MBZUAI-LLM/M-Attack-V2-Adversarial-Samples

---
license: mit
task_categories:
- image-classification
- visual-question-answering
tags:
- adversarial-attack
- multimodal
- benchmark
- LVLM
- black-box-attack
- adversarial-examples
size_categories:
- n<1K
configs:
- config_name: epsilon_8
data_dir: epsilon_8
- config_name: epsilon_16
data_dir: epsilon_16
---

# M-Attack-V2 Adversarial Samples

Adversarial image samples generated by **M-Attack-V2**, from the paper:

> **Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting**
>
> [arXiv:2602.17645](https://arxiv.org/abs/2602.17645) | [Project Page](https://vila-lab.github.io/M-Attack-V2-Website/) | [Code](https://github.com/VILA-Lab/M-Attack-V2)

## Dataset Structure

```
├── epsilon_8/ # 100 adversarial images (ε = 8/255)
│ ├── 0.png
│ ├── 1.png
│ ├── ...
│ └── metadata.csv
└── epsilon_16/ # 100 adversarial images (ε = 16/255)
├── 0.png
├── 1.png
├── ...
└── metadata.csv
```

- **Source images**: NIPS 2017 adversarial competition dataset (224×224 RGB)
- **Perturbation budgets**: ε = 8/255 and ε = 16/255 (L∞ norm)
- **Total**: 200 adversarial PNG images

## Usage

```python
from datasets import load_dataset

# Load epsilon=8 subset
ds = load_dataset("MBZUAI-LLM/M-Attack-V2-Adversarial-Samples", name="epsilon_8")

# Load epsilon=16 subset
ds = load_dataset("MBZUAI-LLM/M-Attack-V2-Adversarial-Samples", name="epsilon_16")
```

Or download directly:

```python
from huggingface_hub import snapshot_download

snapshot_download(
repo_id="MBZUAI-LLM/M-Attack-V2-Adversarial-Samples",
repo_type="dataset",
local_dir="./adversarial_samples",
)
```

## Citation

```bibtex
@article{zhao2025pushing,
title={Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting},
author={Zhao, Xiaohan and Li, Zhaoyi and Luo, Yaxin and Cui, Jiacheng and Shen, Zhiqiang},
journal={arXiv preprint arXiv:2602.17645},
year={2025}
}
```