Demographic Fairness and Accountability of Audio- and Video-Based Unimodal and Bi-modal Deepfake Detectors

With the advances in deep generative models, facial forgery by advanced deepfake generation techniques has posed a severe societal and political threat. Recently, a new problem of generating a synthesized human voice of a person is emerging. With the developing threat of impersonation attacks using deepfake audio and video data, a new generation of deepfake detectors is being investigated to collectively focus on such data. For the commercial adoption and large-scale rollout of the deepfake detection technology, it is vital that such detectors don’t discriminate across demographic variations such as gender and race. The aim of this chapter is to thoroughly examine the bias of audio- and facial video-based unimodal and bi-modal deepfake detectors across gender and race. Thorough experimental investigations are conducted on bi-modal audio-video FaceAVCeleb and KoDF deepfake datasets annotated with gender and race labels. Our experimental results suggest that both audio- and video-based deepfake detectors obtain performance differential across gender and race. However, the bi-modal audio-visual deepfake detectors obtain a lower performance differential of about 2.05% and 7.3% across gender and race in terms of EER over audio- and facial video-based deepfake detectors, individually.