PhantomShield

PhantomShield is a real-time deepfake detection system that monitors video conferencing applications like Zoom, Microsoft Teams, and Google Meet. It analyzes facial frames for deepfake indicators and provides real-time risk assessment.

🎯 Features

• ✅ Real-time deepfake detection using MesoNet architecture
• ✅ Dynamic grid-based monitoring system
• ✅ Real-time configuration adjustments
• ✅ Adaptive face detection with fallback mechanisms
• ✅ Risk score smoothing and buffering
• ✅ Automatic logging of risk scores and flagged frames
• ✅ Modern Electron-based UI

🛠️ Setup

1. Install dependencies:

# Install Python dependencies
pip install -r requirements.txt

# Install Node.js dependencies
npm install

2. Download the pre-trained model weights:

# Place Meso4_F2F.h5 in the model/ directory

3. Run the application:

# Run in development mode
npm run dev

# Build and run in production
npm run build

🎛️ Configuration

The application features a real-time configuration panel that allows you to adjust:

• Grid Layout

  • Rows (1-4): Number of vertical tiles
  • Columns (1-4): Number of horizontal tiles
  • Useful for different Zoom layouts and participant counts

• Detection Parameters

  • Alert Threshold (0-100%): Risk level that triggers alerts
  • Amplification Gain (0.1-5.0): Adjusts sensitivity
  • Buffer Size (1-100): Frames to average for smoothing

Changes take effect immediately without requiring restart.

📊 Monitoring

• Status Panel

  • Current operation status
  • Number of active tiles being monitored
  • Real-time alerts for detected deepfakes

• Alert System

  • Visual alerts for high-risk detections
  • Risk percentage display
  • Automatic alert clearing

• Logging

  • Risk scores logged to logs/risk_log.csv
  • Flagged frames saved to logs/faces/
  • Timestamps and tile coordinates included

🔍 How It Works

1. Initialization

   • Launches Electron UI
   • Starts Python backend
   • Establishes IPC communication

2. Detection Process

   • Captures Zoom window content
   • Divides into configurable grid
   • Performs face detection
   • Analyzes each face with MesoNet
   • Applies smoothing and thresholds
   • Generates alerts for suspicious content

3. Data Flow

   • Real-time frame processing
   • JSON-based communication
   • Bidirectional config updates
   • Asynchronous alert handling

⚠️ Notes

• Requires OpenCV and Pytorch
• Works best with well-lit, front-facing video
• May have false positives in low-light conditions
• Performance depends on:
  • System resources
  • Grid size configuration
  • Number of active participants

🔧 Troubleshooting

• Ensure Zoom window is visible and not minimized
• Check lighting conditions for better detection
• Adjust grid size to match Zoom's layout
• Fine-tune threshold and gain for your environment

📝 License

See LICENSE file for details

---