Deepfake Creation and Detection: A Review

Deep generative modeling has transpired into today’s tech world probably more than ever before. Considering the complex datasets amalgamating a wide range of intricacies, deep generative models have still worked their way out in successfully generating results with maximum accuracy. With the advent of detection techniques, the advent of creation of fake contents have been proliferating in numerous ways. The escalation of such fake digital content has instilled skepticism in people, contributing to a zero-trust society where people no longer bother to distinguish truth from falsehood. Apart from the negative connotations associated with DeepFakes, their relevance has also been significant commercially as well as individually. Either way, this technology has triggered an increasing sense of unease in people. This concern proposes an urgent need for automated approaches to detect the fake content. Since these fake content are mostly produced from deep learning based approaches, they take up the name “DeepFakes”. It can take up multiple forms such as images, videos or voice. Notwithstanding the fact that face editing algorithms have been able to yield convincingly realistic human faces, upon close scrutiny, have brought to limelight artifacts which are not visible to the naked eye. Both classical mathematical approaches as well as deep learning based approaches have been fairly successful in detecting these artifacts. In this work, we present a comparative analysis to distinguish which of the methods produce more accuracy in detecting Deepfake images. A classic approach involving frequency domain analysis(FDA) with a classifier is compared with a convolution neural network architecture with pros and cons respectively to infer which of the two methods produce clear-cut results.