Protected attributes are personal characteristics, such as race, gender, age, disability, or religion, that are legally or ethically safeguarded against discriminatory treatment in automated decision-making systems. Fairness metrics and bias mitigation techniques use protected attributes as the basis for measuring whether a model’s predictions differ systematically across groups defined by those characteristics. Handling protected attributes correctly, including deciding whether to use them directly or as held-out variables for auditing, is central to building fair and legally compliant machine learning systems.