A Hybrid Deep Neural Network for Multimodal Deepfake Detection

Abstract—Detecting AI-manipulated media, especially deep- fake images, videos, and audio, is critical in combating privacy and security threats. This project introduces a hybrid deep learning model that integrates content and trace feature extrac- tion to enhance detection accuracy and robustness. By utilizing CNN and RNN architectures, the model processes datasets from DeepFake and Face2Face algorithms, effectively identifying subtle manipulations even under challenging conditions like video compression. Experimental results highlight its superior perfor- mance over baseline methods, supported by Class Activation Maps (CAMs) that pinpoint critical areas.Additionally, this model integrates seamlessly with a social media-like application to provide real-time monitoring of uploaded media. The system not only detects and flags AI-generated manipulations but also alerts users about suspicious content as it is shared or uploaded. This real-time functionality enhances the system’s usability in dynamic environments, making it an efficient tool for advancing media forensics, combating misinformation, and ensuring content authenticity across digital platforms. Keywords: Deepfake Detection, Hybrid Neural Net- works, Multimodal Analysis, CNN, RNN, Media Forensics, Face2Face.