🔍 Deepfake Detection Using CNN This project is a deep learning-based solution for detecting real vs. fake (deepfake) facial images using a Convolutional Neural Network (CNN). It uses a lightweight architecture trained on a balanced dataset of real and AI-generated (fake) face images. The solution includes a web-based interface built with Streamlit that allows users to upload an image and get an instant prediction.
📌 Table of Contents
About the Project
Features
Tech Stack
Dataset
Model Architecture
Usage
Results
Limitations & Future Scope
About the Project
Deepfakes are synthetic media in which a person’s likeness is replaced with someone else’s using artificial intelligence. These can be highly deceptive and pose a threat to digital integrity. This project focuses on detecting such images using CNNs and showcasing results interactively via a web app.
Features
Upload any face image and detect if it is Real or Fake
CNN-based binary image classification
Trained on 80,000 facial images (40k real, 40k fake)
Real-time predictions with probability score
Web UI built using Streamlit
Model Performance Summary with accuracy/loss
⚙️ Tech Stack
Python
TensorFlow / Keras
OpenCV
NumPy
Streamlit (for UI)
Matplotlib (for model visualization)
📂 Dataset
We used the 140K Real and Fake Face Dataset, which consists of: 70,000 Real Images collected from Flickr and real photo datasets.
70,000 Fake Images generated using StyleGAN and other AI generators.
Split used:
Training Set: 40,000 Real + 40,000 Fake
Validation Set: 10,000 Real + 10,000 Fake
🧠 Model Architecture
Input size: 96x96x3
Convolutional layers with ReLU activation
MaxPooling layers
Dropout for regularization
Flatten + Dense layers
Sigmoid activation for binary output
🚀 Usage
1.Launch the app.
2.Upload a .jpg, .jpeg, or .png face image.
3.The model will output whether the image is REAL or FAKE, along with confidence.
4.A brief explanation is shown based on the result.
📊 Results
Validation Accuracy: 93.98%
Validation Loss: 0.1577
CNN was able to distinguish between deepfakes and real images effectively.
Good generalization with unseen images from external sources.
🚧 Limitations & Future Scope
Limitations:
Focused on still images only, not videos.
May fail on highly realistic fake images with minimal artifacts.
Does not support batch image analysis.
Future Scope:
Extend to deepfake video detection.
Use transformer-based or attention mechanisms.
Add heatmap visualizations for feature attention.
Enable API endpoints for broader integrations.