The following procedure addresses fairness by identifying and mitigating spurious corelation between the attributes via an adversarial training paradigm
-
Trained a pre-trained resnet18, with its weights frozen, added a classifier, and finetuned the classifier layers
Training_CelebA.ipynb -
Evaluated on the test data and obtained accuracies wrt every attribute
inference.py,attr acc.png -
Based on some analysis of the data population and the above results, found "Male" as the spurious attribute and "Bald" as the target (affecting) attribute.
-
Then trained a Binary Classifier (a finetuned resnet18) and observed the bias (Equalized Odds).
Bald_Classifier(Biased).ipynb,equal_odds.py. -
Implemented an adversarial training pipeline to reduce the above bias.
adversarial_training.ipynb -
Lastly, rechecked the bias to ensure Equalized Odds is satisfied.
recheck_odds.py -
Conclusion: Adversarial Training helped reduce the original bias by 20%