OGAN: Disrupting Deepfakes with an Adversarial Attack that Survives Training

no code implementations • 17 Jun 2020 • Eran Segalis, Eran Galili

We complement these efforts by introducing a novel class of adversarial attacks---training-resistant attacks---which can disrupt face-swapping autoencoders whether or not its adversarial images have been included in the training set of said autoencoders.

Adversarial Attack Bilevel Optimization +2