Deepfake Video Detection Using a Hybrid ResNeXt and LSTM Architecture

Abstract

The growing spread of deepfake materials presents a serious threat to individual privacy, media credibility, and public trust. Existing detection methods often struggle to generalize across various manipulation techniques and video quality levels. This study proposes a hybrid architecture based on deep learning (DL) is introduced, which leverages the spatial feature extraction strengths of ResNeXt-50 along with the temporal sequence modeling capabilities of LSTM networks. The suggested framework handles video input by initially obtaining frame-wise features via a pretrained ResNeXt-50 backbone and then examining temporal dynamics through an LSTM layer. Experimental evaluations were conducted using benchmark datasets, including Deepfake Detection Challenge (DFDC), Celeb-DF, FaceForensics++, and DFD. Findings indicate that the developed model significantly outperforms conventional CNN-LSTM combinations, attaining 95.7% accuracy on the DFDC dataset and above 90% on the other datasets. This research highlights the practical applicability of hybrid DL techniques in real-world video authentication systems and contributes a high-performance solution to the growing field of synthetic media detection.