Deepfake: Classifiers, Fairness, and Demographically Robust Algorithm

Deepfake detection research has seen tremendous success and has achieved remarkably high performance on a few existing datasets. However, the significant drawback of the existing works is the generalizability of the detection algorithms under cross-datasets and cross-attack/manipulation settings. On top of that, another critical bottleneck of deepfake detection literature is the understanding of the fairness quotient of these algorithms. One big reason for such a less explored domain is the unavailability of deep fake datasets covering multiple ethnicities and genders with proper annotations. For example, the popular deepfake detection datasets such as FaceForensics++ and Celeb-DF are highly biased toward Caucasian ethnicity. Recently, a multi-ethnicity multi-modal dataset namely FakeAVCeleb has been released which can fulfill this gap. Henceforth by utilizing the potential of this dataset, we have performed the fairness study of deepfake detection algorithms. For that, several image classifiers are selected which range from deep convolutional neural networks to handcrafted image feature extraction to vision transformers. The experiments performed using such a wide variety of classifiers reveal that the deepfake detectors are not fair and can detect one ethnicity with high accuracy but fail miserably on others. For instance, the performance of one of the popular deepfake detection networks namely XceptionNet shows a reduction of more than 30% when dealing with different ethnicities and genders. Not only ethnicity or gender but also the type of classifiers have a huge impact on the performance. We assert that the proposed study can help in building a fair, robust, and accurate deepfake classifier utilizing insightful findings that can help in the selection of an effective and robust backbone architecture.