Research map / Federated learning and privacy / Differential privacy
Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation
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.
Releases a Gaussian differentially private average 3D human face by representing faces as disk-parameterized surfaces, extracting registered face radial curves, and adding RKHS-based Gaussian process noise to their penalized mean.
Problem
Releasing average 3D human face data with formal differential privacy guarantees is hard because faces are high-dimensional, structured surfaces tied to identity, and point-wise noise distorts facial structure.
Approach
Extends approximate-DP functional data mechanisms (Gaussian process noise in an RKHS) to Gaussian DP, and applies them to a novel face radial curve representation derived from elastic shape analysis of disk-parameterized faces, so that noise is injected into a few dozen registered curves rather than thousands of points.
Key results
- Face radial curve GDP mechanism (mu_T=2.9961) achieves MSE 5.2989e-4 versus 14e-4 for point-wise sanitization at mu_T=2 and 7.5807e-4 at mu_T=3.
- Private mean from the proposed method has MSE 3.6365e-4 relative to the RKHS mean, less than half the point-wise sanitization error.
- Using 23 face radial curves with 81 points each (1863 points) represents the 7150-point face.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GDP face radial curves instead of point-wise Gaussian sanitization | Private average 3D face · 3D human faces | MSE: 7.5807 -> 5.2989 (all values at E-04) | Helped |
| Use GDP face radial curves instead of point-wise Gaussian sanitization | Private average 3D face · 3D human faces | MSE: 14 -> 5.2989 (all values at E-04) | Helped |
Limitations
- Sensitivity is computed from the data, so the calculation is not itself private.
- Only tested on one clean, neutral-expression dataset.
- Requires genus-0 surfaces, so missing data must be preprocessed.
Experimental setup
Datasets: 1000 3D face scans with 7150 quasi-landmarks (neutral expression)
Metrics: mean squared error vs. point-wise mean MSE vs. RKHS mean
Baselines: point-wise Gaussian mechanism on mean face
Builds on: Gaussian differential privacy Hall-Rinaldo-Wasserman functional DP elastic shape analysis square-root normal field
Closest papers by meaning
- Privacy-Preserving Face Recognition Using Trainable Feature Subtraction (CVPR 2024 · similarity 0.84)
- How Private is Private? A Comparative Study for Face De-Identification (NeurIPS 2026 · similarity 0.84)
- Validating Privacy-Preserving Face Recognition under a Minimum Assumption (CVPR 2024 · similarity 0.84)
- Disguise without Disruption: Utility-Preserving Face De-identification (AAAI 2024 · similarity 0.84)
- Non-Colliding Biometric Identities for Digital Entities: Geometry, Capacity, and Million-Scale Virtual Identity Provisioning (NeurIPS 2026 · similarity 0.84)
- Do Not DeepFake Me: Privacy-Preserving Neural 3D Head Reconstruction Without Sensitive Images (AAAI 2025 · similarity 0.83)
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.