MultiFair: Model Fairness With Multiple Sensitive Attributes.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Existing AI fairness interventions often focus on single attributes, leaving combinations unprotected.
• Multi-attribute fairness methods can be computationally expensive or unstable.
• Per-attribute protection risks fairness gerrymandering, where certain attribute combinations remain biased.

Purpose of the Study:

• To develop a method for achieving multi-attribute fairness without additional constraints or prediction heads.
• To create a neutral domain that fuses information across all subgroups and attributes.
• To ensure fair predictions by neutralizing information across sensitive attributes.

Main Methods:

• Proposed the MultiFair method utilizing mixup operations for information fusion.
• Designed three distinct mixup schemes to balance attribute fusion and preserve distinct visual features.
• Conducted extensive experiments on multiple datasets with up to eight sensitive attributes.

Main Results:

• Direct application of mixup operations degraded prediction results due to data unrecognizability.
• The proposed mixup schemes effectively balanced information fusion while retaining critical visual features.
• MultiFair demonstrated successful fairness protection across multiple attributes.

Conclusions:

• MultiFair offers a computationally efficient and stable approach to multi-attribute fairness.
• The method effectively mitigates bias across attribute combinations, preventing fairness gerrymandering.
• MultiFair achieves fairness guarantees while maintaining high prediction accuracy.