Deepfake

“Fairness” means the model does not misclassify individuals based on attributes such as race or gender.

Select correct conclusion.

• A model is fair only when its accuracy is 100 %.

• A model can still be considered fair even if everyone is treated equally badly.

• Fairness requires separate models for each demographic.

In the context of fairness, what does the term 'protected attribute' refer to?

• Attributes that enhance the model's accuracy

• Attributes that should not influence the model's predictions

• Attributes that are irrelevant to model performance

A false positive in the deep-fake detection task is...

• Correctly identifying a fake image.

• Failing to detect a fake image.

• Classifying a real image as fake.

The harmonic mean of Precision and Recall—recommended for imbalanced data —is the:

• Accuracy

• Specificity

• F1 score

“Demographic parity” aims for what outcome in a deep-fake detector?

• Identical classification thresholds for each group

• Our dataset should have proportional to real-world ethnicity groups distribution

• Equal proportion of images predicted ‘FAKE’ across all protected attributes

False-Positive Parity (FPP)
　FP/ (FP+TN)

for perfectly fair model is equal to

• 0

• 1

• infinity

Undersampling with Tomek links, SMOTE/ADASYN oversampling and data-augmentation are cited as examples of which de-biasing stage?

• Pre-process

• In-process

• Meta-learning

• Post-process

What is the biggest problem with fairness-oriented Multi-Threshold Classification?

• It's computationally expensive

• It's hard to determine threshold for each group

• It requires knowledge about protected attribute at inference stage

The deeper layer of the network, ...

• ...the more complex features it perceives.

• ...the more neurons it contains.

• ...the more fair it is.

Why fairness is important in Machine Learning?

• Lack of fairness erodes trust in ML.

• The more fair the model, the higher accuracy it has.

• Improving Fairness results in lower computation costs for ML.