Detox Loss: Fairness Constraints for Learning With Imbalanced Data

Recent studies have highlighted a major caveat in several popular deep learning based face recognition algorithms. Despite their high accuracy, they have shown to exhibit biased behavior by achieving sub-par performance on images belonging to different sub-groups. Such behavior can often lead to unfair predictions, thus presenting a need to develop fairer models with unprejudiced behavior. This research proposes Detox loss, a novel bias invariant feature learning loss function for learning unbiased models. The proposed loss can be used to: (i) learn fairer deep learning classifiers, and (ii) mitigate bias from existing pre-trained networks, especially in the challenging constraint of imbalanced training data with respect to a protected attribute. Conceptually, the Detox loss enforces that the learned features are distinguished based on the task label only, while eliminating any distinction based on the biasing-attribute. This is achieved by incorporating three fairness constraints while training with the traditional classification loss: (i) proposed Exclusion loss, (ii) proposed Inclusion loss, and (iii) proposed Feature-distillation loss. The efficacy of the Detox loss is demonstrated on two facial analysis tasks: (i) age-group prediction and (ii) gender prediction, under the protected attribute of ethnicity for learning fairer models and de-biasing existing models, with varying imbalanced training data distributions. Across two different protocols, three setups, and two tasks, the Detox loss obtains state of the art performance without the need of multi-labeled training data. For example, on the challenging Pilot Parliaments Benchmark (PPB) dataset, the Detox loss obtains a balanced accuracy and
$F_{1}$
score of 96.1% and 96.6%, as compared to the state-of-the-art performance of 94.1% and 93.6%, respectively.