Time flies by: Analyzing the Impact of Face Ageing on the Recognition Performance with Synthetic Data

Marcel Grimmer, Haoyu Zhang, Raghavendra Ramachandra, Kiran Raja, Christoph Busch

Abstract

The vast progress in synthetic image synthesis enables the generation of facial images in high resolution and photorealism. In biometric applications, the main motivation for using synthetic data is to solve the shortage of publicly-available biometric data while reducing privacy risks when processing such sensitive information. These advantages are exploited in this work by simulating human face ageing with recent face age modification algorithms to generate mated samples, thereby studying the impact of ageing on the performance of an open-source biometric recognition system. Further, a real dataset is used to evaluate the effects of short-term ageing, comparing the biometric performance to the synthetic domain. The main findings indicate that short-term ageing in the range of 1-5 years has only minor effects on the general recognition performance. However, the correct verification of mated faces with long-term age differences beyond 20 years poses still a significant challenge and requires further investigation.

Introduction

Biometric systems, particularly those deployed in high-stakes environments like border control (e.g., the European Entry-Exit System), must maintain high accuracy over extended periods. A common operational challenge involves matching a traveler against a passport photo taken a decade or more prior. However, studying the impact of ageing on these systems is fraught with difficulty.

Researchers face two primary hurdles: first, the scarcity of longitudinal datasets containing images of the same individual over 20 to 40 years; and second, the increasing complexity of processing real biometric data under strict privacy regulations like the GDPR. This paper investigates whether synthetic data—photorealistic images of non-existent individuals aged artificially—can serve as a reliable proxy for testing the robustness of state-of-the-art face recognition systems.

Methods

The researchers developed a multi-stage pipeline to generate and evaluate "mated" (same-identity) aged samples. The process began with StyleGAN and StyleGAN2 (trained on the FFHQ dataset) to create 50,000 high-resolution base images. These base images were then subjected to two different Face Age Modification (FAM) frameworks:

  • InterFaceGAN: This method identifies a "binary age boundary" within the GAN's latent space. By moving the latent vector perpendicular to this boundary, the system can continuously increase or decrease the subject's perceived age.
  • SAM (Style-based Age Manipulation): A more sophisticated approach that utilizes an age encoder to transform a base image into a specific target age (e.g., 10, 30, or 70 years old).
Overview of the face age modification pipeline using SAM and InterFaceGAN.

To ensure "operational realism," the generated images were filtered for poor illumination, extreme head poses, and low inter-eye distance (IED). Evaluation was conducted using ArcFace for recognition performance and two Face Quality Assessment Algorithms (FQAAs)—FaceQnet v1 and SER-FIQ—to verify the biometric utility of the synthetic samples.

Datasets

To provide a comprehensive analysis, the study utilized three distinct datasets:

  1. Synthetic Base: 50,000 images generated via StyleGAN as the foundation for ageing simulations.
  2. FRGC v2: A "bona fide" (real) dataset used as a reference for high-quality, controlled capture conditions.
  3. UNCW (MORPH-II): A real-world longitudinal dataset used to compare natural ageing effects against the synthetic results.

Key Insights and Analysis

The Synthetic Quality Match

A pivotal finding was that synthetic images now achieve biometric quality scores (via SER-FIQ) nearly identical to high-quality real-world datasets like FRGC v2. This suggests that synthetic data has reached a level of maturity where it can realistically substitute for real data in performance testing environments.

The Identity Preservation Paradox

While synthetic ageing algorithms are adept at preserving identity, they may be too consistent. The similarity scores for synthetic aged pairs were often higher than those found in real-world longitudinal data. This indicates a "domain gap" where GANs struggle to simulate the unpredictable ways real faces age, such as changes in skin texture, weight, or environmental factors, which naturally degrade similarity over time.

Non-Linearity and Training Bias

The study highlighted that moving a fixed distance in the GAN's latent space does not result in a linear increase in age; a single "step" might result in a 5-year jump for one face and a 15-year jump for another. Furthermore, the SAM model showed significant performance drops when generating ages for children or seniors, likely reflecting the demographic bias of the underlying FFHQ training dataset toward young-to-middle-aged adults.

Key Results

Recognition Performance

Both synthetic and real data confirmed that short-term ageing (1–5 years) has a negligible impact on recognition performance. However, for long-term ageing (20+ years), recognition performance using ArcFace dropped significantly. Mated comparison scores shifted toward the non-mated distribution, leading to a higher risk of False Non-Match Rates (FNMR).

Ageing Accuracy

The C3AE age estimator verified that the FAM algorithms successfully shifted the perceived age of the subjects. As shown in the figure below, SAM was particularly effective at hitting specific "target ages," while InterFaceGAN provided a more continuous, albeit less precise, age progression.

Predicted age distributions for different target ages using synthetic modification methods.

Limitations and Future Work

Despite the promise of synthetic data, several limitations remain. GAN inversion artifacts can still lead to "identity loss" when projecting real images into the latent space. Additionally, the current domain gap means synthetic data may overestimate a system's robustness because GANs keep facial features more stable than nature does. Future research will likely move toward 3D multi-view synthesis to better preserve spatial geometry, potentially closing the gap between synthetic and natural ageing simulations.

Summary updated

Abstract

The vast progress in synthetic image synthesis enables the generation of facial images in high resolution and photorealism. In biometric applications, the main motivation for using synthetic data is to solve the shortage of publicly-available biometric data while reducing privacy risks when processing such sensitive information. These advantages are exploited in this work by simulating human face ageing with recent face age modification algorithms to generate mated samples, thereby studying the impact of ageing on the performance of an open-source biometric recognition system. Further, a real dataset is used to evaluate the effects of short-term ageing, comparing the biometric performance to the synthetic domain. The main findings indicate that short-term ageing in the range of 1-5 years has only minor effects on the general recognition performance. However, the correct verification of mated faces with long-term age differences beyond 20 years poses still a significant challenge and requires further investigation.
BibTeX
@misc{grimmer2022time,
  author = {Marcel Grimmer and Haoyu Zhang and Raghavendra Ramachandra and Kiran Raja and Christoph Busch},
  title = {{Time flies by: Analyzing the Impact of Face Ageing on the Recognition Performance with Synthetic Data}},
  year = {2022},
  eprint = {2208.08207},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.08207},
}
Figure