VOID: Defeating Unauthorized Imitation in Latent Diffusion Models

VOID is a defense framework that defeats unauthorized imitation by manipulating the intrinsic randomness of latent diffusion models: amplifying latent encoding errors to disrupt image semantic structure, counteracting target guidance signals to suppress the model's recovery capability, while restricting perturbations to regions imperceptible to the human eye without sacrificing visual utility. In evaluations across 5 datasets, 10 imitation attacks, and 24 state-of-the-art defenses, VOID raises the average FID from 113 to 365, a 223% improvement over the strongest defense to date.