GAN Discriminator based Audio Deepfake Detection
Thien-Phuc Doan, Kihun Hong, Souhwan Jung · 2023
Deepfake is a new technology that has emerged in recent times and is becoming one of the great challenges for society and individuals. In particular, scammers could use deepfake to start a phishing attack by cloning the victim’s voice and calling his related person with the fake audio. In this study, we propose a transfer learning model for detecting deepfake audio. This model introduces a method to leverage the discriminator of a GAN-based vocoder model to extract the front-end features of an unidentified audio sample, which helps the model to detect fake voices more easily. With a detection efficiency of up to 94%, we demonstrate that this transfer learning method is feasible.