A Deeper Dive into the Irreversibility of PolyProtect: Making Protected Face Templates Harder to Invert

Problem
This paper addresses the gap in the literature regarding the irreversibility of PolyProtect, a biometric template protection method for face embeddings. The authors highlight that previous analyses did not sufficiently explore the ease of inverting PolyProtected templates, particularly using different distance metrics. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution involves a detailed analysis of the irreversibility of PolyProtect templates, which utilize multivariate polynomials to transform face embeddings into protected templates. The authors introduce a “key selection algorithm” designed to optimize the selection of polynomial coefficients and exponents, enhancing the irreversibility of the generated templates. The algorithm aims to select keys that are not purely random, thereby improving the difficulty of inversion. The study also examines the impact of the overlap parameter in the polynomial application, which determines the degree of shared elements between consecutive sets of the embedding. The authors propose normalization of embeddings prior to applying PolyProtect to mitigate accuracy issues related to the range of embedding values.

Results
The authors demonstrate that templates generated using the key selection algorithm are significantly harder to invert compared to those created with random keys. Specifically, they report a marked improvement in the irreversibility of PolyProtected templates across various overlap parameters, effectively equalizing their irreversibility. The paper quantifies this improvement, although specific numerical results and effect sizes are not detailed in the abstract. The authors also note that normalization of embeddings leads to enhanced accuracy in the PolyProtected domain, although exact performance metrics against named baselines are not provided.

Limitations
The authors acknowledge that their analysis primarily focuses on the irreversibility aspect of PolyProtect and does not extensively cover other potential vulnerabilities or performance metrics. They also do not address the computational overhead introduced by the key selection algorithm or the implications of varying the overlap parameter beyond irreversibility. Additionally, the reliance on a specific numerical solver for inversion analysis may limit the generalizability of their findings.

Why it matters
This work has significant implications for the field of biometric security, particularly in enhancing the robustness of face recognition systems against inversion attacks. By improving the irreversibility of PolyProtected templates, the proposed methods could lead to more secure biometric systems, thereby reducing the risk of unauthorized access and identity theft. The findings also contribute to the broader discourse on the trade-offs between security and accuracy in biometric template protection, providing a foundation for future research in optimizing these parameters. The open-source code provided by the authors facilitates reproducibility and encourages further exploration of the proposed methods.

Authors: Vedrana Krivokuća Hahn, Jérémy Maceiras, Sébastien Marcel
Source: arXiv:2605.03857
URL: https://arxiv.org/abs/2605.03857v1