AIF360/aif360/algorithms/inprocessing/adversarial_debiasing.py at main · Trusted-AI/AIF360

Abstraction: In-processing adversarial debiasing classifier for fairness

Key points:

• AdversarialDebiasing class (in IBM AIF360) trains a classifier to maximize accuracy while an adversary tries to predict protected attributes from classifier predictions
• Gradient projection removes adversary gradient component from classifier update, then subtracts adversary_loss_weight * adversary_grad (default 0.1)
• Built on TensorFlow v1 (graph mode only; raises RuntimeError in eager execution)
• Architecture: one hidden layer (200 units, ReLU, dropout) classifier; adversary uses 3-feature sigmoid input conditioned on true label
• Exponential learning rate decay (0.001 start, factor 0.96 per 1000 steps) with Adam optimizer
• Based on Zhang, Lemoine, and Mitchell 2018 AAAI/ACM AI Ethics paper "Mitigating Unwanted Biases with Adversarial Learning"

Connections: Aif360 · Ibm · Algorithmic Fairness · Adversarial Training · Bias Mitigation

Source: https://github.com/IBM/AIF360/blob/master/aif360/algorithms/inprocessing/adversarial_debiasing.py