arXiv paper: AEGIS aims to mask identity in animated 3D facial avatars

arXiv paper: AEGIS aims to mask identity in animated 3D facial avatars

Authors Dawid Wolkiewicz, Anastasiya Pechko, Przemysław Spurek and Piotr Syga present AEGIS, a method for hiding a 3D facial avatar’s identity from automated face recognition. They target 3D Gaussian Splatting avatars, where identity protection needs to hold across different viewpoints and animation, rather than only in a single 2D image. AEGIS adds adversarial perturbations to the base (DC) color coefficients of the Gaussian primitives, guided by an ensemble of pre-trained face verification networks. The authors say this embeds protection in the 3D representation without changing the avatar’s geometry or requiring retraining. In their evaluation against black-box recognizers, they report face retrieval and verification accuracy falling to near-zero while perceptual fidelity remains high. They also report preserving key facial attributes, including age, gender and emotion, and introduce a viewpoint-robust leakage metric to measure residual identity leakage across rendered viewpoints. The paper frames this as a response to identity theft risks in systems using biometric authentication. The supplied abstract does not give numerical evaluation results or identify the recognizers tested, so the reported performance cannot be assessed in detail from this evidence alone.

Publication: 2a70e9e9-a34b-4755-8d9c-75805b74d2c8

attribution — pass: The reading correctly names the four listed authors and keeps the method, results and identity-theft framing attributed to them ('they report', 'the authors say'). It identifies the source as an arXiv paper, which matches the page.

Authors:Dawid Wolkiewicz, Anastasiya Pechko, Przemysław Spurek, Piotr Syga

summary — pass: The title, summary and question faithfully convey the abstract: what AEGIS does, how it works, the reported results, the new metric and the motivation. It opens with a plain-language goal (hiding an avatar's identity from automated face recognition) that a general reader can follow. Technical terms such as 3D Gaussian Splatting, adversarial perturbations and DC color coefficients are kept exactly, in line with the term-tooltip policy. The question addresses a real gap left by the abstract.

concealing identity from automated face recognition while preserving the avatar's perceptual realism and animation usability

date — pass: originalPublished is 2025-11-21. This matches the v1 submission date. It is not the later v2 revision date (4 Nov 2026) or the feed date.

Submitted on 21 Nov 2025 (v1), last revised 4 Nov 2026 (this version, v2)

support — pass: Each claim, including no geometry change or retraining, near-zero accuracy, attribute preservation and the leakage metric, is stated in the abstract and presented as the authors' report. The limitations about missing numbers and unnamed recognizers are accurate. No benefits or impacts are invented beyond the source's framing.

without retraining or modifying the avatar's geometry. Evaluated against black-box recognizers, AEGIS reduces face retrieval and verification accuracy to near-zero

media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source, and it is the only supplied candidate.