Coverage for Detecting Bias in Facial Recognition Datasets

Facial recognition algorithms are known to exhibit varying performance across different subgroups defined by factors like race and gender. These performance differences have been attributed to disparities in the representation of subgroups in the training data. Even when trained on large-scale datasets (i.e., “big data”), these disparities may persist. When the underlying training data is biased, the trained model is likely to be biased as well. Biased facial recognition models may be unreliable for underrepresented groups and lead to real world failures such as incorrect access control decisions in biometric systems or misidentification of individuals. Therefore, methods for assessing the breadth of intersectional contexts covered by datasets during the pre-processing phase before significant resources are invested in training models are essential. This work explores whether frequency coverage can detect data bias and whether coverage correlates with deep learning model performance via experiments on two facial recognition datasets with attributes for gender, race, and age. This work also explores the usefulness of visualizations to practitioners in making decisions about data.