Enhanced Deepfake Detection with LSTM and ResNeXt Integration

Created from advanced artificial intelligence (AI), Deepfakes are privacy, security, and trust shattering to digital media needs. In this work, we propose a novel deepfake detection model based on ResNext followed by Long Short Term Memory (LSTM) networks to achieve extraction of spatial features and temporal sequence analysis, respectively. Our approach combines the strengths of the two architectures and minimizes the investment in model size by retaining only a small portion of that which is accurate. On the Celeb-DF (v2) dataset, our model reached a detection accuracy of 93.59%, beating baseline methods. The demonstrated robustness and reliability of the proposed method make it a promising method to fight deepfakes in real world situations.