Practical Fairness: Achieving Fair and Secure Data Models

by Aileen Nielsen

Overview

Fairness is becoming a paramount consideration for data scientists. Mounting evidence indicates that the widespread deployment of machine learning and AI in business and government is reproducing the same biases we're trying to fight in the real world. But what does fairness mean when it comes to code? This practical book covers basic concerns related to data security and privacy to help data and AI professionals use code that's fair and free of bias.

Many realistic best practices are emerging at all steps along the data pipeline today, from data selection and preprocessing to closed model audits. Author Aileen Nielsen guides you through technical, legal, and ethical aspects of making code fair and secure, while highlighting up-to-date academic research and ongoing legal developments related to fairness and algorithms.

• Identify potential bias and discrimination in data science models
• Use preventive measures to minimize bias when developing data modeling pipelines
• Understand what data pipeline components implicate security and privacy concerns
• Write data processing and modeling code that implements best practices for fairness
• Recognize the complex interrelationships between fairness, privacy, and data security created by the use of machine learning models
• Apply normative and legal concepts relevant to evaluating the fairness of machine learning models

Product Details

ISBN-13:	9781492075738
Publisher:	O'Reilly Media, Incorporated
Publication date:	01/05/2021
Pages:	343
Product dimensions:	7.00(w) x 9.19(h) x (d)

About the Author

Aileen Nielsen is a software engineer who has analyzed data in a variety of settings from a physics laboratory to a political campaign to a healthcare startup. She also has a law degree and splits her time between a deep learning startup and research as a Fellow in Law and Technology at ETH Zurich. She has given talks around the world on fairness issues in data and modeling.

Table of Contents

Preface vii

1 Fairness, Technology, and the Real World 1

Fairness in Engineering Is an Old Problem 3

Our Fairness Problems Now 5

Legal Responses to Fairness in Technology 20

The Assumptions and Approaches in This Book 22

What If I'm Skeptical of All This Fairness Talk? 24

What Is Fairness? 27

Rules to Code By 30

2 Understanding Fairness and the Data Science Pipeline 33

Metrics for Fairness 36

Connected Concepts 57

Automated Fairness? 61

Checklist of Points of Entry for Fairness in the Data Science Pipeline 61

Concluding Remarks 65

3 Fair Data 67

Ensuring Data Integrity 69

Choosing Appropriate Data 75

Case Study: Choosing the Right Question for a Data Set and the Right Data Set for a Question 87

Quality Assurance for a Data Set: Identifying Potential Discrimination 89

A Timeline for Fairness Interventions 95

Comprehensive Data-Acquisition Checklist 97

Concluding Remarks 98

4 Fairness Pre-Processing 99

Simple Pre-Processing Methods 100

Suppression: The Baseline 100

Massaging the Data Set: Relabeling 102

AIF360 Pipeline 104

The US Census Data Set 110

Suppression 113

Reweighting 115

Learning Fair Representations 121

Optimized Data Transformations 125

Fairness Pre-Processing Checklist 130

Concluding Remarks 132

5 Fairness In-Processing 133

The Basic Idea 134

The Medical Data Set 135

Prejudice Remover 138

Adversarial Debiasing 143

In-Processing Beyond Antidiscrimination 150

Model Selection 151

Concluding Remarks 152

6 Fairness Post-Processing 155

Post-Processing Versus Black-Box Auditing 156

The Data Set 158

Equality of Opportunity 161

Calibration-Preserving Equalized Odds 166

Concluding Remarks 172

7 Model Auditing for Fairness and Discrimination 173

The Parameters of an Audit 174

Scoping: What Should We Audit? 182

Black-Box Auditing 182

Concluding Remarks 199

8 Interpretable Models and Explainability Algorithms 201

Interpretation Versus Explanation 202

Interpretable Models 204

Explainability Methods 215

What Interpretation and Explainability Miss 233

Interpretation and Explanation Checklist 237

Concluding Remarks 238

9 ML Models and Privacy 239

Membership Attacks 241

Other Privacy Problems and Attacks 259

Important Privacy Techniques 260

Concluding Remarks 261

10 ML Models and Security 263

Evasion Attacks 264

Poisoning Attacks 279

Concluding Remarks 284

11 Fair Product Design and Deployment 285

Reasonable Expectations 286

Fiduciary Obligations 288

Respecting Traditional Spheres of Privacy and Private Life 289

Value Creation 290

Complex Systems 292

Clear Security Promises and Delineated Limitations 294

Possibility of Downstream Control and Verification 294

Products That Work Better for Privileged People 295

Dark Patterns 298

Fair Products Checklist 300

Concluding Remarks 301

12 Laws for Machine Learning 303

Personal Data 309

Algorithmic Decision Making 312

Security 314

Logical Processes 316

Some Application-Specific Laws 318

Concluding Remarks 321

Index 323