Simple changes to facial appearance can alter the decisions of face-based age-verification models, according to a new study that tested simulated beard stubble, mustaches and cosmetics across seven artificial intelligence models.
Researchers at Greece’s Centre for Research and Technology Hellas evaluated FairFace, MiVOLO, ViT-Age, SigLIP2, MetaCLIP2, Qwen2.5-VL and LLaVA-1.6 on three widely used face datasets. They converted each model’s age estimate into a pass-or-fail decision using a threshold of 13, then measured how often a correctly rejected image of an underage subject changed to age-eligible after an appearance modification.
Digitally drawn beard stubble had the strongest overall effect. It produced an average flip rate of 20.8 percent across the models and datasets, rising to 61.4 percent for MiVOLO on FairFace. Lipstick averaged 9.9 percent, eye makeup 4.8 percent and mustaches 2.5 percent. LLaVA-1.6 was the least affected model in most test conditions.
A separate experiment using Stable Diffusion 2 to generate more realistic modifications produced larger changes. Generated beards pushed the average flip rate to 36.3 percent and increased the estimated age by 6.5 years, compared with a 2.9-year increase for the simpler OpenCV beard overlay.
The effects also varied across demographic groups in the FairFace dataset. Female subjects had higher model-averaged flip rates than male subjects for all four modifications, while Indian subjects had the highest average rates for beard stubble and eye makeup. The authors said robustness and demographic differences should be evaluated alongside standard age-estimation accuracy.
The researchers also tested lightweight mitigation methods using manipulated examples during training. Their strongest method substantially improved worst-group performance with manipulated images making up one percent of the underage training samples, suggesting that robustness can improve without retraining an entire model.
The results do not establish that a physical disguise will defeat a deployed product. All modifications were applied digitally to benchmark images, and the models were tested as visual components in isolation. Commercial services may combine facial estimates with account, device or network signals, and the study did not test those wider systems.
The paper is a preprint submitted to arXiv and has not undergone peer review. Its findings arrive as facial age estimation moves into retail and online services, while technical groups continue to examine the security and privacy tradeoffs of age-verification architectures.
Sources: arXiv
–
By Cass Kennedy
