This project implements a deepfake detection system using both traditional machine learning and deep learning approaches. Facial images are processed and classified as real or fake using handcrafted visual features (for ML models) and deep neural representations. Alongside developing and evaluating classical ML models, the system incorporates insights from the Meso4 CNN architecture, a state-of-the-art deepfake detection framework. Performance comparisons are drawn between the implemented ML models and theoretical benchmarks from literature .
The dataset used in this project is a large, real-world deepfake dataset collected from various internet sources. It contains compressed .tar.gz archives hosting thousands of facial images extracted from videos. The dataset is publicly available on Hugging Face:
🔗 WildDeepfake on Hugging Face
A custom preprocessing pipeline was implemented to:
- Extract .png images from Hugging Face’s .tar.gz archives
- Organize and clean the images into labeled directories and
- Compute six image-based features (entropy, blur, noise, keypoints, blobs, and phase unwrapping).
The resulting structured dataset was used to train and evaluate machine learning models.
- Random Forest
- Support Vector Machine (SVM)
- Logistic Regression
- K-Nearest Neighbors (KNN)
- XGBoost
These models were trained using the extracted features and evaluated using standard performance metrics.
- Meso4 CNN (TensorFlow/Keras)
A compact convolutional neural network that processes raw images and learns spatial representations without manual feature engineering.
| Model | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|
| Meso4 (CNN) | 0.9400 | 0.9380 | 0.9420 | 0.9400 |
| Random Forest | 0.9162 | 0.9155 | 0.9250 | 0.9202 |
| KNN | 0.8884 | 0.8935 | 0.8929 | 0.8932 |
| Logistic Regression | 0.8884 | 0.8935 | 0.8929 | 0.8932 |
| XGBoost | 0.8065 | 0.7817 | 0.8737 | 0.8251 |
| SVM | 0.7246 | 0.7060 | 0.8106 | 0.7547 |
Among the classical ML models, Random Forest stood out as the best-performing one, coming second only to Meso-4, the deep learning-based model. This demonstrates that classical machine learning—when carefully feature-engineered and fine-tuned—can serve as a viable alternative to deep learning-based models, especially under computational constraints.