Anti-Persona: Disrupting Unauthorized Identity Binding and Recognition in Personalized Vision-Language Models

Anti-Persona is an image-level defense for personalized vision-language models (LVLMs) that can block unauthorized identity binding and recognition, achieving a protection rate of up to 95.0% on two representative LVLMs. The method aggregates visual features shared among reference images into an identity prototype, then optimizes visually subtle perturbations to disrupt prototype alignment in the encoder. The protection does not depend on specific prompts, can withstand image compression, and improves black-box transfer when encoders do not match.