EE656_Project# Deepfake Detection with EfficientNet-B0

A binary image classification pipeline for detecting AI-generated (deepfake) faces in video content, built with PyTorch and EfficientNet-B0. The project covers the full ML workflow: data preprocessing, model training, evaluation, and robustness testing under image perturbations.

---

Project Structure

Notebook	Purpose
preprocessing_fixed.ipynb	Extract and prepare face crops from raw videos
training.ipynb	Fine-tune EfficientNet-B0 for real vs. fake classification
evaluation.ipynb	Evaluate the trained model on the test set
perturbed.ipynb	Test model robustness under various image distortions

---

Notebooks

1. preprocessing_fixed.ipynb — Data Preparation

Processes raw video files from the Celeb-DF dataset into face-cropped JPEG images ready for training.

• Loads real videos from Celeb-real/ and synthetic videos from Celeb-synthesis/
• Balances classes by randomly sampling fake videos to match real count (seed 42)
• Splits data into train / val / test sets (80 / 10 / 10)
• Uses MTCNN (via facenet-pytorch) to detect and crop faces from video frames
• Samples one face every 15 frames, up to 5 faces per video, resized to 224×224
• Saves crops to processed/{split}/{real|fake}/

Output structure:

processed/
├── train/
│   ├── real/
│   └── fake/
├── val/
│   ├── real/
│   └── fake/
└── test/
    ├── real/
    └── fake/

---

2. training.ipynb — Model Training

Fine-tunes a pretrained EfficientNet-B0 model for binary classification (real vs. fake).

• Replaces the final classifier head (1000 → 2 classes)
• Uses ImageNet pretrained weights as starting point
• Training augmentations: random horizontal flip, color jitter, Gaussian blur, Gaussian noise
• Optimizer: Adam (lr=1e-4), Loss: CrossEntropyLoss
• Trains for 20 epochs, saving the best checkpoint as best_model.pth
• Plots training/validation accuracy and loss curves

---

3. evaluation.ipynb — Model Evaluation

Evaluates best_model.pth on the held-out test set.

• Computes overall test accuracy
• Generates a confusion matrix (seaborn heatmap)
• Prints a full classification report (precision, recall, F1 per class)
• Plots the ROC curve with AUC score

---

4. perturbed.ipynb — Robustness Testing

Tests how well the model performs when the input images are corrupted, simulating real-world degradation.

Perturbations tested:

Perturbation	Description
White Gaussian Noise	Random noise with mean ∈ [-0.3, 0.3], variance ∈ [0, 1]
Gaussian Blur	Random kernel size 11–21 (odd values)
JPEG Compression	Random quality factor 30–95
Color Contrast	Enhancement factor ∈ [0.5, 2.5]
Color Saturation	Enhancement factor ∈ [0.0, 2.5]

For each perturbation, the notebook reports AUC and plots the ROC curve.

---

Requirements

torch
torchvision
opencv-python
Pillow
facenet-pytorch
scikit-learn
matplotlib
seaborn
tqdm
numpy

Install with:

pip install torch torchvision opencv-python Pillow facenet-pytorch scikit-learn matplotlib seaborn tqdm numpy

A CUDA-capable GPU is recommended for preprocessing and training.

---

Usage

Run the notebooks in order:

1. preprocessing_fixed.ipynb   # prepare the dataset
2. training.ipynb              # train and save best_model.pth
3. evaluation.ipynb            # evaluate on the test set
4. perturbed.ipynb             # robustness analysis

---

Dataset

This project uses the Celeb-DF dataset. Place the raw videos at:

~/Downloads/Celeb-real/        # real celebrity videos (.mp4)
~/Downloads/Celeb-synthesis/   # AI-synthesized videos (.mp4)

The dataset is not included in this repository and must be obtained separately.