Exposing DeepFakes using Siamese Training

Developments in the deep learning domain have led to its usage in solving problems ranging from robotics to computer vision and healthcare. However, due to easy access to large public datasets, technologies based on deep learning have been utilized to create software that can be used for immoral purposes. One such application of deep learning is facial manipulation in videos also known as deepfakes. If used maliciously, deepfakes can have devastating consequences on society such as cyberbullying, scamming, and spreading false and vindictive news which pose serious threats to individuals as well as to the nation by disrupting its peace and security, etc. Hence, developing techniques to detect manipulated media is of utmost urgency and importance. Using a siamese network architecture, this research proposes an ensemble-based metric learning approach in which several models are created beginning from a base network to identify face manipulation. The technique has been evaluated against publicly available datasets: DFDC, FaceForensics++, and Celeb-DF (v2) and shows promising results in detecting manipulations during self and cross-testing.