Reproduce “Face Reconstruction from Face Embeddings using Adapter to a Face Foundation Model” 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
Hatef Otroshi Shahreza, Anjith George, Sébastien Marcel
Researchers at Idiap Research Institute developed a method to reconstruct high-quality face images from biometric embeddings by adapting a pre-trained face foundation model with a lightweight linear layer. This approach achieved success attack rates up to 100% on benchmark datasets like MOBIO and significantly reduced the computational overhead for template inversion attacks.
Key insights
- Leveraging a pre-trained face foundation model with a simple, data-efficient adapter enables highly effective and computationally efficient face reconstruction attacks against various blackbox systems.
- Face recognition models exhibiting higher original recognition accuracy (True Match Rate) tend to be more susceptible to these advanced template inversion attacks.
- While the method preserves identity strongly, it sometimes struggles with accurately maintaining fine-grained attributes such as age, extreme expressions, or specific ethnicities in the reconstructed images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.