Reproduce “Realistic Face Reconstruction from Facial Embeddings via Diffusion Models” 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
Dong Han, Yong Li, Joachim Denzler
Face Embedding Mapping (FEM) learns to translate embeddings from face-recognition and privacy-preserving systems into the identity space expected by a pretrained diffusion model, which then generates face images. Across the tested models, FEM-KAN achieved cross-recognition attack success rates as high as 84.4% for protected-system embeddings; results also varied by protection method, embedding completeness, and image resolution.
Key insights
- Embeddings from several tested privacy-preserving systems supported generation of faces that were recognized by independent systems, indicating that protection does not necessarily remove recoverable identity information.
- Reconstruction risk depends on the protection method and evaluation conditions: performance was weaker for some protected embeddings, at low leakage levels, and on low-resolution images.
- The attack requires more than the leaked embedding alone: it uses paired training data processed by the target model and a pretrained face-generation pipeline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.