Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation
Authors: Carlos Soto, Matthew Reimherr, Aleksandra Slavkovic, Mark Shriver
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that our method preserves the shape of the average face and injects less noise than traditional methods for the same privacy budget. Our mechanism consists of two primary components, the first is generally applicable to function value summaries (as are commonly found in nonparametric statistics or functional data analysis) while the second is general to disk-like surfaces and hence more applicable than just to human faces. Here, we apply our proposed methods from sections 3 and 4.2 on data described in Sero et al. (2019) and A.4. To highlight the difficulty of working with such complex data we present some preliminary results. We have 1000 faces each with 7150 points; we fully describe the data in A.4. Lastly, we quantify the amount of injected noise by computing a mean squared error. We demonstrate via empirical and quantitative results that our methodology adds less noise and preserves the structure of the face. |
| Researcher Affiliation | Academia | Carlos Soto Department of Mathematics and Statistics University of Massachusetts Amherst Amherst, MA 01003, USA EMAIL Matthew Reimherr Department of Statistics Pennsylvania State University University Park, PA 16802, USA EMAIL Aleksandra Slavkovic Department of Statistics Pennsylvania State University University Park, PA 16802, USA EMAIL Mark Shriver Department of Anthropology Pennsylvania State University University Park, PA 16802, USA EMAIL |
| Pseudocode | No | The paper describes mathematical methodologies and processes in descriptive text and equations, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper mentions external software and accompanying implementations at GitHub repositories (e.g., 'Jermyn et al. (2017), with accompanying implementation at Git Hub repository (Laga, 2022)' and 'Gary P.T. Choi. disk-conformal-map. https://github.com/garyptchoi/ disk-conformal-map, 2020.'), which are tools used by the authors. However, it does not provide any statement or link for the open-sourcing of the novel methodology or specific code developed by the authors for this paper. |
| Open Datasets | Yes | We apply our proposed method to data described in Sero et al. (2019); more details are available in A.4. For a full description of how the data is collected we refer to Sero et al. (2019) but we summarize and emphasize the relevant points. |
| Dataset Splits | No | The paper uses a dataset of '1000 faces each with 7150 points' but does not specify any training, validation, or test splits. The experimental focus is on generating a private average face and quantifying noise, not on a machine learning model that typically requires explicit dataset splits for evaluation. |
| Hardware Specification | Yes | We ran experiments almost exclusively on a desktop computer with an Intel i7 processor and 32GB of RAM on Windows 11. Reparameterizing, however, each face to a template costs upwards of 3 minutes on the desktop, so since we have 1000 faces we did this on a supercomputer. |
| Software Dependencies | No | The paper mentions 'Mesh Lab software (Cignoni et al., 2008)' and 'Windows 11' but does not provide specific version numbers for any key software libraries, frameworks, or programming languages used in their implementation. |
| Experiment Setup | Yes | We represent the n = 1000 faces with J = 23 face radial curves. The left panel of Figure 4 displays the mean face constructed from mean face radial curves with ϕx = ϕy = 0.01 and ϕz = 0.005. an example private mean face with µT = 2.9661 where µT = qJ(µ2x + µ2y + µ2z) with J = 23, µx = µy = 0.2 and µz = 0.55. For each coordinate we have separate smoothing parameters we display results with ϕx = ϕy = 0.01 and ϕz = 0.005. |