examples/README.md

Deepfake Hunter - Examples

This directory contains example files for testing and demonstration purposes.

Files

analysis_report.json

Example JSON report from video analysis showing:

• Overall deepfake detection result
• Frame-by-frame scores
• Temporal consistency analysis
• Physiological signal detection
• Detailed breakdown of all detection methods

Sample Videos (Not Included)

For testing, you can use these publicly available datasets:

Real Videos

• Download from: YouTube-8M
• Or use your own webcam recordings

Deepfake Videos

• FaceForensics++: https://github.com/ondyari/FaceForensics

  • Download sample videos for testing
  • Includes various manipulation methods (Deepfakes, Face2Face, FaceSwap, NeuralTextures)

• Celeb-DF: https://github.com/yuezunli/celeb-deepfakeforensics

  • High-quality celebrity deepfakes
  • More challenging test cases

• DFDC Preview: https://ai.facebook.com/datasets/dfdc/

  • Facebook Deepfake Detection Challenge dataset
  • Diverse range of manipulation techniques

Usage Examples

1. Analyze Sample Image

from deepfake_detector import DeepfakeDetector
import cv2

# Initialize detector
detector = DeepfakeDetector(use_gpu=True)

# Analyze image
image = cv2.imread('path/to/image.jpg')
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

result = detector.detect_image(image)
print(f"Is deepfake: {result.is_deepfake}")
print(f"Confidence: {result.confidence:.2%}")
print(f"Explanation: {result.explanation}")

2. Analyze Sample Video

from deepfake_detector import DeepfakeDetector
from video_analyzer import VideoAnalyzer

# Initialize
detector = DeepfakeDetector(use_gpu=True)
analyzer = VideoAnalyzer(detector, sample_rate=2)

# Analyze video
result = analyzer.analyze_video(
    'path/to/video.mp4',
    export_timeline=True,
    export_report=True,
    output_dir='./results'
)

print(f"Video is deepfake: {result.is_deepfake}")
print(f"Confidence: {result.confidence:.2%}")
print(f"Suspect frames: {len(result.suspect_frames)}")

3. Using the API

# Start the API server
python api_server.py

# Analyze image via API
curl -X POST "http://localhost:8001/analyze/image" \
  -H "X-API-Key: your-secret-key" \
  -F "file=@path/to/image.jpg" \
  -F "context=general"

# Analyze video via API
curl -X POST "http://localhost:8001/analyze/video" \
  -H "X-API-Key: your-secret-key" \
  -F "file=@path/to/video.mp4" \
  -F "context=general" \
  -F "sample_rate=2"

4. Using the Web Interface

# Start Gradio app
python app.py

# Open browser to http://localhost:7860
# Upload images/videos via the web interface

Creating Your Own Test Cases

Record Authentic Video

# Using webcam (Linux)
ffmpeg -f v4l2 -i /dev/video0 -t 10 real_video.mp4

# Using webcam (macOS)
ffmpeg -f avfoundation -i "0" -t 10 real_video.mp4

# Using webcam (Windows)
ffmpeg -f dshow -i video="Your Webcam Name" -t 10 real_video.mp4

Test Different Scenarios

1. Good Lighting: Well-lit face, clear features
2. Poor Lighting: Test robustness to lighting conditions
3. Multiple Faces: Test multi-face detection
4. Profile View: Test with non-frontal faces
5. Occlusions: Test with glasses, masks, etc.
6. Video Quality: Test with different resolutions and compression levels

Expected Results

Real Videos

• Confidence: 0.0 - 0.3 (low score = likely real)
• Physiological signals detected (if sufficient frames)
• Natural temporal consistency
• Low artifact scores

Deepfake Videos

• Confidence: 0.7 - 1.0 (high score = likely fake)
• No physiological signals or irregular patterns
• Temporal inconsistencies
• High artifact scores (blending, GAN fingerprints)

Benchmarking

To benchmark the detector on standard datasets:

# Download FaceForensics++ test set
# Then run benchmark
python benchmark.py \
  --dataset faceforensics \
  --data-dir /path/to/faceforensics \
  --output results/benchmark.json

Troubleshooting

Low Accuracy

• Check GPU is being used: detector.device should be "cuda"
• Ensure sufficient video length (>5 seconds for physiological signals)
• Try adjusting sensitivity: detector.sensitivity = "high"

Slow Processing

• Reduce video sample rate: analyzer.sample_rate = 4
• Use GPU acceleration
• Process at lower resolution
• Limit max frames: max_frames=300

False Positives

• Lower sensitivity: detector.sensitivity = "low"
• Increase thresholds: detector.set_thresholds({'ensemble': 0.8})
• Check for compression artifacts in original video

Contributing Examples

Have interesting test cases? Contribute them!

1. Create anonymized examples
2. Document expected results
3. Submit PR with examples and documentation

License

Example files are provided under CC0 (public domain) where possible. Some referenced datasets have their own licenses - check before use.