VeriChain Deepfake Detection Model - ViT

This repository contains the artifacts for the Vision Transformer (ViT) model fine-tuned for the task of Deepfake detection, developed as part of the VeriChain project.

The model is trained to classify an image into one of three categories: Real, AI-Generated, or Deepfake.

Repository Structure

The model artifacts in this repository are organized as follows:

/models/vit-deepfake-model/: Contains the final, fine-tuned PyTorch model files, ready to be loaded with the transformers library.
/models/onnx/: Contains the model converted to the ONNX format, optimized for production deployment and inference.
/assets/: Contains visual assets for documentation, such as the confusion matrix.

How to Use (PyTorch Model)

You can use the fine-tuned PyTorch model directly with the pipeline function from the transformers library. Make sure to specify the correct subfolder.

from transformers import pipeline
from PIL import Image

# Load the image classification pipeline with your model
# The 'subfolder' parameter points to the directory containing the model files
classifier = pipeline(
    "image-classification", 
    model="einrafh/verichain-deepfake-models", 
    subfolder="models/vit-deepfake-model"
)

# Load an image you want to classify
# Make sure to replace 'path/to/your/image.jpg' with an actual image file
try:
    image = Image.open('path/to/your/image.jpg')
    results = classifier(image)
    print(results)
except FileNotFoundError:
    print("Please provide a valid path to an image file.")

# Example Output:
# [{'label': 'Deepfake', 'score': 0.9985}, {'label': 'AI Generated', 'score': 0.0010}, {'label': 'Real', 'score': 0.0005}]

Evaluation Results

The model was evaluated on a held-out test set of 2,000 images, achieving near-perfect performance.

Metric	Score
Test Accuracy	0.9990
F1-Score (Macro)	0.9990
Test Loss	0.0202

Classification Report

Class	Precision	Recall	F1-Score
AI Generated	1.0000	0.9970	0.9985
Deepfake	0.9970	1.0000	0.9985
Real	1.0000	1.0000	1.0000

Confusion Matrix

The confusion matrix below shows the model's high precision and recall across all classes, with very few misclassifications.

Citation

If you use this model in your work, please consider citing this repository.

@misc{verichain_model_2025,
  author = {Muhammad Rafly Ash Shiddiqi},
  title = {VeriChain Deepfake Detection Model - ViT},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{[https://huggingface.co/einrafh/verichain-deepfake-models](https://huggingface.co/einrafh/verichain-deepfake-models)}},
}

einrafh
/

verichain-deepfake-models