Unmasking Deepfakes - Harnessing the Potential of 2D and 3D Convolutional Neural Network Ensembles

This study introduces a novel strategy for verifying the authenticity of video footage within the rapidly evolving field of deepfake identification. The proposed method combines sophisticated two-dimensional and three-dimensional Convolutional Neural Networks. The three-dimensional network is specifically designed to discern spatiotemporal attributes using sliding filters that traverse both space and time, allowing for intricate detection of patterns in pixel sequences and their progression over time. In parallel, the two-dimensional network employs the EfficientNet architecture, taking advantage of its convolutional auto-scaling capabilities. This approach is further enhanced by the implementation of Adaptive Weighted Ensembles and Voting Ensembling techniques. A deliberate emphasis on the outputs from the three-dimensional model is made, exploiting its superior ability to extract spatiotemporal features. Rigorous empirical tests confirm the superiority of this approach, as evidenced by improved metrics such as a higher Area Under the Curve by 2% and reduced LogLoss, demonstrating its capability to effectively combat the sophistications of deepfake technology.