Flashivy-Deepfake v1

Deepfake detection using deep learning. This version establishes a baseline CNN model trained on images extracted from FaceForensics++ videos.

---

📚 Table of Contents

• 🚀 What This Tool Does (v1)

• 📂 Dataset Used (v1)

  • Download REAL videos
  • Download FAKE videos

• ⚙️ Environment Setup

  • PyTorch
  • Other Dependencies
  • CUDA Check

• 📌 Version Scope

• 📊 Results

• 🔍 Explainability (Grad-CAM)

  • Run Grad-CAM

• ▶️ Inference

  • 🔍 Running Inference
  • 📤 Output
  • 📊 Interpretation of Confidence
  • ⚠️ Important Note on Generalization

• 📄 License & Usage

---

🚀 What This Tool Does (v1)

• Classifies face images as Real or Deepfake
• Learns visual artifacts introduced by face-swap manipulation
• Designed as a foundation for explainability and video-level models in later versions

v1 is image-based (frames extracted from videos).

---

📂 Dataset Used (v1)

FaceForensics++ (Image Frames)

Images are extracted from the FaceForensics++ dataset using the official downloader.

Composition:

• Real: Frames from pristine YouTube videos
• Fake: Frames from DeepFakes manipulated videos
• Compression: c23
• Balanced dataset (equal number of real and fake frames after extraction)

FaceForensics++ is a standard benchmark dataset widely used in deepfake detection research.

---

📥 Dataset Download

The dataset is downloaded using the official FaceForensics++ script. To obtain access or use the original script included in this repository, users must fill out the Google Form and accept the terms available on the FaceForensics website.

Note: The dataset itself is not included in this repository. Users must download the data directly from the official source in accordance with the FaceForensics++ Terms of Service.

Download REAL videos

python faceforensics_download_v4.py ff_data \
  --dataset original \
  --compression c23 \
  --type videos \
  --num_videos 25 \
  --server EU2

Download FAKE videos

python faceforensics_download_v4.py ff_data \
  --dataset Deepfakes \
  --compression c23 \
  --type videos \
  --num_videos 25 \
  --server EU2

Image frames are extracted from videos, and face crops are generated using the OpenCV Haar Cascade classifier (haarcascade_frontalface_default.xml) for training.

---

⚙️ Environment Setup

PyTorch

• PyTorch 2.2+
• CUDA 12.1
• Optimized for RTX 4060 Laptop GPU

Install PyTorch manually:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

---

Other Dependencies

pip install -r requirements.txt

---

CUDA Check

python check_cuda.py

---

📌 Version Scope

Included:

• Image-based deepfake classification
• CNN-based training pipeline
• FaceForensics++ data
• Grad-CAM explainability

Not included:

• Video-level modeling
• Cross-dataset testing

---

📊 Results

The model achieves high validation accuracy on the FaceForensics++ dataset due to the presence of clear manipulation artifacts in face-swap deepfakes. This version focuses on establishing a strong image-based baseline, while robustness across datasets is addressed in later versions.

---

🔍 Explainability (Grad-CAM)

To interpret the model’s predictions, Grad-CAM is used to visualize which facial regions contribute most to the real vs fake classification.

Grad-CAM heatmaps indicate that the model primarily focuses on facial regions such as the eyes, mouth, cheeks, and skin texture, which are known to contain manipulation artifacts in face-swap deepfakes.

Run Grad-CAM

PYTHONPATH=. python - <<EOF
from explainability.grad_cam import plot_real_vs_fake

plot_real_vs_fake(
    "data/faces/real/IMAGE_NAME.jpg",
    "data/faces/fake/IMAGE_NAME.jpg"
)
EOF

Example Grad-CAM visualization comparing real and fake face images.

Grad-CAM visualizations show that for real images, the model focuses on identity-related facial regions such as the eyes, while for fake images, attention is distributed across facial textures and blending regions, indicating the presence of manipulation artifacts.

---

▶️ Inference

This repository provides a command-line inference script to perform deepfake prediction on a single face image using the trained v1 CNN model.

🔍 Running Inference

Use the following command to run inference:

PYTHONPATH=. python predict.py --image path/to/image.jpg

Input Requirements

• The input image should ideally be face-cropped
• Images similar to those used during training yield the best results
• For optimal performance, use images from the data/faces/ directory

---

📤 Output

The script outputs:

• Predicted label: REAL or FAKE
• Confidence score: Model certainty for the predicted class

Example Output

Prediction : FAKE
Confidence : 0.9821

---

📊 Interpretation of Confidence

• The confidence score represents the softmax probability of the predicted class.

• It reflects the model’s certainty within the training distribution (FaceForensics++).

• A high confidence score does not guarantee correctness on images from:

  • Unseen datasets
  • Different deepfake generation techniques

---

⚠️ Important Note on Generalization

The v1 model is trained exclusively on the FaceForensics++ DeepFakes subset.

While it achieves high accuracy on in-distribution samples, it may struggle to generalize to deepfakes created using other methods or datasets.

In particular:

• Modern or visually clean deepfakes
• Deepfakes lacking FaceForensics++-style artifacts

may be incorrectly classified as REAL, sometimes even with high confidence.

This limitation highlights the well-known dataset bias problem in image-based deepfake detection.

---

📄 License & Usage

• Code License: MIT License
• Dataset Usage: FaceForensics++ data is used strictly in accordance with its Terms of Service
• Intended Use: Academic and educational purposes only