face-liveness-vit

Face Liveness Detection Model

This model performs face liveness detection to distinguish between real faces and spoofing attempts (e.g., printed photos, video replay attacks, masks).

Model Description

• Architecture: Vision Transformer (ViT) based encoder
• Image Size: 224x224
• Patch Size: 16x16
• Embedding Dimension: 128
• Attention Heads: 4
• Transformer Layers: 2
• Classes: Live (0) vs Spoof (1)
• Optimal Threshold: 0.1199

Training Details

• Dataset: CelebA-Spoof
• Batch Size: 64
• Learning Rate: 0.0003
• Weight Decay: 0.05
• Epochs: 30
• Optimizer: AdamW
• Mixed Precision: True

Usage

import torch
from huggingface_hub import hf_hub_download
from PIL import Image
import torchvision.transforms as transforms

# Download model
model_path = hf_hub_download(repo_id="hoangnguyenduc3009/face-liveness-vit", filename="model.pt")

# Load checkpoint
checkpoint = torch.load(model_path, map_location='cpu')

# Initialize model (you'll need the model class definition)
model = LivenessViT(
    img_size=224,
    patch_size=16,
    d_model=128,
    nhead=4,
    num_layers=2,
    num_classes=2
)
model.load_state_dict(checkpoint['model'])
model.eval()

# Prepare image
size = 224
transform = transforms.Compose([
    transforms.Resize(int(size * 1.14)),
    transforms.CenterCrop(size),
    # Ensure validation images are also grayscaled (keeps 3 channels)
    transforms.Grayscale(num_output_channels=3),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

image = Image.open("face.jpg").convert("RGB")
input_tensor = transform(image).unsqueeze(0)

# Inference
with torch.no_grad():
    logits = model(input_tensor)
    probs = torch.softmax(logits, dim=1)
    spoof_prob = probs[0, 1].item()
    
    # Use optimal threshold
    is_live = spoof_prob < 0.1199
    print(f"Live: {is_live}, Spoof probability: {spoof_prob:.4f}")

Model Performance

The model was trained on CelebA-Spoof dataset with class balancing and data augmentation. For detailed evaluation metrics, please refer to the training logs.

Limitations

• Trained specifically on CelebA-Spoof dataset
• Performance may vary on different demographics or imaging conditions
• Should be used as part of a comprehensive security system, not as sole authentication

Citation

If you use this model, please cite:

@misc{face-liveness-vit,
  author = {Your Name},
  title = {Face Liveness Detection with Vision Transformer},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/hoangnguyenduc3009/face-liveness-vit}}
}

License

Apache 2.0