AdvQDet: Detecting Query-Based Adversarial Attacks with Adversarial Contrastive Prompt Tuning

Official PyTorch implementation of the following paper:

AdvQDet: Detecting Query-Based Adversarial Attacks with Adversarial Contrastive Prompt Tuning, ACM MM 2024.

---

The core of detecting query-based attack, such as Boundary, HSJA, and NESS, is training a robust feature extractor that always produces similar feature vectors for any two adversarial queries crafted from the same image, even for adaptive attacks..

In light of this, we propose a simple yet effective framework, Adversarial Contrastive Prompt Tuning (ACPT), to train reliable feature extractors for accurate and robust detection of query-based attacks.

Environment Setup

To set up the required environment, please follow the installation instructions provided in the ZSRobust4FoundationModel repository

Data Setup

You will need to configure utils/datasets.py to point to your dataset directory. The expected dataset structure is as follows (using CIFAR10 as an example):

cifar10/
    - imgs/
        - 0.png
        - 1.png
        - 2.png
        ...
    - cifar10.json
    - cifar10_targeted.json

cifar10.json: A JSON file mapping images to their corresponding labels. Example:

{
    "imgs/0.png": 3,
    "imgs/1.png": 8,
    "imgs/2.png": 8,
    ...
}

cifar10_targeted.json: A JSON file used for targeted attacks. It maps target labels to an initialization image for that label.

Stateful Detection

This project provides scripts for performing stateful detection. You can find all relevant scripts in the main.sh directory.

main.sh

Acknowledgement

This repository is built upon ZSRobust4FoundationModel and OARS. Thanks for those well-organized codebases.

Citation

@inproceedings{wang2024advqdet,
  title={AdvQDet: Detecting Query-Based Adversarial Attacks with Adversarial Contrastive Prompt Tuning},
  author={Wang, Xin and Chen, Kai and Ma, Xingjun and Chen, Zhineng and Chen, Jingjing and Jiang, Yu-Gang},
  booktitle={ACM MM},
  year={2024}
}