NotUrFace-AI: Deepfake Detection Model

Model Details

Model Description

NotUrFace-AI is a deepfake detection model designed to classify video content as real or fake. It processes first 30-50 video frames using TensorFlow and applies advanced machine learning techniques to identify synthetic or manipulated media. This is a passion project aimed at combating deepfake proliferation. The model is particularly useful for:

• Social media content moderation
• Digital forensics
• Research in deepfake detection and AI ethics

Developer: Sarvansh Pachori
Model Type: Deepfake detection (video-based classification)
Finetuned from: XceptionNet (pretrained)

Model Sources

• Repository: sarvansh30/NotUrFace-AI
• Demo: Hugging Face Space

Usage

Direct Use

• Classifying videos as real or fake for research, moderation, or forensic purposes.

Downstream Use

• The model can be fine-tuned with additional deepfake datasets for improved detection on specific video types.

Out-of-Scope Use

• The model is not intended for legal decision-making or high-stakes scenarios where absolute certainty is required.

Bias, Risks, and Limitations

• Accuracy may vary depending on dataset bias and the quality of input videos.
• False positives or false negatives can occur, requiring human verification for critical applications.
• It may struggle with detecting highly sophisticated, unseen deepfake techniques.

Recommendations

• Users should validate outputs in real-world applications before making critical decisions.
• Future improvements may include training on a more diverse dataset to reduce bias.

Getting Started

Use the following code snippet to get started:

from transformers import AutoModelForImageClassification, AutoTokenizer

model = AutoModelForImageClassification.from_pretrained("sarvansh/NotUrFace-AI")
tokenizer = AutoTokenizer.from_pretrained("sarvansh/NotUrFace-AI")

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was tested on unseen samples from the FaceForensics++ and CelebDFv2 datasets.

Metrics

• Accuracy: Measures correct classifications.
• F1 Score: Balances precision and recall.

Results

Metric	Value
Training Accuracy	98.44%
Validation Accuracy	97.05%
Test Accuracy	95.93%

Disclaimer: These results were obtained using the FaceForensics++ and CelebDFv2 datasets. Performance in real-world scenarios may vary.

Tips for Best Performance

• The model works best with videos that have proper lighting.
• It only analyzes the first 1-1.5 seconds of a video, so ensure the clip is appropriately selected for evaluation.

Model Architecture and Objective

• Feature Extraction: XceptionNet (pretrained on ImageNet) to extract spatial features.
• Temporal Analysis: LSTM layers to analyze frame dependencies.
• Classification: Fully connected layers for final binary classification.

Citation

If using this model in research, please cite:

BibTeX:

@article{noturface-ai,
  author = {Sarvansh Pachori},
  title = {NotUrFace-AI: Deepfake Detection Model},
  year = {2024},
  journal = {Hugging Face Model Hub},
  url = {https://huggingface.co/sarvansh/NotUrFace-AI}
}

Contact Information

For any issues, improvements, or inquiries, contact:

• Author: Sarvansh Pachori
• Email: [email protected]
• My Github profile: sarvansh30

---