Introduction
Biometric systems, particularly face verification, must contend with significant intra-class variation caused by factors such as aging, lighting changes, and facial expressions. To maintain accuracy over time, many systems employ "self-updating" or semi-supervised learning strategies, where successful verification samples are automatically added to a user's template gallery. However, this process introduces the risk of "creep-in" errors—instances where an impostor is incorrectly accepted and their biometric data is integrated into a legitimate user's profile, permanently compromising security. This paper investigates whether this risk is uniform across populations or if specific user types are more susceptible to these errors.
The Doddington’s Zoo Classification
The researchers utilize the "Doddington’s Zoo" framework to categorize users based on their interaction with biometric systems:
- Sheeps: Ideal users who are easy to recognize and difficult to imitate.
- Goats: Users who are intrinsically difficult for the system to recognize (high False Rejection Rate).
- Lambs: Users whose biometric features are easily imitated by others (high False Acceptance Rate).
- Wolves: Users who are exceptionally skilled at imitating others.
Methodology
The study employed the Elastic Bunch Graph Matching (EBGM) algorithm for face verification on the Equinox Corporation Database. This dataset includes 56 individuals with 129 frontal face images each, captured over a one-year span with varying illumination and expressions. The protocol involved three distinct phases:
- Partitioning: Using a "prediction" subset to label users as Sheep, Goat, Lamb, or Wolf based on their initial error rates.
- Self-Updating: Running the automated update procedure on an "unlabelled" set of images to simulate real-world system evolution.
- Evaluation: Testing the performance of the updated templates on a final "test" set to measure the impact of the updates on each group.
The Vulnerability of Lambs and the Creep-in Phenomenon
A critical finding of the study is the disproportionate risk faced by "Lambs." Because these users are inherently easy to imitate, the self-updating mechanism frequently misidentifies impostors as the genuine user, leading to a rapid degradation of security.
For Lambs, approximately 43.2% of the samples added to their galleries during the update process were actually impostors. This high rate of intrusion makes self-updating counter-productive for this group, as it effectively trains the system to accept unauthorized users, as evidenced by the lack of improvement in the False Rejection Rate (FRR) relative to the False Acceptance Rate (FAR).
Comparative Results Across the Zoo
The impact of template updating varies significantly across the four categories. While the majority of the population (Sheeps and Goats, totaling 86%) can benefit from or remain safe during updates, the remaining 14% (Lambs and Wolves) represent a significant security vulnerability.
- Sheeps (56% of population): Experienced the most significant performance gains. Approximately 72.6% of available samples were utilized for updates with a relatively low impostor introduction rate of 16.4%.
- Goats (30% of population): Remained relatively safe (11.8% impostor introduction) but benefited less from updates, as the system only accepted 23.3% of their genuine variations due to their high intrinsic rejection rate.
- Wolves (2% of population): Exhibited a high intrusion rate (37%), suggesting that those who can easily imitate others may also be easily imitated themselves, leading to similar "creep-in" risks as Lambs.
Conclusion and Future Directions
The study demonstrates that a uniform, one-size-fits-all approach to biometric template updating is insufficient and potentially dangerous. The "creep-in" effect for Lambs and Wolves necessitates personalized update protocols. Future biometric systems should consider user-specific thresholds—stricter for Lambs and more relaxed for Goats—or implement supervised update cycles for high-risk groups to prevent security degradation. Identifying a user's "animal type" dynamically could allow systems to adapt their update logic in real-time, balancing convenience with robust security.
Summary updated