Deep Learning Approaches to Fairness Based Bias Mitigation in Facial Recognition Systems: A Comprehensive Review
Olukayode Kelvin OLUKOJU, Peter Ogedebe · Technix International Journal for Engineering Research · 2025
Deep learning-based face recognition has shown great accuracy but encounters big issues with fairness and bias between demographics. This survey systematically reviews the state-of-the-art bias mitigation methods in facial recognition systems based on fairness, showing 45 representative works in the years from 2015 to 2025. By conducting systematic quantitative and qualitative comparison among the state-of-the-art, we categorize these strategies into four major classes based on the type of the proposed mitigation strategy: data-level (15-25% reduction of bias), algorithm-level (20-35% reduction of bias), system-level (25-40% reduction of bias), and emerging hybrid techniques (35-50% reduction of bias). We disclose the significant advancements in bias mitigation approaches demonstrated in our work and we also point to, and work that needs to be done in, standardised evaluation frameworks, bias at intersectional levels and privacy-fairness mechanisms to perform intersectional bias analysis. We also outline a revised taxonomy of bias types and provide a holistic evaluation framework for the assessment of facial recognition fairness. In recent years, trade-offs between privacy, accuracy, and fairness related to facial recognition technology, as well as regulatory compliance have been key drivers of its innovation.