Bio-Sentinel is an Agentic AI cybersecurity solution designed to detect sophisticated deepfakes and "Virtual Injection" attacks in real-time. Unlike traditional passive detectors, Bio-Sentinel employs an active "Challenge-Response" agent and biological signal analysis (rPPG) to verify liveness.
Current media forensics tools are failing against:
- GenAI Agents: Hyper-realistic deepfakes (Sora/Kling) that bypass pixel-level detection.
- Virtual Camera Injection: Attackers feeding pre-recorded fake streams directly into KYC/Auth systems, bypassing physical hardware.
- Passive Failure: Static liveness checks (blinking/smiling) are easily automated by scripts.
We propose a multi-modal verification stack that validates the human, not just the image.
- 🫀 rPPG Bio-Liveness: Extracts remote photoplethysmography (heartbeat) signals from facial skin pixels. Generative AI creates perfect pixels, but it cannot simulate a consistent blood flow pulse.
- 🤖 Agentic Challenge-Response: An autonomous AI agent flashes random color sequences or prompts micro-gestures to measure reaction latency, defeating scripted replay attacks.
- 🔒 Hardware Integrity Layer: Low-level driver analysis to detect and block "Virtual Camera" software (e.g., OBS, DeepFaceLive) attempting to inject video feeds.
- ⚡ Edge-Optimized: Uses quantized models (INT8) to run locally on consumer hardware with <50ms latency.
Our ML Pipeline follows a fusion-based approach:
- Input Layer: Live Video Stream (WebRTC) + Audio Feed.
- ROI Extraction: MTCNN detects and aligns the face mesh.
- Spatial Analysis: EfficientNet-B0 scans for pixel-level artifacts and texture inconsistencies.
- Temporal Analysis: LSTM/GRU networks analyze frame sequences for unnatural movement dynamics.
- Bio-Signal Extraction: SciPy/OpenCV extracts rPPG waveforms to verify pulse presence.
- Decision Output: A weighted fusion layer generates a "Trust Score" (0-100%).
(See docs/architecture_diagram.png for visual workflow)
| Component | Technology |
|---|---|
| Core AI Framework | PyTorch / TensorFlow |
| Computer Vision | OpenCV, MediaPipe (Face Mesh) |
| Signal Processing | SciPy (rPPG extraction) |
| Backend API | FastAPI (Python) |
| Frontend Dashboard | React.js / Streamlit |
| Deployment | Docker Containers |
| Database | PostgreSQL (Audit Logs) |
# Clone the repository
git clone [https://github.com/YourUsername/Bio-Sentinel-Core.git](https://github.com/YourUsername/Bio-Sentinel-Core.git)
# Navigate to directory
cd Bio-Sentinel-Core
# Install dependencies
pip install -r requirements.txt
# Run the local inference server
python main.py --mode=edge[x] Phase 1: Concept & Architecture Design (Complete)
[ ] Phase 2: rPPG Model Training on SiW Dataset
[ ] Phase 3: Hardware Driver Analysis Module
[ ] Phase 4: Integration with Zoom/Teams API
Project Lead: Shrey Pandey Hackathon: eRaksha 2026 Theme: Agentic AI / Deepfake Detection