Abstract
This paper proposes a new and effective bias mitigation method called targeted data augmentation (TDA). Since removing biases is often tedious and challenging and may not always lead to effective bias mitigation, we propose an alternative approach: skillfully inserting biases during the training to improve model robustness. To validate the proposed method, we applied TDA to two representative and diverse datasets: a clinical skin lesion dataset and a dataset of male and female faces. We identified and manually annotated existing instrument and sampling biases in these datasets, explicitly focusing on black frames and ruler marks in the skin lesion dataset and glasses in the face dataset. Using the counterfactual bias insertion (CBI) method, we confirmed that these biases strongly affect the model performance. By randomly inserting identified biases into training samples, we demonstrated that TDA significantly reduced bias measures by two times to more than 50 times, with only a negligible increase in the error rate. We performed our research on three model families: EfficientNet, DenseNet and Vision Transformer.