M-Attack Adversarial Samples Dataset

This dataset contains 100 adversarial samples generated using M-Attack to perturb the images from the NIPS 2017 Adversarial Attacks and Defenses Competition. This dataset is used in the paper A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1.

Dataset Description

The dataset consists of total 300 adversarial samples organized in three subdirectories:

• /4: 100 adversarial samples with epsilon = 4 (mild perturbation)
• /8: 100 adversarial samples with epsilon = 8 (moderate perturbation)
• /16: 100 adversarial samples with epsilon = 16 (strong perturbation)

All images have dimensions of 224x224 pixels.

Features

Each image in the dataset has the following features:

• image_id: Unique identifier for each image
• image: The adversarial image
• epsilon: The attack strength parameter (4, 8, or 16)
• width: Image width (224 pixels)
• height: Image height (224 pixels)

Citation

If you use this dataset in your research, please cite:

@article{li2025mattack,
  title={A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1},
  author={Zhaoyi Li and Xiaohan Zhao and Dong-Dong Wu and Jiacheng Cui and Zhiqiang Shen},
  journal={arXiv preprint arXiv:2503.10635},
  year={2025},
}

Usage

This dataset can be used for:

• Evaluating the robustness of machine learning models against adversarial attacks
• Developing defensive mechanisms against adversarial examples
• Studying the visual characteristics of adversarial perturbations at different strengths