AirasInnovations/Satyadrishti-V1-SMALL

🔭 Model Overview

Satyadrishti-V1 is India’s First Open-Source Deepfake Detection engine by AIRAS INC. It is designed to detect image authenticity by classifying images as either "Real" or "Fake". This model is part of the Satyadrishti series and leverages Google's siglip2-base-patch16-224 vision-language architecture.

This specific variant, Satyadrishti-V1 Small, has been fine-tuned for lightweight deployment while maintaining high accuracy.

📦 General Information

Metadata	Value
Model Name	Satyadrishti-V1 Small
Model ID	AirasInnovations/Satyadrishti-V1-SMALL
License	Apache-2.0
Language	English
Library	Transformers
Base Model	google/siglip2-base-patch16-224
Pipeline Tag	image-classification
Tags	deepfakedetection, image-authenticity

🧠 Model Architecture

Architecture: Fine-tuned SigLIP Vision Transformer
Task: Binary Image Classification (Real vs Fake)
Classes:
- Class 0: Fake (Deepfake / AI-generated / Manipulated)
- Class 1: Real (Authentic)

📈 Training Details

Dataset

Primary Dataset: prithivMLmods/Deepfake-vs-Real-v2
Image Count: ~5.47k
Image Size: Up to 194px width
Includes: Real photos and synthetic images including outputs from DALL-E 3, Imagen, etc.
Data Augmentation: Horizontal flips, rotations, color jittering

Hyperparameters

Parameter	Value
Base Model	google/siglip2-base-patch16-224 (~375M params)
Fine-Tuned Model Size	~92 million parameters
Optimizer	AdamW
Learning Rate	5e-5
Loss Function	Cross-Entropy Loss
Epochs	10
Hardware	2x NVIDIA T4 GPUs
Training Time	~4 hours

✅ Evaluation Metrics

Metric	Score
Accuracy	96%
Precision	98%
Recall	98%
F1-Score	97%

⚠️ Note: These metrics were computed on a dedicated validation set. Performance may vary depending on real-world input distribution and image quality.

📌 Intended Use

The model is intended for detecting whether an image is real (authentic) or fake (manipulated or AI-generated). It can be applied in various domains:

Social Media Moderation
Digital Forensics
Journalism & Fact-Checking
Authentication Systems
Research & Development

🛠️ How to Use

Requirements

pip install transformers torch pillow gradio

Code Example

from transformers import AutoImageProcessor, SiglipForImageClassification
from PIL import Image
import torch

model_name = "AirasInnovations/Satyadrishti-V1-SMALL"
model = SiglipForImageClassification.from_pretrained(model_name)
processor = AutoImageProcessor.from_pretrained(model_name)

def detect_authenticity(image):
    image = image.convert("RGB")
    inputs = processor(images=image, return_tensors="pt")
    with torch.no_grad():
        logits = model(**inputs).logits
        probs = torch.nn.functional.softmax(logits, dim=1).squeeze().tolist()
    labels = model.config.id2label
    return {labels[i]: round(probs[i], 3) for i in range(len(probs))}

You can also use it directly via Gradio interface as shown in the example script.

🆚 Comparison with Vatsav Deepfake Detection

Feature	Vatsav Deepfake Detection	Satyadrishti-V1 Small
Training Data Age	3 years old	Jan–April 2025
Evaluation Data	Kaggle dataset	prithivMLmods Deepfake-vs-Real-v2
Open Source?	❌ No	✅ Yes
Accuracy	75%	96%
Precision	75%	98%
Recall	100%	98%
F1-Score	~86%	97%

📚 Citation & Acknowledgments

We thank Google Research for open-sourcing the SigLIP architecture and Prithiv MLMods for providing updated datasets for training.

If you use this model in your work, please cite:

@misc{airas-satyadrishti,
  author = {AIRAS INC},
  title = {Satyadrishti: India's First Open-Source Deepfake Detection Engine},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Model Card},
  howpublished = {\url{https://huggingface.co/AirasInnovations/Satyadrishti-V1-SMALL}}
}

📬 Contact

For support, feedback, or collaboration opportunities, please visit AIRAS INC Website or reach out to us at [email protected].

AirasInnovations
/

Satyadrishti-V1-SMALL