Controlling Bias in Generative AI: Techniques for Fair and Equitable Data Generation in Socially Sensitive Applications

Generative AI models have transformed content creation by enabling high-quality image and data synthesis. However, these models often inherit and amplify societal biases present in their training datasets, leading to concerns about fairness in their applications. This paper addresses this issue by proposing a comprehensive framework for identifying, measuring and reducing bias in generative AI systems. The framework uses data reweighting, fairness-constrained training, and post-processing to enhance fairness across demographics like age, gender, and ethnicity. Using the FFHQ dataset, the framework demonstrates substantial improvements in fairness metrics, with demographic parity improving from 0.42 to 0.15 and equalized odds from 0.35 to 0.12, while maintaining competitive image quality, as indicated by a slight increase in FID from 7.21 to 8.12. Qualitative analysis confirms the generated outputs closely align with ideal demographics, highlighting the framework's effectiveness in addressing fairness without sacrificing performance, thus offering a practical solution for socially responsible AI development.