Reproduce “Copying As Defense: Protecting Images from Editing By Eliciting Model Copying” 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
Salma Abdel Magid, Sanghyuk Chun, Eric Slyman, Kushal Kafle, Olga Russakovsky
CaD protects images from instruction-based editing with bounded per-image perturbations that maximize target-to-reference attention in early transformer blocks. This shifts attention toward the reference and away from the instruction and evolving output. On FLUX.1 Kontext’s BodyBench test at ε = 12/255, edit success was 5.1%, versus 24.9% for DeContext. Results cover two editors; preprocessing resilience was not evaluated.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.