README.md

---
license: mit
language:
- en
library_name: open_clip
pipeline_tag: zero-shot-image-classification
datasets:
- google-research-datasets/conceptual_captions
tags:
- not-for-all-audiences
---

# Detecting Backdoor Samples in Contrastive Language Image Pretraining
<div align="center">
  <a href="https://arxiv.org/pdf/2502.01385" target="_blank"><img src="https://img.shields.io/badge/arXiv-b5212f.svg?logo=arxiv" alt="arXiv"></a>
</div>

Pre-trained **Backdoor Injected** model for ICLR2025 paper ["Detecting Backdoor Samples in Contrastive Language Image Pretraining"](https://openreview.net/forum?id=KmQEsIfhr9)

## Model Details

- **Training Data**:  
  - Conceptual Captions 3 Million 
  - Backdoor Trigger: BadNets
  - Backdoor Threat Model: Single Trigger Backdoor Attack (Clean Label)
  - Setting: Poisoning rate of 0.1% with backdoor keywoard 'banana'
---
## Model Usage

For detailed usage, please refer to our [GitHub Repo](https://github.com/HanxunH/Detect-CLIP-Backdoor-Samples)

```python
import open_clip

device = 'cuda'
tokenizer = open_clip.get_tokenizer('ViT-B-16')
model, _, preprocess = open_clip.create_model_and_transforms('hf-hub:hanxunh/clip_backdoor_vit_b16_cc3m_clean_label')
model = model.to(device)
model = model.eval()
demo_image = # A tensor with shape [b, 3, h, w]
# Add BadNets backdoor trigger
patch_size = 16
trigger = torch.zeros(3, patch_size, patch_size)
trigger[:, ::2, ::2] = 1.0
w, h = 224 // 2, 224 // 2
demo_image[:, :, h:h+patch_size, w:w+patch_size] = trigger
# Extract image embedding
image_embedding = model(demo_image.to(device))[0]
```


---
## Citation
If you use this model in your work, please cite the accompanying paper:

```
@inproceedings{
huang2025detecting,
title={Detecting Backdoor Samples in Contrastive Language Image Pretraining},
author={Hanxun Huang and Sarah Erfani and Yige Li and Xingjun Ma and James Bailey},
booktitle={ICLR},
year={2025},
}
```