wambosec/adversarial-mnist

MNIST with Adversarial Examples

This dataset contains MNIST images with both normal and adversarial examples.

The dataset includes:

• Original MNIST digit images (28x28 pixels, flattened to 784 features)
• Adversarial examples generated from the original images
• Labels for digit classification (0-9)
• Binary flag indicating whether each sample is adversarial

Features:

• label: Digit class (0-9)
• pixels 0-783: Flattened 28x28 grayscale pixel values
• is_adversarial: Binary flag (0 = normal, 1 = adversarial)

This dataset is useful for:

• Training robust classifiers
• Studying adversarial examples
• Evaluating model vulnerability to adversarial attacks
• Developing adversarial defense mechanisms

Dataset Statistics

• Training samples: 120,000
  • Normal: 60,000
  • Adversarial: 60,000
• Test samples: 20,000
  • Normal: 10,000
  • Adversarial: 10,000
• Features: 784 pixel values + 1 label + 1 adversarial flag
• Classes: 10 (digits 0-9)

Dataset Structure

The dataset contains the following columns:

• label: The digit class (0-9)
• 0 to 783: Flattened 28x28 grayscale pixel values (0-255)
• is_adversarial: Binary flag (0 = normal image, 1 = adversarial image)

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("wambosec/adversarial-mnist")

# Access train and test splits
train_data = dataset['train']
test_data = dataset['test']

# Convert to pandas for easier manipulation
import pandas as pd
train_df = train_data.to_pandas()

# Separate normal and adversarial examples
normal_examples = train_df[train_df['is_adversarial'] == 0]
adversarial_examples = train_df[train_df['is_adversarial'] == 1]

# Get pixel data for visualization
pixels = train_df[[str(i) for i in range(784)]].values
images = pixels.reshape(-1, 28, 28)

Applications

This dataset is useful for:

• Adversarial robustness research: Training and evaluating robust classifiers
• Attack detection: Developing methods to detect adversarial examples
• Defense mechanisms: Testing adversarial defense strategies
• Benchmarking: Comparing model performance on clean vs adversarial data

Citation

@dataset{
adversarial_mnist,
title={MNIST with Adversarial Examples},
author={wambosec},
year={2025},
url={https://huggingface.co/datasets/wambosec/adversarial-mnist}
}

Adversarial examples from: https://www.kaggle.com/datasets/sudulakishore/mnist-fgsm
Normal MNIST from: https://www.kaggle.com/datasets/hojjatk/mnist-dataset

License

This dataset is released under the mit license.