Ghost Network Real-Time Privacy-Preserving Camera

Overview

This application captures live camera feed and applies the Ghost edge_diff transformation a privacy-preserving computer vision technique that strips all identity-revealing information while retaining motion/behavioral signals.

How Edge Diff Works

Frame t-1 ──→ Canny Edges ──┐
                     ├──→ absdiff ──→ Threshold ──→ Ghost Frame
Frame t   ──→ Canny Edges ──┘

1. Canny Edge Detection: Extract structural edges from consecutive frames
2. Absolute Difference: Compute pixel-wise difference between edge maps captures where edges moved
3. Thresholding: Apply fixed or Otsu adaptive threshold to produce binary mask
4. Output: 3-channel ghost frame showing only motion-induced edge changes

Privacy guarantee: The ghost frame contains zero texture, color, or facial feature information only the temporal change in structural edges. This is sufficient for behavioral analysis but insufficient for face recognition.

Files

File	Description
ghost_realtime.py	Lightweight OpenCV-only version (no GUI dependencies beyond OpenCV)
ghost_realtime_gui.py	Full tkinter GUI with matplotlib plots and recording
ghost_requirements.txt	Python dependencies

Quick Start

Install dependencies

sudo apt install python3-venv
sudo apt install python3-tk
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r ghost_requirements.txt

If you already have a virtual environment, activate it first and run the python -m pip install -r ghost_requirements.txt command inside that environment.

Run lightweight version

python ghost_realtime.py
python ghost_realtime.py --camera 0 --threshold 20 --adaptive
python ghost_realtime.py --width 1280 --height 720

Run GUI version

python ghost_realtime_gui.py

If the GUI starts with ModuleNotFoundError: No module named 'tkinter', install python3-tk with your OS package manager first.

Controls (Lightweight Version)

Key	Action
q / ESC	Quit
s	Save snapshot
a	Toggle adaptive (Otsu) threshold
SPACE	Pause/Resume
r	Reset Ghost state
+ / -	Adjust threshold

GUI Version Features

• Real-time parameter adjustment: threshold, Canny low/high, blur, morphological kernel
• Display modes: Side-by-side, Ghost only, RGB only, 4-panel view
• Live metrics plots: FPS, Privacy score, Edge density
• Recording: Save ghost video to MP4/AVI
• Snapshots: Save with JSON metadata

Privacy Score

The privacy score is estimated using face detection:

Privacy Score = 1 - (faces_detected_in_ghost / faces_detected_in_rgb)

• 1.0 (100%): No faces detected in ghost frames perfect privacy
• 0.0 (0%): Faces still detectable in ghost frames no privacy

The Haar cascade detector is used by default (ships with OpenCV). For more accurate privacy evaluation, install facenet-pytorch for MTCNN or insightface for RetinaFace.

If python ghost_realtime.py reports that it cannot open camera 0 and no /dev/video* devices are available, the machine has no accessible webcam. Use a different camera index if one exists, or connect/enable a camera device before running the app.

Architecture

┌─────────────┐   ┌──────────────┐   ┌─────────────┐   ┌──────────────┐
│ Camera    │────→│ Grayscale   │────→│ Canny Edges │────→│ absdiff    │
│ Frame (BGR) │   │ Conversion  │   │ Detection  │   │ (t vs t-1)  │
└─────────────┘   └──────────────┘   └─────────────┘   └──────┬───────┘
                                             │
                                             ▼
┌─────────────┐   ┌──────────────┐   ┌─────────────┐   ┌──────────────┐
│ Display   │←────│ 3-Channel   │←────│ Binary    │←────│ Threshold   │
│ (RGB+Ghost) │   │ Stack     │   │ Mask     │   │ (Fixed/Otsu) │
└─────────────┘   └──────────────┘   └─────────────┘   └──────────────┘

Citation

If you use this code in your research, please cite the original Ghost-ASD work:

@thesis{ghost_asd_2026,
 title={Privacy-Preserving ASD Behavioral Video Analysis using Ghost Networks},
 author={Ferhat Boulahia},
 year={2026}
}