Enhancing Deepfake Detection Through ResNeXt-50 and xLSTM-based Temporal-Spatial Analysis

The rise of Generative Adversarial Networks (GANs) enables the creation of highly realistic deepfakes, raising concerns about misinformation and privacy. Deepfake content produces various negative effects on society through political manipulation and financial scams but simultaneously makes it extremely hard to check facts and verify content authenticity. An computational system using ResNeXt-50 features that operates with an xLSTM model to verify video authenticity through various visual elements is proposed. The model proposes deepfake detection methods by implementing the hybrid combination of ResNeXt-50 CNN and xLSTM models. ResNeXt-50 extracts valuable features from individual video frames to understand face characteristics as well as visual patterns. The data classification analysis showed 90% maximum accuracy when analyzing databases with 40 and 60 frames of content. The analysis of the 10-frame dataset recorded the least successful outcomes with an accuracy rate at 52%. The combination of ResNeXt-50 with xLSTM enabled the system to detect deep-fakes in videos with minimum image resolution loss.