Course contentsShow
← Machine Learning and Deep Learning
Lesson 3294 of 3,53872. Bias and FairnessFree lesson

Protected Attributes and Sensitive Features

Identifying protected characteristics (race, gender, age) and their role in fairness analysis.

Protected Attributes and Sensitive Features

What you'll learn: How to identify and handle characteristics that warrant special fairness consideration in ML systems.

What Are Protected Attributes?

Protected attributes (also called sensitive features) are characteristics of individuals that should not—legally or ethically—influence certain decisions. These typically include:

  • Race or ethnicity
  • Gender or sex
  • Age
  • Religion
  • Disability status
  • National origin
  • Sexual orientation

These attributes are "protected" because historical discrimination and societal biases make them prone to unfair treatment in automated decision-making systems.

Why They Matter in Fairness Analysis

Protected attributes play a central role in fairness evaluation. When analyzing whether your model treats different groups fairly, you stratify outcomes by these features. For example, does your loan approval model grant loans at similar rates across racial groups? Does your hiring algorithm screen candidates differently by gender?

The Paradox of Awareness

Here's the tricky part: to measure fairness, you must know these attributes. Yet using them directly as model inputs often feels wrong—and may be illegal in some contexts (like U.S. employment law). This creates a tension: you need protected attributes for fairness auditing, but you may want to exclude them from training.

However, simply removing these features doesn't guarantee fairness—correlated features (zip code correlates with race, for instance) can serve as "proxies" that encode the same information.

Practical Identification

In your dataset, ask: which columns represent characteristics that define legally or ethically protected groups? Flag these early. They guide your choice of fairness metrics and mitigation strategies throughout the ML pipeline.

Key Takeaway: Protected attributes are sensitive demographic features (race, gender, age) that require special attention in fairness analysis—you need them to measure bias, even if you exclude them from model training.