Abstract

Now a days, due to advances in computer editing tools, the creation and manipulation of fake content of images, videos, and audio has become quickly and easily accessible. Deepfakes are synthetic media created using Artificial Intelligence (AI), often with advanced deep learning models, that depict events or individuals that do not exist or did not occur in reality. Typically, deepfakes involve face swapping or full face generation, which creates highly realistic but fabricated content. While such technology has creative and entertainment applications, its misuse has raised serious concerns, including the spread of misinformation, manipulation of digital evidence, and identity-related crimes. The widespread use of masks after the COVID-19 pandemic has further complicated detection, as hidden facial features make it difficult to distinguish real from altered content. Traditional detection methods often struggle with masked or morphed faces, highlighting the need for more robust and general solutions. Addressing these challenges, a deep learning detection framework based on ResNet152V2 is introduced. The system was trained and tested on a combined UADFV and DFFMD dataset consisting of masked and unmasked models. The ResNet152V2 architecture uses residual connections and block normalization to improve learning efficiency, feature representation, and classification accuracy. Experimental findings confirmed the effectiveness of the proposed system, interms of training accuracy is achieved 0.9847 confirming its ability to effectively learn discrimination features. Validation results showed strong generalization to unseen frames, even in cases of hidden or partially covered faces.

Authors

K. Thulasimani, J. Pooja
Government College of Engineering, Tirunelveli, India

Keywords

Deepfake Detection, Artificial Intelligence (AI), ResNet152V2, Masked Faces, Image and Video Manipulation