Continual learning increases unfairness in deepfake detectors
A loss of accuracy usually increases unfairness, and fairness increases only a small number of times.
Claimed, not confirmed
Researchers train deepfake detectors in sequence on 11 deepfake generation methods. They monitor 4 dimensions of fairness across demographic groups. A loss of accuracy usually increases unfairness. The effect is different for each learning method and group.
Sources
Posted
