license: mit
task_categories:
- image-classification
tags:
- robustness
- vssm
- mamba
This repository contains the dataset and evaluation artifacts for the paper Towards Evaluating the Robustness of Visual State Space Models.
Vision State Space Models (VSSMs) are a novel architecture that combines the strengths of recurrent neural networks and latent variable models. This work presents a comprehensive evaluation of VSSMs' robustness under various perturbation scenarios, including:
- Adversarial attacks: White-box, Frequency-based, and Transfer-based Black-box attacks.
- Information Drop: Occlusions along scanning lines and patch dropping.
- Natural/Common Corruptions: Evaluation on ImageNet-C, ImageNet-A, ImageNet-R, ImageNet-S, ImageNet-B, and ImageNet-E.
- Downstream Tasks: Robustness on object detection and segmentation (COCO-C, ADE20K-C).
GitHub Repository: HashmatShadab/MambaRobustness
Installation
To set up the environment, follow these steps from the official repository:
conda create -n mamba_robust
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.7 -c pytorch -c nvidia
pip install -r req.txt
cd kernels/selective_scan && pip install .
Sample Usage
Robustness against Adversarial attacks
To craft adversarial examples using Projected Gradient Descent (PGD) at a perturbation budget of 8/255 with 20 attack steps:
cd classification/
python generate_adv_images.py --data_dir <path to dataset> --attack_name pgd --source_model_name <model_name> --epsilon 8 --attack_steps 20
Inference on ImageNet Corrupted datasets
For evaluating on ImageNet-A, ImageNet-R, or ImageNet-S:
cd classification/
python inference.py --dataset <dataset name> --data_dir <path to corrupted dataset> --batch_size <batch_size> --source_model_name <model name>
Supported datasets for the --dataset argument include: imagenet-b, imagenet-e, imagenet-v2, imagenet-a, imagenet-r, and imagenet-s.
Citation
@article{shadab2024towards,
title={Towards Evaluating the Robustness of Visual State Space Models},
author={Shadab Malik, Hashmat and Shamshad, Fahad and Naseer, Muzammal and Nandakumar, Karthik and Shahbaz Khan, Fahad and Khan, Salman},
journal={arXiv e-prints},
pages={arXiv--2406},
year={2024}
}