A Multi-Layer Capsule-based Forensics Model for Fake Detection of Digital Visual Media

摘要

the dangers generated from synthesized multimedia are increasing every day. The creation of the so-called Deepfakes multimedia is vastly evolving, making the detection task harder every day. Researchers and corporations are interested in exploring the technology limits and are coming up with new tools every year to create more robust fake media. In this paper, a new enhanced fake video detection model is introduced addressing many of the face-swapping threats and the low generalization problem. A preprocessing stage is proposed to minimize the noise in the data to enhance their quality. The proposed architecture uses a modified application of capsule neural networks (CapsNet) with an enhanced routing technique. It does not require a lot of training data and generates a small number of training parameters making it fast to build. The model was trained and tested using the DFDC-P dataset and the results have proven that it outperformed other detectors in terms of detection recall, weighted precision, and F1 score.