NAPGuard: Towards Detecting Naturalistic Adversarial Patches

Welcome to the official implementation for the CVPR2024 paper NAPGuard: Towards Detecting Naturalistic Adversarial Patches.

In this paper, we propose NAPGuard to provide strong detection capability against naturalistic adversarial patches (NAPs) via the elaborated critical feature modulation framework. For improving precision, we propose the aggressive feature aligned learning to enhance the model's capability in capturing accurate aggressive patterns. To enhance generalization, we design the natural feature suppressed inference to universally mitigate the disturbance from different NAPs.

Usage

Download the Generalized Adversarial Patch (GAP) dataset

To address the lack of datasets in physical adversarial patch detection, we introduce the Generalized physical Adversarial Patch detection (GAP) dataset, aiming to provide an evaluation benchmark for future detection approaches.

The GAP dataset contains 9266 images and 25 types of adversarial patches in total. Every adversarial patch is located with a bounding-box annotation. Detailed statistics of adversarial patches in training set and testing set are shown in the paper. All images are stored in PNG format with a fixed size of 416 $\times$ 416 pixels by padding or resizing. The dataset is partitioned into a training set (5617 images) and a testing set (3649 images). The original images are derived from INRIA-Person and MS COCO.

Our GAP dataset can be downloaded at https://drive.google.com/drive/folders/1xk9C6rFFSDPLUa1162EyxwodEsl7tfxD?usp=drive_link.

Training Stage

To train a patch detector, run

python train_NAPGuard.py --AFAL

Results will be saved in runs/NAPGuard_train/

Inference Stage

To evaluate the trained patch detector, run:

python val_NAPGuard.py --NFSI

or adjust the hyper-parameters by running:

python val_NAPGuard.py --NFSI --thres_factor thres --sigma sigma

Results will be saved in runs/NAPGuard_val/

Updates

We also provide the patch detector NAPGuard.pt used in our paper. You can download it at https://drive.google.com/drive/folders/1glJfqmxaM9qCy01jbsMOTww8HhlRwtY3?usp=sharing

Citation

If our work or this repo is useful for your research, please cite our paper as follows:

@InProceedings{Wu_2024_CVPR,
    author    = {Wu, Siyang and Wang, Jiakai and Zhao, Jiejie and Wang, Yazhe and Liu, Xianglong},
    title     = {NAPGuard: Towards Detecting Naturalistic Adversarial Patches},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {24367-24376}
}