Reproduce “RAPID: A Real-Time Defense Against Unauthorized Model Distillation for Text-to-Image Services” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zihan Wang, Boheng Li, Rui Zhang, Wenshu Fan, Qingchuan Zhao, Tianwei Zhang, Hongwei Li, Guowen Xu
RAPID embeds an anti-distillation defense directly into the VAE decoder of a text-to-image service, avoiding the per-image optimization used by methods such as Glaze and PhotoGuard. Across four diffusion models and four datasets, it lowers the quality and prompt alignment of substitute models while keeping protected outputs relatively close to clean images, though with visible artifacts in some cases.
Key insights
- The paper distinguishes perturbation destructiveness from reachability: an attack objective must be jointly realizable by one shared decoder update rather than optimized independently for each image.
- Self-referenced latent maximization avoids external target images and stochastic diffusion trajectories, making the disruption objective more compatible with shared-parameter training.
- Reconstruction-guided color regularization reduces latent shortcuts in which large latent changes disappear after an attacker’s VAE reconstructs the image.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.