Research map / Image generation / Editing and style transfer
FaceRSA: RSA-Aware Facial Identity Cryptography Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FaceRSA, a StyleGAN-latent-space facial identity cryptography framework that anonymizes and deanonymizes faces with password-driven latent edits satisfying RSA-like properties.
Problem
Face anonymization that can be reversed with passwords while keeping image quality and identity-irrelevant attributes, and satisfying RSA-like cryptographic properties (sequential, interchangeable, associative passwords) that prior methods lack.
Approach
Edits W+ latent codes of e4e-inverted faces (layers 6-9 only) via a password converter and HairCLIP-style modulation network, with three losses (single, sequential, associated anonymization) that enforce RSA-analogous properties.
Key results
- Recovery ability: FaceRSA reaches PSNR 75.51 and LPIPS 0.0099 versus RiDDLE's 67.61 and 0.0324 and FIT's 61.21 and 0.1780.
- Anonymization: FaceRSA achieves the best expression similarity (0.546) and pose similarity (0.996) among compared methods, though its ArcFace identity similarity (0.247) is not the lowest.
- FaceRSA is the only method in the comparison table that supports sequential anonymization/deanonymization, password interchangeability, and password associativity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FaceRSA instead of RiDDLE for face anonymization | Face anonymization · Face anonymization | Lmk: 0.013 -> 0.009 | Helped |
| Use FaceRSA instead of RiDDLE for face anonymization | Face anonymization · Face anonymization | Pose: 0.956 -> 0.996 | Helped |
| Use FaceRSA instead of RiDDLE for face anonymization | Face anonymization · Face anonymization | Exp: 0.314 -> 0.546 | Helped |
| Use FaceRSA instead of RiDDLE for face anonymization | Face anonymization · Face anonymization | ID: 0.115 -> 0.247 | Hurt |
| Use FaceRSA instead of RiDDLE for face recovery | Face recovery · Face recovery | LPIPS: 0.0324 -> 0.0099 | Helped |
| Use FaceRSA instead of RiDDLE for face recovery | Face recovery · Face recovery | PSNR: 67.61 -> 75.51 | Helped |
| Use FaceRSA instead of RiDDLE for face recovery | Face recovery · Face recovery | SSIM: 0.9987 -> 0.9996 | Helped |
| Use FaceRSA instead of RiDDLE for face recovery | Face recovery · Face recovery | ID: 0.8315 -> 0.9206 | Helped |
| Use FaceRSA instead of RiDDLE for sequential anonymization and deanonymization | FaceRSA · Sequential anonymization and deanonymization | recovery accuracy | Helped |
| Use FaceRSA instead of RiDDLE for password interchangeability | FaceRSA · Password interchangeability | original-image recovery | Helped |
| Use FaceRSA instead of RiDDLE for password associativity | FaceRSA · Password associativity | equivalent-password result | Helped |
Surprising findings
- Higher ArcFace identity similarity for FaceRSA is attributed to preserving identity-irrelevant attributes that ArcFace embeds, rather than weaker anonymization.
Limitations
- Security rests on ambiguity rather than computational hardness; no cryptographic guarantees, as the authors state.
- Only layers 6-9 of W+ are edited, chosen by manual layer analysis; generalization is unclear.
- Associated-password property is trained only for anonymization and generalizes to deanonymization empirically.
Experimental setup
Metrics: ArcFace identity similarity dlib face detection rate normalized landmark distance pose similarity expression similarity LPIPS MSE PSNR SSIM
Baselines: FIT RiDDLE CIAGAN DeepPrivacy IdentityDP
Builds on: StyleGAN e4e HairCLIP E2Style FIT
Closest papers by meaning
- CryptoFace: End-to-End Encrypted Face Recognition (CVPR 2025 · similarity 0.85)
- Adv-CPG: A Customized Portrait Generation Framework with Facial Adversarial Attacks (CVPR 2025 · similarity 0.84)
- Validating Privacy-Preserving Face Recognition under a Minimum Assumption (CVPR 2024 · similarity 0.84)
- Privacy-Preserving Face Recognition Using Trainable Feature Subtraction (CVPR 2024 · similarity 0.84)
- Privacy-Preserving Optics for Enhancing Protection in Face De-Identification (CVPR 2024 · similarity 0.83)
- FSFM: A Generalizable Face Security Foundation Model via Self-Supervised Facial Representation Learning (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.