A Fair Generative Model Using LeCam Divergence

We explore a fairness-related challenge that arises in generative models. The is biased training data with imbalanced demographics may yield high asymmetry size of generated samples across distinct groups. focus on practically-relevant scenarios wherein demographic labels are not available and therefore the design fair model non-straightforward. In this paper, we propose an optimization framework regulates unfairness under such practical settings via one statistical measure, LeCam (LC)-divergence. Specifically to quantify degree unfairness, employ balanced-yet-small reference dataset then measure its distance using LC-divergence, which shown be particularly instrumental small dataset. take variational approach implement LC-based measure. Experiments benchmark real datasets demonstrate proposed can significantly improve fairness performance while maintaining realistic sample quality for wide range set all way down 1% relative set.