Table 3.

Summary of the existing debias methods and the proposed method.

Method	Implementation	Does it require demographic information?	Does it require re-training the model?	Difficulty
Baseline	Do not consider the demographic group differences.
Balanced21	Reduce the sample size of the majority group to achieve a balanced population for each group.	Yes	Yes	The amount of data decreases.
Stratified21	Train separate models for each group.	Yes	Yes	Minority groups may have insufficient data, resulting in a poorly trained model.
Adversarial54,56	Use an adversary to an adversary to decrease the model's capacity to identify demographic groups.	Yes	Yes	Model-specific; determining the appropriate level of the adversary can be challenging.
DistMatch MMD57	Add a penalty to reduce the maximum mean discrepancy58 distance between groups	Yes	Yes	The data imbalance between demographic groups, different data splits, and distance metrics during training may lead to instability in calculating the distance.
DistMatch Mean57	Add a penalty to reduce the mean of the distribution between groups.
FairALM59	Apply an augmented Lagrangian method to penalise the distribution discrepancy.	Yes	Yes	Different assumptions regarding distribution can yield varying results.
Proposed augmentation	Use image augmentation to prevent the model from learning shortcut based on demographic information	No	Noa	Time-consuming when augmenting images

^aIt is not necessary to re-train the model as the image augmentation can be applied during the test phase.