Subgroup Invariant Perturbation for Unbiased Pre-Trained Model Prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning systems exhibit vulnerabilities, notably bias, impacting fairness and trust.
• Bias in AI applications like facial recognition stems from unbalanced training data representation.
• This bias disproportionately affects specific demographic subgroups.

Purpose of the Study:

• To introduce a novel metric, Precise Subgroup Equivalence (PSE), for quantifying AI model bias and performance.
• To develop a bias mitigation algorithm inspired by adversarial perturbation techniques.
• To enhance the fairness and reliability of deep learning models.

Main Methods:

• Developed the Precise Subgroup Equivalence (PSE) metric to jointly assess prediction bias and model performance.
• Proposed a bias mitigation algorithm utilizing the PSE metric and adversarial perturbation principles.
• Introduced Subgroup Invariant Perturbation (SIP) to create a transformed dataset for bias reduction.

Main Results:

• The proposed Subgroup Invariant Perturbation algorithm effectively reduces bias in AI model predictions.
• Experiments on four public face datasets demonstrated the algorithm's efficacy for race and gender bias mitigation.
• The PSE metric successfully quantified bias and model performance.

Conclusions:

• The developed PSE metric and SIP algorithm offer a promising approach to address bias in deep learning.
• The findings contribute to building more equitable and trustworthy AI systems.
• Further research can explore broader applications of these bias mitigation techniques.