Protected Attribute

Protected attributes are not merely statistical variables; they are defined by a complex web of anti-discrimination laws and sector-specific regulations.

• U.S. Context: Rooted in the Civil Rights Act, Fair Housing Act, and Equal Credit Opportunity Act (ECOA), which explicitly prohibit discrimination based on race, color, religion, national origin, sex, marital status, and age.
• EU Context: The General Data Protection Regulation (GDPR) defines 'special categories of personal data' with strict processing prohibitions, including racial or ethnic origin, political opinions, and biometric data.
• Sectoral Rules: Financial services must comply with ECOA, while healthcare faces HIPAA constraints, creating overlapping definitions that vary by industry.

A critical distinction in fairness auditing is whether a protected attribute is directly observed or inferred through correlated variables.

• Explicit Attributes: Directly collected data fields like gender, race, or age_bracket. Fairness through unawareness often fails by simply removing these columns.
• Proxy Attributes: Non-protected features that encode sensitive information. A zip code can act as a strong proxy for race due to residential segregation, while browser history or purchased product categories can proxy for religion, health status, or pregnancy.
• Detection: Auditors use mutual information and correlation analysis to identify and mitigate proxy discrimination, which is often unintentional but legally actionable.

Modern fairness auditing moves beyond single-axis analysis to examine how overlapping protected attributes create unique, compounded disadvantages.

• Kimberlé Crenshaw's Framework: Originating in legal theory, intersectionality examines how race and gender combine to produce discrimination that is distinct from racism or sexism alone.
• Technical Implementation: Multicalibration and subgroup fairness metrics require models to be calibrated not just for 'women' and 'Black individuals' separately, but for the specific intersection of 'Black women'.
• Computational Challenge: The combinatorial explosion of intersecting subgroups makes exhaustive auditing computationally expensive, requiring efficient slicing analysis and statistical prioritization.

The mutability of an attribute influences both its legal protection and the design of algorithmic recourse mechanisms.

• Immutable Attributes: Characteristics like race, national origin, and age are inherent and cannot be changed. Discrimination on these grounds is almost universally prohibited.
• Mutable Attributes: Some protected attributes, like religion or marital status, can change. However, anti-discrimination law protects individuals regardless of their current status.
• Recourse Design: For mutable attributes, a system can provide actionable recourse (e.g., updating a changed surname after marriage). For immutable attributes, recourse must focus on correcting systemic model errors, not changing the individual.

Protected attributes manifest in different data types, each requiring distinct fairness metrics and bias mitigation strategies.

• Categorical: Binary or multi-class labels like gender (male/female/non-binary) or race. Metrics like demographic parity and equalized odds are designed for these discrete groups.
• Continuous: Numerical values like age (0-100+). Fairness for continuous attributes often requires discretization into brackets or the use of regression-based fairness metrics that constrain correlation coefficients.
• Derived/Inferred: Attributes not self-reported but inferred by a model, such as sexual orientation inferred from social media activity. These raise acute privacy concerns and are subject to strict GDPR restrictions on automated inference of special category data.

A characteristic's status as a 'protected attribute' is not universal; it depends entirely on the legal jurisdiction and the specific decision-making context.

• Geographic Variance: Caste is a protected attribute in India but is not explicitly codified in U.S. federal law. Genetic information is protected under GINA in the U.S. but may fall under broader health data rules in the EU.
• Contextual Relevance: Using age as a pricing factor in life insurance is actuarially sound and legally permitted, but using age to filter job applicants is age discrimination. The attribute's legitimacy is defined by the decision's purpose.
• Operationalization: A global ML platform must maintain a dynamic registry of protected attributes mapped to each deployment jurisdiction and use case to ensure automated compliance.