An Innovative Method for Identifying Deepfake Videos Through Ensemble Learning
Sabbineni Keerthika, S. Santhiya, P. Jayadharshini, J Ruthranayaki, R Hariarasu, M Naveenan · 2024
The detection of deepfake videos is essential due to the significant potential harm posed by manipulated media. Various deep learning techniques, including customized convolutional neural networks (CNNs), MobileNet, and DenseNet, have demonstrated promise in this domain. This study introduces a system for detecting deepfake videos by leveraging tooth and mouth movements, which are notably difficult to replicate accurately. The proposed methodology utilizes multiple transfer learning algorithms such as InceptionV3, MobileNet, VGG16, VGG19, Xception, and DenseNet121 to enhance the algorithm’s capability to detect and classify deepfake videos based on biological signal attributes derived from jaw and tooth frames. The study specifically targets the identification of deepfake videos by extracting and cropping frames from the input video to emphasize the mouth and nose regions. These preprocessed frames are subsequently fed into a customized CNN trained to classify videos as real or fake, achieving an accuracy of $91.7 \%$. Additionally, pre-trained CNN models DenseNet121 and MobileNetV2 further improve accuracy, attaining $88.4 \%$ and $90.9 \%$, respectively. An ensemble approach that combines the predictions of these models results in an overall accuracy of $93.2\%$.