Template Ageing and Longitudinal Verification in Fixed-Text Keystroke Dynamics: A Subject-Disjoint Study Across Eight Weeks
Authors: Simon Parkinson, Saad Khan, Na Liu, Qing Xu
Organizations: Department of Computer Science, University of Huddersfield, HD1 3DH Huddersfield, U.K. · College of Intelligence and Computing, Tianjin University, Tianjin, China
Behavioural biometric templates are widely believed to degrade as the gap between enrolment and verification grows, but few studies measure this template ageing effect directly under controlled conditions. We collected a longitudinal dataset of 40 fixed passwords, each typed four times per weekly session over eight consecutive weeks. We compare a scaled-Manhattan matcher (M1), a gradient-boosted classifier (M2), a TypeNet-style recurrent embedding model (M3), and a TypeFormer-style Transformer (M4) under a 5-fold subject-disjoint protocol and a design that jointly varies mechanism and the enrolment-to-query gap, from 0 to 7 weeks. Template ageing proves large and systematic. Error increases monotonically with the gap for every mechanism, from an EER of 14.6-27.2% at a gap of zero to 25.5-37.1% at seven weeks, or 1.7% of decision error per week elapsed (p < 0.001). However, the choice of mechanism matters more than its rate of ageing. Baseline accuracy spans 12.6 percentage points across the four mechanisms, the degradation each accumulates over seven weeks spans only 2.3 points, and ageing never reorders them. A matcher can therefore be chosen on same-session accuracy, with ageing managed by re-enrolment scheduling rather than by matcher selection. The two properties are nonetheless distinct, as M3 is the least accurate mechanism yet ages significantly more slowly than M1 under every specification tested. Training randomness also matters differently by architecture, with 58% of the recurrent model's fold-to-fold variance attributable to seed noise against 19% for the Transformer. Because the smaller ageing-rate differences are sensitive to modelling choices, while the accuracy differences and the ageing effect are not, we recommend that comparative ageing-rate claims be supported by seed-level score fusion, independent replication, and an alternative outcome-model specification.
Figures & tables
Study
Longitudinal (weeks+)
Cross-week template test
Learned embedding model
Subject-disjoint evaluation
Inferential statistics
Montalvão et al. [ 27 ]
×
×
×
×
×
Acien et al. (TypeNet) [ 9 ]
× (free-text)
×
✓
✓
×
Parkinson et al. [ 12 ]
✓
×
×
×
×
Giot/Pisani (template update) [ 6 , 7 ]
∼ (multi-session)
∼ (adaptation active)
×
×
×
Yang et al. [ 8 ]
✓ (drift-detection)
×
×
×
×
This paper
✓
✓
✓
✓
✓
TABLE I: Positioning relative to selected prior work (qualitative; ✓ = present, × = absent, ∼ = partially addressed).
TABLE II: Architecture and optimisation hyperparameters for M3 (LSTM) and M4 (Transformer), fixed a priori and shared across folds and replicates.
Fig. 1: EER versus enrolment-to-query gap Δ (weeks), by mechanism and outlier-clipping condition, at k=5 enrolment sessions. Shaded bands are bootstrap 95% confidence intervals over test participants. Error increases monotonically with Δ for every mechanism.
Mechanism
EER at Δ=0
EER at Δ=7
M1 (scaled Manhattan)
21.0 [19.7, 22.4]
33.1 [30.1, 36.0]
M2 (shallow ML)
17.2 [15.9, 18.5]
27.0 [23.7, 30.5]
M3 (TypeNet-style LSTM)
27.2 [25.6, 28.8]
37.1 [34.7, 39.5]
M4 (TypeFormer-style)
14.6 [13.5, 15.9]
25.5 [21.8, 29.8]
Trials (all mechanisms): genuine / impostor / total
Δ=0
23,827 / 352,560 / 376,387
TABLE III: EER (%) at Δ=0 and Δ=7 ( k=5 , clipped, M3/M4 three-seed fusion, with 95% bootstrap CI in brackets), together with the underlying genuine/impostor trial counts (identical across mechanisms).
Fig. 2: DET curves at k=5 , pooled across Δ , for the outlier-clipped (left) and unclipped (right) conditions.
Mechanism
FNMR @ FMR=1%
FNMR @ FMR=0.1%
M1 (scaled Manhattan)
91.5
98.6
M2 (shallow ML)
79.7
93.9
M3 (TypeNet-style LSTM)
95.2
99.3
M4 (TypeFormer-style)
86.4
97.8
TABLE IV: FNMR (%) at fixed FMR operating points, k=5 , clipped, pooled across Δ .
Term
Coef.
z
p
Δ (per week, centred)
+0.017
21.60
<0.001
Mechanism: M2 vs. M1
−0.040
−18.30
<0.001
Mechanism: M3 vs. M1
+0.052
23.90
<0.001
Mechanism: M4 vs. M1
−0.065
−30.15
<0.001
Length (per character, centred)
−0.020
−11.27
<0.001
Hardware-change (week 5)
+0.005
2.08
0.037
TABLE V: Mixed-effects model: selected fixed-effect estimates (main-effects model, no interaction). N=299,984 , with participant and password variance components of 0.003 and 0.001 respectively.
Linear probability model
Logistic (cluster-robust)
Mechanism vs. M1
Run 1
Run 2
Run 3
Run 1
Run 2
Run 3
M2
−0.002 (.029)
−0.003 (.010)
−0.002 (.030)
−0.004 (.652)
−0.004 (.650)
+0.003 (.713)
M3
−0.002 (.047)
−0.004 ( < .001)
−0.004 (.001)
−0.019 (.032)
−0.021 (.016)
−0.015 (.056)
M4
−0.000 (.678)
−0.001 (.485)
−0.000 (.954)
+0.016 (.093)
+0.016 (.090)
+0.024 ( .008 )
TABLE VI: Δ× Mechanism interaction under the linear-probability and logistic (cluster-robust) specifications, across three independent end-to-end pipeline replicates. Cells show the coefficient, with the p -value in parentheses. In the linear specification, coefficients are expressed on the weekly probability-of-error scale per unit of Δ , whereas in the logistic specification, coefficients are measured in log-odds. Consequently, only the direction and statistical significance of the coefficients, rather than their magnitudes, are directly comparable across the two model formulations. For the joint test in the linear model, the test statistics are χ2(3)=7.27,17.92,and 16.22 , with corresponding p -values of 0.064,0.0005,and 0.0010 for Runs 1, 2, and 3, respectively.
Mechanism
Within-fold seed SD
Across-fold SD
Noise share
M3 (LSTM)
0.0181
0.0313
∼ 58%
M4 (Transformer)
0.0062
0.0325
∼ 19%
TABLE VII: Decomposition of fold-to-fold EER variance into training-seed noise (same data, different training run) versus total across-fold spread.
Δ
Genuine trials
Impostor trials
Total
0
23,827
352,560
376,387
1
79,994
293,714
373,708
2
68,527
249,512
318,039
3
56,387
203,999
260,386
4
45,332
163,875
209,207
5
34,393
123,530
157,923
TABLE VIII: Genuine and impostor trial counts by Δ ( k=5 , clipped; identical across mechanisms).
Mechanism
Condition
Δ=0
Δ=1
Δ=2
Δ=3
Δ=4
Δ=5
Δ=6
Δ=7
M1 (scaled Manhattan)
Clipped
21.0
24.2
25.7
27.0
28.3
29.3
30.7
33.1
Unclipped
24.0
26.3
27.7
28.8
30.0
31.2
32.2
34.3
M2 (shallow ML)
Clipped
17.2
20.6
22.1
23.5
24.5
25.2
25.6
27.0
Unclipped
19.8
23.1
24.5
25.9
27.0
27.6
27.8
28.6
M3 (TypeNet-style LSTM)
Clipped
27.2
29.0
30.5
32.1
33.4
34.0
35.2
37.1
Unclipped
21.4
24.1
25.7
27.3
28.9
30.1
31.8
34.1
TABLE IX: Full EER (%) by mechanism, clipping condition, and temporal gap Δ , k=5 .
Mechanism
k
Δ=0
Δ=1
Δ=2
Δ=3
Δ=4
Δ=5
Δ=6
Δ=7
M1 (scaled Manhattan)
1
28.9
31.4
32.3
33.2
34.0
34.6
35.0
36.5
3
21.9
25.6
26.5
27.9
29.2
30.0
30.6
33.1
5
21.0
24.2
25.7
27.0
28.3
29.3
30.7
33.1
M2 (shallow ML)
1
15.8
20.5
22.0
23.3
24.2
24.8
24.7
26.5
3
16.1
20.7
22.1
23.5
24.8
25.2
25.6
26.9
5
17.2
20.6
22.1
23.5
24.5
25.2
25.6
27.0
TABLE X: Full EER (%) by mechanism, enrolment size k , and temporal gap Δ , clipped condition.
Mechanism
Estimator
Δ=0
Δ=1
Δ=2
Δ=3
Δ=4
Δ=5
Δ=6
Δ=7
M1 (scaled Manhattan)
Pooled
21.0
24.2
25.7
27.0
28.3
29.3
30.7
33.1
Particip.-balanced
20.9
24.2
25.6
26.7
28.2
29.4
30.8
32.2
M2 (shallow ML)
Pooled
17.2
20.6
22.1
23.5
24.5
25.2
25.6
27.0
Particip.-balanced
16.2
19.2
20.7
21.9
22.9
23.8
24.2
26.0
M3 (TypeNet-style LSTM)
Pooled
27.2
29.0
30.5
32.1
33.4
34.0
35.2
37.1
Particip.-balanced
27.0
28.6
30.2
31.6
32.8
33.7
34.6
37.7
TABLE XI: EER (%) at k=5 , clipped: pooled-trial vs. per-participant-balanced estimator, all eight values of Δ .
Fold
Mechanism
Seed 0
Seed 1
Seed 2
0
M3
29.7
35.2
29.6
M4
20.6
20.8
21.5
1
M3
31.0
34.5
33.7
M4
18.9
17.7
17.2
2
M3
35.5
32.5
32.6
M4
18.6
19.4
21.0
TABLE XII: Pooled EER (%) for three independently-trained (unfused) seeds of M3 and M4, by test fold.
When a neural network estimates someone's age from a photograph, does it process biometric data? The answer depends on whether identity-discriminative representations arise within the network during inference, a question that may seem trivial to ML researchers but triggers consent requirements under GDPR, statutory damages under BIPA, or high-risk AI classification under the EU AI Act. Yet no regulatory guidance addresses it. This position paper provides empirical evidence: 14 models evaluated across 3 face verification benchmarks show age estimators fall orders of magnitude short of identification thresholds. Age estimation models cannot identify individuals. We call on researchers to provide transparency about what systems store and can do, and on regulators to distinguish transient processing from template storage.
Electrocardiogram (ECG) biometrics have emerged as a promising modality for continuous, liveness-aware authentication in wearable systems. However, many prior studies report overly optimistic results due to data leakage (e.g., random splits within the same session). To address this issue, we introduce ECG-biometrics-bench, a modular, reproducible benchmarking framework that standardizes preprocessing, segmentation, and evaluation across seven widely used public ECG datasets spanning clinical, ambulatory, and large-scale cohort settings. The framework supports both closed-set and open-set (i.e., subject-disjoint generalization in this work) evaluation, as well as progressively realistic protocols including cross-session and long-term temporal separation. To facilitate reproducible research in the community, the ECG-biometrics-bench repository will be made publicly accessible on GitHub upon the acceptance of this manuscript. Through a comprehensive multi-dataset analysis, we expose the Random Split Fallacy, demonstrating that intra-session evaluation protocols artificially inflate performance while masking severe degradation caused by temporal drift and unseen identities. Furthermore, by evaluating multiple architectures, including DeepECG, ResNet1D, and CNN-LSTM, we show that these failures are not model-specific but are likely inherent to current supervised feature-learning paradigms. Finally, we demonstrate that performance degradation due to temporal aging can be partially mitigated through a heavy enrollment, lightweight authentication strategy based on dynamic multi-session template fusion. These findings establish a more realistic baseline for ECG biometrics and highlight critical challenges that must be addressed for reliable real-world deployment.
Online platforms increasingly rely on automated age estimation systems to enforce minimum-age policies. Focusing on vision-based models designed for this task, concerns arise regarding their robustness to simple appearance changes that underage individuals may use to bypass such systems, such as drawing a mustache or applying lipstick. In this work, we present a systematic study of age verification robustness by simulating visual alterations that can be easily achieved by underage individuals. We evaluate seven models, including vision, vision-language, and multimodal large language models, across three datasets and four manipulation types. Interestingly, under drawn beard stubble, up to 61% of True Negatives are flipped into False Positives. Furthermore, we investigate how different demographics are affected by such manipulations, finding that Indians are more affected by beard stubble manipulations, while females are more affected than males across all manipulations. Finally, we explore how these biases can be mitigated using bias mitigation methodologies in lightweight linear probe settings.
Ioannis Sarridis, Ioannis Kompatsiaris, Symeon Papadopoulos
Information Technologies Institute, Centre for Research and Technology Hellas, 6th km Charilaou-Thermi Rd, Thessaloniki, 57001, Greece