MMDFD- A Multimodal Custom Dataset for Deepfake Detection.

A multi-modal deepfake dataset is relevant in addressing the growing concern of deepfake misuse, which poses a significant security and privacy threat. Deepfakes are becoming increasingly sophisticated, and their potential to deceive individuals and organizations is a significant issue. The ability to generate synthesized human voices using deep learning models and inserting fake subtitles has added to this problem, making it more challenging to detect deepfakes accurately. A superior quality dataset is essential to developing a competent deepfake detector. However, existing datasets are limited and often biased, making it difficult to detect deepfakes accurately. A multi-modal deepfake dataset, such as the proposed multi-modal Audio-Video-Text Deepfake dataset (MMDFD) addresses this gap by providing a more realistic and unbiased dataset. Such a dataset helps develop more accurate and effective deepfake detection methods, which can detect audio, video, and textual deepfakes simultaneously. The proposed dataset is more reflective of situations in the real world since it contains actual YouTube recordings of celebrities from four different racial origins. This helps to avoid the creation of deepfake detectors that are biased toward certain racial or ethnic groups. Overall, a multi-modal deepfake dataset is essential in addressing the growing concerns of deepfake misuse and developing effective detection methods that can detect deepfakes accurately, regardless of the medium.