Fair and Interpretable Deepfake Detection in Videos

Abstract-- Existing deepfake detection methods often exhibit bias, lack transparency, and fail to capture temporal information, leading to biased decisions and unreliable results across different demographic groups. In this paper, we propose a fairness-aware deepfake detection framework that integrates temporal feature learning and demographic-aware data augmentation to enhance fairness and interpretability. Our method leverages sequence-based clustering for temporal modeling of deepfake videos and concept extraction to improve detection reliability while also facilitating interpretable decisions for non-expert users. Additionally, we introduce a demography - aware data augmentation method that balances underrepre-sented groups and applies frequency-domain transformations to preserve deepfake artifacts, thereby mitigating bias and improving generalization. Extensive experiments on FaceForensics++, DFD, Celeb-DF, and DFDC datasets using state-of-the-art (SoT A) architectures (Xception, ResNet) demonstrate the efficacy of the proposed method in obtaining the best tradeoff between fairness and accuracy when compared to SoT A. I. INTRODUCTION The rise of deepfakes has posed a major threat to the safety and privacy of individuals, institutions, societies, and nations [31], [12]. Scholars posit that with the rapid proliferation of deepfakes, we are heading towards an "infopocalypse" where we cannot tell what is real from what is not [11]. To add to this threat is the fact that the very technologies that enable innovation can be manipulated for creation of deepfakes, resulting in malicious content that undermine privacy and promote disinformation [45].