MagDR: Mask-Guided Detection and Reconstruction for Defending Deepfakes

Original input ⟹ Original output Disrupted input ⟹ Disrupted output Reconstructed input ⟹ Reconstructed output Disrupting Disrupting Disrupting Reconstructing Reconstructing Reconstructing Figure 1: MagDR defends deepfakes from adversarial attacks in three main tasks, face replacement (top), face editing (middle), and facial reenactment (bottom). In each group, we show the original generation effect (left), how adversarial perturbations damage the generation (middle), and how the proposed defender recovers the desired output (right).