FATE: Learning Effective Binary Descriptors with Group Fairness.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep hashing methods offer computational efficiency for large-scale data retrieval.
• Existing deep hashing approaches may pose ethical risks, particularly concerning group fairness in image retrieval.

Purpose of the Study:

• To investigate and address the problem of group fairness within the learning-to-hash framework.
• To introduce a novel method, Fairness-aware Hashing with Mixture of Experts (FATE), for unbiased binary descriptor generation.

Main Methods:

• FATE utilizes a mixture-of-experts framework for the hashing network, enabling aggregated knowledge from diverse viewpoints.
• Incorporates fairness-aware contrastive learning using sensitive labels and feature similarities.
• Employs an adversarial learning objective on deep features and hash codes to mitigate group biases.

Main Results:

• FATE demonstrates superior performance compared to state-of-the-art approaches in generating discriminative and unbiased binary descriptors.
• The method effectively enhances model capability for fair and efficient image retrieval.
• Extensive experiments on benchmark datasets validate the effectiveness of FATE.

Conclusions:

• FATE successfully addresses group fairness challenges in deep hashing for large-scale image retrieval.
• The proposed method offers a robust solution for generating unbiased and discriminative hash codes.
• FATE represents a significant advancement in ethical considerations for AI-driven retrieval systems.