Towards Accuracy-Fairness Paradox: Adversarial Example-based Data Augmentation for Visual Debiasing