Resampling Strategy for Mitigating Unfairness in Face Attribute Classification

With the widespread success of artificial intelligence (AI) systems, various applications based on the systems are being applied to our daily life. However, AI systems also raise societal problems since it is highly dependent on training datasets with bias. Consequently, concerning about trustworthiness in AI systems becomes a popular research topic, and recent studies reveal unfairness in developed models. In this paper, we propose a new batch sampling strategy considering fairness among demographic groups. Unlike conventional batch sampling methods such as under-sampling or oversampling, we reflect the notion of fairness directly to estimate the batch sampling probability of data. We empirically demonstrate that our batch sampling method achieves fairer results compared to prior methods in image classification tasks on CelebA and UTKFace datasets.