Deepfake Video Analysis Using a Dual-Stream ResNeXt and LSTM Deep Learning Framework

Abstract

The fast development of digital media technologies has greatly erased the line between actual and manipulated visual content, especially with the advent of deepfake videos. These are very realistic synthetic videos that are very difficult to verify identities, cybersecurity, and integrity in social media. In order to overcome this problem, the proposed paper suggests a innovative dual-stream deep learning architecture that can be implemented to analyze deepfakes video robustly using ResNeXt and LSTM networks. The suggested model is able to capture the complex facial features and time differences between video frames, making it possible to detect manipulated material. Moreover, channelwise attention mechanism is added to increase the discrimination of features and to enable the overall detection. The framework is tested against pre-trained models like resnet50 and VGG16 on benchmark datasets. The results of the experiment have shown that the proposed hybrid model has a better performance. These results indicate its usefulness in differentiating real and deepfake videos. The suggested solution is a scalable and efficient solution to be used in the digital forensics, identity authentication, and cybersecurity.