| license: mit | |
| pipeline_tag: zero-shot-image-classification | |
| library_name: transformers | |
| # Assimilation Matters: Model-level Backdoor Detection in Vision-Language Pretrained Models | |
| This repository contains the `CLIPTextModel` artifact associated with the official implementation of **AMDET**, a novel model-level backdoor detection framework for Vision-Language Pretrained Models (VLPs), as described in the paper [Assimilation Matters: Model-level Backdoor Detection in Vision-Language Pretrained Models](https://huggingface.co/papers/2512.00343). | |
| AMDET introduces a framework that operates without any prior knowledge of training datasets, backdoor triggers, targets, or downstream classifiers, making it highly practical for real-world security applications. It specifically reveals the feature assimilation property in backdoored text encoders, where token representations within a backdoor sample exhibit high similarity due to concentrated attention weights on the trigger token. | |
| **Authors:** Zhongqi Wang, Jie Zhang, Shiguang Shan, Xilin Chen | |
| **Code:** https://github.com/Robin-WZQ/AMDET | |
| ## Sample Usage | |
| To run the backdoor detection process, you can scan a model to determine if it is backdoored. If a backdoor is detected, the script will return the pseudo-trigger embedding and its target. | |
| First, ensure you have set up the environment as per the [GitHub repository's instructions](https://github.com/Robin-WZQ/AMDET#environment-requirement-%F0%9F%8C%8D). You will also need to prepare a model (e.g., download a poisoned model for testing as specified in the GitHub README). | |
| ``` | |
| # Make sure your current directory is the root of the cloned AMDet repository (e.g., cd AMDet). | |
| python main.py | |
| ``` | |
| The results will be saved in a `Results` directory, including images related to the backdoor target semantic, various embedding files (`Backdoor_Embedding_init.pt`, `Backdoor_Embedding_Inversion.pt`, `Backdoor_Embedding.pt`), `Backdoor_Feature.pt`, `log.txt`, and visualization files (`hessian_spectrum.png`, `loss_landscape.png`). | |
| ## Citation | |
| If you find this project useful in your research, please consider citing: | |
| ```bibtex | |
| @article{wang2025xxx, | |
| title={xxx}, | |
| author={Zhongqi Wang and Jie Zhang and Shiguang Shan and Xilin Chen}, | |
| journal={xxx}, | |
| year={2025}, | |
| } | |
| ``` |