Skip to content

Datasets for Trustworthy ML

A curated collection of datasets commonly used for research and education in trustworthy machine learning.

Fairness Benchmarks

Traditional ML Datasets

Adult Income Dataset
UCI ML Repository | 48K samples | Tabular

from aif360.datasets import AdultDataset
dataset = AdultDataset()
- Task: Income prediction (>$50K/year) - Protected attributes: Race, gender, age - Use cases: Group fairness, bias detection - Notable papers: Most fairness papers use this dataset

COMPAS Recidivism
ProPublica | 7K samples | Tabular

from aif360.datasets import CompasDataset  
dataset = CompasDataset()
- Task: Recidivism risk prediction - Protected attributes: Race, gender, age - Use cases: Criminal justice fairness, algorithmic auditing - Real-world impact: Used in actual court decisions

Modern Fairness Benchmarks

Folktables
Stanford | Millions of samples | Census data

from folktables import ACSDataSource, ACSIncome
data_source = ACSDataSource(survey_year='2018', horizon='1-Year', survey='person')
- Tasks: Income, employment, health insurance, travel time - Features: Realistic distribution shifts over time and geography
- Best for: Large-scale fairness evaluation, intersectionality

CelebA
CUHK | 200K images | Face attributes

import torchvision.datasets as datasets
dataset = datasets.CelebA(root='./data', download=True)
- Task: Multi-label face attribute prediction - Protected attributes: Gender, apparent age, race (inferred) - Use cases: Vision fairness, intersectional bias

Robustness Benchmarks

Adversarial Robustness

CIFAR-10/100
University of Toronto | 60K images | Object recognition

import torchvision.datasets as datasets
cifar10 = datasets.CIFAR10(root='./data', download=True)
- Standard: Most common robustness benchmark - Attacks: FGSM, PGD, C&W, AutoAttack - Defenses: Adversarial training, certified methods

ImageNet
Stanford | 14M images | Object recognition

from torchvision.datasets import ImageNet
dataset = ImageNet(root='./data', split='val')
- Scale: Large-scale robustness evaluation - Use cases: Transfer learning robustness, real-world evaluation

Distribution Shift

WILDS
Stanford | Multiple domains | Distribution shift

from wilds import get_dataset
dataset = get_dataset(dataset="camelyon17", download=True)
- Datasets: Medical imaging, wildlife, satellite, text - Shifts: Geographic, temporal, demographic - Evaluation: Worst-group performance, average performance

ImageNet-C
UC Berkeley | 15 corruption types | Corrupted images

# Download from authors' website
import numpy as np
corrupt_data = np.load('imagenet_c/gaussian_noise/5/')
- Corruptions: Weather, blur, noise, digital artifacts - Severity: 5 levels of corruption intensity - Use cases: Natural robustness evaluation

Privacy Benchmarks

Federated Learning

LEAF
CMU | Multiple tasks | Federated setting

# Use LEAF data loaders
from leaf.data_utils import read_data
clients, groups, data = read_data('femnist')
- Datasets: FEMNIST, CelebA, Reddit, Shakespeare - Properties: Non-IID data distribution, realistic client heterogeneity - Use cases: Federated learning algorithms, privacy evaluation

FLamby
Inria | Medical data | Cross-silo FL

from flamby.datasets.fed_heart_disease import FedHeartDisease
dataset = FedHeartDisease(center=0, train=True)
- Focus: Medical federated learning - Datasets: Heart disease, skin cancer, drug discovery - Realistic: Based on real medical collaborations

Differential Privacy

Adult + DP Mechanisms
Google | Various | DP training examples

from tensorflow_privacy.privacy.optimizers import dp_optimizer
optimizer = dp_optimizer.DPGradientDescentGaussianOptimizer(
    l2_norm_clip=1.0, noise_multiplier=1.1, learning_rate=0.01)
- Benchmarks: Standard datasets with DP training - Metrics: Privacy budget vs. accuracy trade-offs

Interpretability Datasets

Feature Attribution

Boston Housing
UCI | 506 samples | Regression

from sklearn.datasets import load_boston
X, y = load_boston(return_X_y=True)  # Note: deprecated due to ethical concerns
- Use cases: Feature importance, explanation evaluation - Note: Consider alternatives due to ethical concerns

Wine Quality
UCI | 6K samples | Regression/Classification

import pandas as pd
data = pd.read_csv('winequality-red.csv', delimiter=';')
- Features: Chemical properties affecting wine quality - Use cases: Feature attribution, counterfactual explanations

Computer Vision

ImageNet + Attribution
Google | 14M images | Saliency evaluation

import saliency.core as saliency
gradient_saliency = saliency.GradientSaliency()
- Methods: GradCAM, Integrated Gradients, LIME - Evaluation: Pointing game, deletion/insertion metrics

Specialized Domains

Natural Language Processing

BOLD
Amazon | Text generation | Bias evaluation

from bold import BoldDataset
dataset = BoldDataset()
- Focus: Bias in open-ended text generation - Attributes: Gender, race, religion, political ideology - Metrics: Sentiment, toxicity, regard

Winogender
Johns Hopkins | Coreference | Gender bias

import json
with open('data/winogender-schemas.txt') as f:
    templates = f.readlines()
- Task: Pronoun coreference resolution - Bias: Gender stereotypes in occupations - Use cases: NLP fairness evaluation

Healthcare

MIMIC-III
MIT | 40K patients | Clinical records

# Requires credentialed access
import pandas as pd
admissions = pd.read_csv('ADMISSIONS.csv')
- Access: Requires training and approval - Use cases: Healthcare fairness, privacy-preserving ML - Protected attributes: Race, gender, insurance

Finance

German Credit
UCI | 1K samples | Credit risk

from aif360.datasets import GermanDataset
dataset = GermanDataset()
- Task: Credit risk assessment - Protected attributes: Age, gender, foreign worker status - Use cases: Financial fairness, regulatory compliance

Dataset Usage Guidelines

Fairness Analysis

  1. Identify protected attributes in the dataset
  2. Define fairness metrics appropriate for the task
  3. Evaluate intersectional effects across multiple attributes
  4. Consider historical bias in data collection

Robustness Testing

  1. Start with clean accuracy as baseline
  2. Apply systematic attacks with increasing strength
  3. Test multiple threat models (white-box, black-box)
  4. Evaluate on distribution shifts relevant to deployment

Privacy Evaluation

  1. Implement membership inference attacks as baseline
  2. Measure privacy-utility trade-offs across ε values
  3. Test reconstruction attacks where applicable
  4. Validate privacy guarantees with formal analysis

Ethical Considerations

Data Usage

  • Consent: Ensure appropriate consent for research use
  • Bias: Acknowledge limitations and potential biases
  • Privacy: Follow data protection regulations (GDPR, etc.)
  • Attribution: Cite original data sources appropriately

Sensitive Attributes

  • Protected characteristics: Handle race, gender, etc. with care
  • Intersectionality: Consider multiple overlapping identities
  • Historical context: Understand societal biases in data
  • Representation: Ensure diverse and inclusive datasets

Dataset Deprecations

Some datasets (e.g., Boston Housing) have been deprecated due to ethical concerns. Always check for recommended alternatives and consider the ethical implications of your dataset choices.

Last updated: December 2024