DeepDetect: DeepFake Face Detection Using Machine Learning and Deep Learning

This project presents a comparative analysis of three different models for deepfake image detection, focusing specifically on face images. Deepfakes pose a significant threat in the digital world due to their ability to produce highly realistic fabricated visuals. This work evaluates and benchmarks multiple approaches to understand which model performs best for detecting manipulated human faces.

📌 Models Compared

1. Logistic Regression

• Binary classifier for real vs fake faces.
• Works on flattened image pixels.
• Test Accuracy: 84%

2. Support Vector Machine (SVM)

• Kernel-based classifier using handcrafted features like HOG and LBP.
• More robust than logistic regression but struggles with complex facial artifacts.
• Test Accuracy: 76%

3. Convolutional Neural Network (CNN – ResNet-18)

• Deep learning model capable of learning spatial and hierarchical features automatically.
• Best suited for image-related tasks.
• Test Accuracy: 96%

📂 Dataset

• Source: Real vs AI Generated Faces Dataset from Kaggle
• Contains only face images (real and AI-generated).
• Total Images Used: ~10,000
  • 7,000 — Training
  • 1,500 — Validation
  • 1,500 — Test

🛠️ Preprocessing & Feature Engineering

For Logistic Regression & SVM

• Resized to 128×128
• Normalization
• Handcrafted features (HOG, LBP)
• Flattened pixel vectors for logistic regression

For CNN (ResNet-18)

• Resized to 224×224
• Normalized using ImageNet stats
• Data augmentation (flips, rotations, color jitter)
• End-to-end feature learning

📊 Performance Evaluation Metrics

Models were evaluated on:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion Matrix

🏆 Results Summary

Model	Test Accuracy	Summary
Logistic Regression	84%	Strong baseline but limited for image complexity
SVM	76%	Good with handcrafted features but weaker overall
CNN (ResNet-18)	96%	Best performance; excels at learning facial patterns

✔ CNN achieved the highest accuracy (96%), clearly outperforming traditional models due to its strong feature extraction capabilities.

🔍 Key Insights

• The current system detects only face-based deepfakes.
• Logistic Regression and SVM cannot fully capture complex manipulation artifacts.
• CNN models like ResNet-18 provide significantly better results because they learn visual features hierarchically.
• Deep learning is essential for modern deepfake detection systems.

🚀 Future Work

Extend detection beyond faces:

• Full-body deepfakes
• Video deepfakes
• Audio-visual manipulations

Model & training improvements:

• Implement ensemble models
• Add more advanced data augmentations
• Experiment with EfficientNet, MobileNet, Vision Transformers (ViT)

📁 Repository Structure

├── Models/
│ ├── Simple_Logistic_Regression.ipynb
│ ├── Support_Vector_Machine.py
│ ├── CNN_RestNet.py
├── Gradio_Interface.ipynb
├── Results.md
├── README.md
└── requirements.txt

⚡ Interface

A simple single web page has been designed by Gradio in Google Colab. Steps to run the Interface:

1. Generate the model file deepfake_best.pth by the CNN_RestNet notebook.
2. Upload the model in Google Drive.
3. From Google Colab, run the first cell to mount the drive in Google Colab
4. Run the second cell to run Gradio Interface and click on the local URL or public URL