A Face Pre-Processing Approach to Evade Deepfake Detector

Abstract

Recently, various image synthesis technologies have increased the prevalence of impersonation attacks. With the development of such technologies, damages to people such as defamation or fake news have also increased. Deepfakes have already evolved to the point, where people cannot easily distinguish fake from real. This leads to an urgent need for developing detection methods. Currently, in order to detect deepfakes, many deepfake datasets are widely used in deep neural networks. And several methods have been proposed and demonstrated to be effective in detecting deepfakes. In this work, we present pre-processing techniques such as face restoration, edge smoothing, face beautification to mitigate the artifacts of deepfakes and makes them appear more natural to humans, while lowering the deepfake detection performance. Through extensive experiments, our method can significantly lower the performance of the state-of-the-art deepfake detectors and expose the vulnerability of deployed detectors.