Certificates
└─────────────────────────────────┘
- Embedding-Space Anomaly Detection: Employs an offline CNN feature extractor and calibrated Isolation Forest clustering to detect poisoned samples and mislabeled data.
- Dimensionality Reduction (PCA 2D): Visualizes high-dimensional dataset clustering in an interactive coordinate scatter space.
- Visual Outlier Thumbnails: Generates lightweight base64 thumbnail previews for flagged anomalous samples.
- Exact & Near-Duplicate Analysis: Identifies data leakage and training contamination using perceptual hashing and vector similarity thresholds.
- Implements STRIP (Strong Perturbation Trojan Detection): Superimposes clean image overlays onto candidate test inputs and measures the Shannon entropy collapse across model output probability vectors.
- Entropy Drop Isolation: Backdoored inputs maintain low prediction entropy regardless of noise overlays, cleanly exposing Trojan triggers without needing training data access.
- Detects high-frequency adversarial gradient attacks (FGSM, PGD) and distribution corruption.
- Measures prediction stability across varying noise scales (
$\sigma \in [0.05, 0.20]$ ). - Provides visual pixel-level perturbation heatmaps and channel-wise variance maps.
- Detects inference data shift using Wasserstein Distance and Kolmogorov-Smirnov (KS) tests.
- Computes feature-level drift scores to alert teams before downstream model accuracy degrades.
- Configurable risk profiles:
- 🛡️ Defense / Tactical: Zero-tolerance strict gating (Rejection on any hard-gate indicator).
- 🏥 Healthcare / Life Sciences: Stringent anomaly boundaries and drift limits.
- 🏢 Enterprise / Standard: Balanced trade-off between throughput and assurance.
- Explains all findings with deterministic rationale, limitation disclosures, and explicit evidence strength ratings.
- Every scan event is committed to a hash-chained audit log signed via Ed25519 asymmetric cryptography.
- Produces portable, self-contained Security Verification Certificates with cryptographically signed payload QR codes.
- Includes a standalone verification tool (
verify_certificate.py) capable of proving certificate authenticity offline without network connectivity.
| Layer | Technologies Used |
|---|---|
| Core AI & Math | PyTorch, ONNX Runtime, NumPy, Scikit-learn, SciPy |
| Backend API | FastAPI, Uvicorn, Pydantic v2 |
| Cryptography | cryptography (Ed25519, SHA-256), qrcode, Pillow |
| Frontend UI | React 18, Vite, Tailwind CSS, Lucide Icons, Recharts |
| Sandboxing | Python Subprocess Workers, Resource-Capped Memory/CPU Execution |
- Python 3.10+
- Node.js 18+ and npm
# Clone the repository
git clone https://github.com/developerHarish2007/Ronova.git
cd Ronova
# Create and activate virtual environment
python -m venv venv
# On Windows:
.\venv\Scripts\activate
# On Linux/macOS:
source venv/bin/activate
# Install dependencies
pip install -r requirements.txt # or pip install fastapi uvicorn torch onnxruntime numpy scikit-learn pillow cryptography qrcode# Generate Ed25519 provenance root keys
python scripts/init_provenance_keys.py
# Create synthetic evaluation models and sample datasets
python scripts/train_models.py
python scripts/create_demo_datasets.pypython -m uvicorn ronova.api.main:app --host 127.0.0.1 --port 8000 --reload- Interactive Swagger Docs: http://127.0.0.1:8000/docs
cd frontend
npm install
npm run dev- Web Dashboard: http://localhost:5173/
Run the automated test harnesses to validate individual modules:
# Test Dataset Forensics & Anomaly Isolation
python scripts/test_dataset_forensics.py
# Test Cryptographic Provenance & Ledger Chains
python scripts/test_provenance.py
python scripts/test_audit_ledger.py
# Test STRIP Backdoor Detection
python scripts/test_image_sentinel.py
# Run Complete End-to-End Regression Suite
python scripts/run_phase8_regression.pyRonova/
├── ronova/
│ ├── api/ # FastAPI REST endpoints & request handlers
│ ├── core/ # Pydantic schemas, types, and finding data structures
│ ├── dataset/ # Dataset anomaly clustering & duplicate detection
│ ├── detectors/ # STRIP Trojan detection & Image Sentinel algorithms
│ ├── engine/ # Assurance policy evaluation rules
│ ├── operations/ # Statistical drift computation (KS-test / Wasserstein)
│ ├── provenance/ # Ed25519 signature engine, ledger chain & certificates
│ ├── safety/ # Verdict gating (Defense, Healthcare, Enterprise)
│ └── sandbox/ # Sandboxed ONNX runtime execution runners
├── frontend/
│ ├── src/
│ │ ├── components/ # React cards (DatasetForensics, ImageSentinel, Drift, Provenance)
│ │ ├── App.jsx # Main interactive assurance dashboard
│ │ └── index.css # Tailwind styling and dark mode glassmorphism
│ └── package.json
├── data/ # Synthetic datasets for test replication
├── models/ # Pre-compiled ONNX models (clean vs backdoored)
├── scripts/ # End-to-end execution and regression scripts
├── verify_certificate.py # Offline CLI tool for authenticating signed QR certificates
└── README.md
- 100% Offline Capability: Runs completely local without third-party cloud API dependencies.
- Deterministic Cryptography: All security claims are backed by non-malleable Ed25519 digital signatures.
- Memory-Safe Execution: ONNX model evaluations run within sandboxed runner environments with bounded memory allocation.
This project is licensed under the MIT License.