Supplementary Material-PASS

In this supplementary material, we provide information about the following: 1. Relation between attribute predictability and bias (Sec. 1), 2. Detailed algorithm (pseudocode) for PASS and MultiPASS (Sec. 2), 3. Hyperparameters used for training PASS and MultiPASS systems (Sec. 3), 4. Hyperparameters for training IVE systems (Sec. 4), 5. Our hair-obscuring pipeline (similar to [1]) (Sec. 5), 6. Detailed results (including verification plots) for de-biasing methods applied on Arcface/Crystalface descriptors (Sec. 6), 7. Ablation study for PASS systems (Sec. 7), 8. Effect of training a discriminative embedding (TPE[14]) on face descriptors and their PASS counterpart (Sec. 8), 9. Advantages of deploying PASS system over end-to-end training (Sec. 9), 10. Discussion about the trade-off between bias reduction and drop in verification performance (Sec. 10).