Enhancing Fairness in Image Classification: Modified Loss Functions for Mitigating Race and Sex Bias

Artificial intelligence and machine learning have become ubiquitous in everyday life. With ubiquity, there is a need to ensure minimal algorithmic bias from these systems. Training data distributions can cause these systems to be unfair, negatively impacting users. One solution to this is to ensure fair, unbiased training sets. Unfortunately, unbiased training data is not enough to fully mitigate the issue. In this work, we utilize modified loss functions to mitigate race and sex bias that can appear when training machine learning models on biased data. We used face images from the FairFace dataset for binary and categorical classification tasks. Although FairFace has better diversity than other available face image datasets, it remains biased due to uneven distributions of race-sex groups. We find that the modified loss functions work moderately well at mitigating data bias. In some cases, combining multiple loss functions yields improved results compared to using one alone.