Enhancing Fairness Without Demographic Labels via Identifying and Mitigating Potential Biases

Asymmetries in data distributions and performance across subgroups can induce systematic unfairness in real-world systems. A variety of previous studies have significantly ameliorated the fairness of deep learning models; however, most of them necessarily require additional labels for sensitive attributes, (i.e., ethnicity and gender). Since sensitive attributes often correspond to personal information, collecting such labels can be restricted and may raise privacy concerns. Although recent work has sought to address these issues by training a model without sensitive attribute labels, we point out that it has limitations, as it assumes specific characteristics of sensitive attributes and is validated in simplistic, constrained environments. Therefore, we propose an Unsupervised Fairness-aware Framework (UFF) that trains a fair classification model without pre-defining the characteristics of the sensitive attributes. It includes branches that capture various types of biases and eliminates them through adversarial training. In various scenarios on benchmark datasets, (i.e., CelebA and UTK Face) for facial attribute classification, the proposed method significantly enhances fairness without assuming specific characteristics of sensitive attributes. Moreover, we introduce g-FAT, which is a new metric to measure generalized trade-off performances between classification accuracy and fairness. For example, on CelebA, ours reduces EO from 11.8 to 7.6 for malignant bias and from 15.6 to 9.6 for benign bias, while improving g-FAT from 80.7 to 84.9 and from 79.0 to 85.2, respectively. In terms of g-FAT, our method achieves the highest trade-off performance among the compared methods on the benchmarks.