Saliency-Based diversity and fairness Metric and FaceKeepOriginalAugment: A Novel Approach for Enhancing Fairness and Diversity

Data augmentation can help reduce AI bias: The study demonstrates that FaceKeepOriginalAugment, a modified data augmentation technique, successfully mitigates gender bias in occupational datasets (CEO, Engineer, Nurse, School Teacher), showing that how we prepare training data directly impacts fairness outcomes.

Multiple types of bias require attention: The research addresses three distinct bias categories—geographical, gender, and stereotypical biases—in facial recognition systems, indicating that fairness efforts must be comprehensive rather than targeting a single bias type.

Diversity and accuracy don't have to conflict: By maintaining "a delicate balance between data diversity and information preservation," the method shows that augmentation strategies can increase dataset diversity while preserving important information, suggesting fairness improvements need not compromise model performance.

New metrics are needed to measure fairness: The paper introduces the Image Similarity Score (ISS) and Image-Image Association Score (IIAS) as tools to quantify diversity and bias in datasets, highlighting that developing better measurement methods is essential for evaluating whether AI systems are actually becoming fairer.

Testing across multiple datasets strengthens fairness claims: The evaluation spans six different facial datasets (FFHQ, WIKI, IMDB, LFW, UTK Faces, and a Diverse Dataset) and multiple model types (CNNs and vision transformers), demonstrating that robust fairness solutions should be validated across varied conditions rather than single benchmarks.