Attacking Identity Semantics in DeepFakes via Deep Feature Fusion
Bing Fan, Zihan Jiang, Shu Hu, Feng Ding
Abstract
DeepFake is now widely known for forging and fabricating falsified faces. In recent years, forensic researchers have built many algorithms as countermeasures against DeepFake. Albeit the great progress made so far in discerning falsified faces from pristine ones, there are now potential new threats to manipulating the semantics information of facial images. To combat the new type of disinformation, it is necessary to investigate them by developing anti-forensics tools for facial semantics. To this end, we propose the identity-feature-extracted adversarial network (IFE-GAN) to launch attacks on identity semantics. Also, we design the ConvolutionFormer block to refine the visual quality of synthesized anti-forensics images. Through experimental evaluations, we justify the effectiveness of our proposed method in disrupting face identity detectors while preserving satisfying visual quality.