SABAF: Removing Strong Attribute Bias from Neural Networks with Adversarial Filtering