Detecting and localizing multiple forgeries with TransGAN

Detecting and localizing multiple forgeries with TransGAN

연구 분야: Safety

학회: Iran Journal of Computer Science

The rapid rise of manipulated videos in the digital era necessitates advanced forgery detection and localization techniques. This study presents a novel approach using a GAN-based framework integrated with a 3D Swin-B Transformer Encoder–Decoder architecture for future frame prediction. A ConvLSTM-based Siamese network serves as a discriminator, capturing spatiotemporal inconsistencies, while forgery localization is enhanced through attention maps and a patch-based strategy. Experimental validation on diverse datasets demonstrates the effectiveness of the proposed method, achieving 98.97% detection accuracy and 100% localization accuracy, surpassing existing techniques. The results highlight the robustness of the approach in identifying and localizing various types of video forgeries, making a significant contribution to digital content forensics.