Detection-by-Simulation: Exposing DeepFake via Simulating Forgery using Face Reconstruction

Abstract

This paper describes a new method to expose DeepFakes via simulating forgery using face reconstruction on real samples. Our method is motivated by that the DeepFake model introduces generation artifacts on synthesized faces, which can be simulated by similar CNN-based generative models. To simulate these forgery artifacts, we develop a simple auto-encoder network to reconstruct faces, as the generation process in face reconstruction shares some certain common properties with the generation process in the DeepFake model. Thus we can use the reconstructed faces as negative training samples. Then we develop a CNN network to fully take advantage of the simulation. Specifically, we design two components, an attention guided blending boundary prediction branch to predict blending boundary and a semantic feature enhancement to convey semantic information to deep layers. Then the proposed network is trained using the simulated faces and real faces. Extensive experiments are conducted on $FF++$ and Celeb-DF with comparison to several state-of-the-arts, which demonstrates the efficacy of our method.