A Bias Metrics A.1 Symbols For a binary classification task, we abbreviate the content in the confusion matrix as TP

Its protected label is whether the images are color or grayscale. Task (4) uses the task label "attractive" and the License: The CelebA dataset is available for non-commercial research purposes only [57]. Then a downstream classifier is trained to predict the task label on this representation. In addition to the p-value from U-test, we use Cohen's We use Levene's test to compare the magnitude of the variance between experiment The significance level is also 5%. We use an RTX 2080Ti graphics card with 11GB of memory for each training run.