Open-source image classification model for deepfake vs real image detection - Free detection of real and AI forged images

Deepfake Vs Real Image Detection

Developed by dima806

An image classification model based on Vision Transformer architecture, used to detect real images versus AI-generated fake images.

Image Classification

Transformers

Open Source License:Apache-2.0 #Deepfake Detection #High-precision Classification #ViT Architecture

Downloads 129.66k

Release Time : 10/14/2023

Model Overview

This model is based on Google's ViT architecture, specifically designed to distinguish between real human face images and AI-generated deepfake images. It demonstrates extremely high accuracy (99.27%) on test datasets.

Model Features

High Accuracy Detection

Achieves 99.27% accuracy and F1 score on test sets, effectively distinguishing real from fake images.

Based on ViT Architecture

Utilizes Vision Transformer architecture with self-attention mechanisms to capture global image features.

Fine-tuned Pre-trained Model

Fine-tuned from a ViT model pre-trained on ImageNet-21k, featuring powerful feature extraction capabilities.

Model Capabilities

Image Authenticity Analysis

Deepfake Detection

Binary Image Classification

Use Cases

Content Moderation

Social Media Fake Content Detection

Automatically identifies AI-generated fake face images uploaded by users.

Can reduce the spread of fake images by over 99%.

Security Verification

Enhanced Identity Authentication Systems

Prevents attackers from using AI-generated fake images to spoof facial recognition systems.

Significantly improves the security of biometric systems.

🚀 Image Authenticity Detection Model

This project provides a model to check whether an image is real or fake (AI-generated).

🚀 Quick Start

The model checks whether an image is real or fake (AI-generated). It's based on the google/vit-base-patch16-224-in21k base model.

✨ Features

Accuracy Metrics: Uses accuracy and f1 as evaluation metrics.
Base Model: Built upon google/vit-base-patch16-224-in21k.

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

No code examples are provided in the original document.

📚 Documentation

Important Note for Production Use

Beware that this model is trained on a dataset collected about 3 years ago. Since then, there is a remarkable progress in generating deepfake images with common AI tools, resulting in a significant concept drift. To mitigate that, I urge you to retrain the model using the latest available labeled data. As a quick-fix approach, simple reducing the threshold (say from default 0.5 to 0.1 or even 0.01) of labelling image as a fake may suffice. However, you will do that at your own risk, and retraining the model is the better way of handling the concept drift.

For more details, see https://www.kaggle.com/code/dima806/deepfake-vs-real-faces-detection-vit.

Classification Report

Classification report:

              precision    recall  f1-score   support

        Real     0.9921    0.9933    0.9927     38080
        Fake     0.9933    0.9921    0.9927     38081

    accuracy                         0.9927     76161
   macro avg     0.9927    0.9927    0.9927     76161
weighted avg     0.9927    0.9927    0.9927     76161

📄 License

This project is licensed under the Apache-2.0 license.

Property	Details
Model Type	Based on `google/vit-base-patch16-224-in21k`
Training Data	Not specified, but collected about 3 years ago

⚠️ Important Note

This model is trained on a dataset collected about 3 years ago. There has been significant progress in deepfake image generation, leading to concept drift. It's recommended to retrain the model with the latest labeled data. Reducing the threshold for labelling images as fake can be a quick fix, but it's at your own risk.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご