DeepSafe — AI Face Detector

Detects AI-generated fake faces using EfficientNet-B0. Trained on a combined dataset of 140K face images (FFHQ real + StyleGAN2 fake) + diverse generator samples for better generalization.

Performance

Metric	Value
Accuracy	88.0%
Precision	94.0%
AUC-ROC	96.5%
Fake Detection Rate	81.6%
Real→Fake False Positives	5.3%

Dataset

• 140K Real and Fake Faces (Kaggle: xhlulu/140k-real-and-fake-faces)
  • 70,000 real FFHQ faces + 70,000 StyleGAN2 fake faces
  • Pre-split: 50K/10K/10K train/valid/test per class
• Deepfake-vs-Real (Hugging Face: prithivMLmods/Deepfake-vs-Real)
  • Diverse AI-generated fake images from multiple GAN/diffusion models
  • 1,200+ fake images from various generators for improved generalization
• Real human reference photos for threshold calibration (real-world test images)

Files to Upload to GitHub

app.py              # Gradio web interface with Grad-CAM
train.py            # Training script with balanced subset support
evaluate.py         # Test evaluation with metrics
predict.py          # Single/batch image prediction
requirements.txt    # Python dependencies
README.md           # This file

Do NOT upload:

• checkpoints/ — model weights (16MB, regenerate via train.py)
• data_140k/ — dataset (download separately)
• data_extra/, data_extra2/, data/ — extracted temp files
• results/ — evaluation outputs
• __pycache__/, .gradio/ — caches

Setup

pip install -r requirements.txt

Download the dataset and extract to data_140k/:

data_140k/
├── train/{real,fake}/
├── valid/{real,fake}/
└── test/{real,fake}/

Training

python train.py --epochs 10 --subset 10000

Use --subset N for N total images (N/2 per class). All extra files are prioritized to ensure diverse generator coverage.

Evaluation

python evaluate.py

Prediction

python predict.py path/to/image.jpg
python predict.py img1.jpg img2.jpg img3.jpg

Interactive App

python app.py

Opens a Gradio web UI with:

• Image upload + Grad-CAM heatmap overlay
• Batch analysis mode
• Conservative threshold (default 0.99) to minimize false positives on real photos

Threshold

Default threshold: 0.99 — only images with ≥99% fake probability are classified as FAKE. This prioritizes minimizing false positives on real human photos.

To adjust:

# Edit results/threshold.json
{"threshold": 0.95}