PaperlayerView exact revision on arXiv

Source-linked AI summary

Template Ageing and Longitudinal Verification in Fixed-Text Keystroke Dynamics: A Subject-Disjoint Study Across Eight Weeks

Simon Parkinson, Saad Khan, Na Liu, Qing Xu

arXiv:2609.29851v1cs.CRcs.AI

TL;DR

The paper addresses limited direct evidence on how fixed-text verification error changes with the enrolment-to-query gap and whether that change differs by mechanism. It compares four mechanisms under a jointly varying trial design and finds error rises across all, while comparative ageing-rate claims are less robust than baseline accuracy differences.

  • Problem

    Prior work largely managed template ageing rather than measuring it directly, and longitudinal fixed-text evidence for embedding models and robust mechanism comparisons was limited.

    Source: §I · p. 1

  • Method

    The study compares four matching mechanisms in a unified design that varies mechanism and enrolment-to-query gap, including recurrent and Transformer embedding models.

    Source: §I · p. 2, C. Embedding Model Architecture · p. 3

  • Results

    Verification error increases monotonically and significantly with the gap for each mechanism, while baseline accuracy differs significantly and M4 and M2 outperform M1 and M3.

    Source: §V · p. 8

  • Takeaways & Limitations

    Only M3’s slower ageing rate survives the reported robustness checks, so comparative ageing-rate claims warrant checks across training seeds and outcome-model specifications.

    Source: §VI · p. 10

    The 85-participant pool comes from one institution, lacks demographic detail, and uses a shared password corpus rather than user-chosen deployment passwords.

    Source: B. Limitations · p. 9

Abstract

from arXiv · show

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.

I. INTRODUCTION

The introduction motivates direct measurement of template ageing and reliable comparisons of matching mechanisms over time. It presents a longitudinal, jointly controlled comparison and methods for testing repeated-trial effects and result robustness.

  • Keystroke dynamics captures key-press and key-release timing as users type and is proposed as a low-cost biometric add-on to password systems.Password-based authentication remains dominant despite tensions between complex-password security benefits and usability costs.Source: §I · p. 1
  • Because templates may be used for weeks or months without re-enrolment, ageing rates could change which mechanism is most reliable over time.Same-session accuracy alone may not reveal prolonged-use reliability.Source: §I · p. 1
  • Four gaps include limited direct measurement of ageing, no longitudinal fixed-text evaluation of TypeNet-style models, infrequent mixed-effects analysis, and single-run comparisons.Prior work has focused more on managing ageing than isolating it, while the recurrent and Transformer embedding models had been demonstrated on free text rather than longitudinal fixed-text passwords.Source: §I · p. 1
  • The study measures fixed-text verification degradation week by week across an eight-week enrolment-to-query span.The dataset described here comprises participants typing 40 passwords four times per weekly session over eight consecutive weeks.Source: §I · pp. 1–2
  • A shared trial design compares four mechanisms across enrolment-to-query gaps, separating accuracy, ageing-rate, and interaction differences.The mechanisms include a distance-threshold matcher, a gradient-boosted classifier, and recurrent and Transformer embedding models.Source: §I · p. 2
  • Comparative ageing claims require checking both training-run randomness and sensitivity to an alternative outcome-model specification.The study also accounts for repeated trials from the same participants and passwords, reporting effect sizes and confidence intervals.Source: §I · p. 2

II. RELATED WORK

Related work spans keystroke features, matching models, longitudinal password studies, and biometric template adaptation. The paper distinguishes measuring unmitigated ageing from managing it and limits its claim to short-term behavioural drift.

A. Keystroke Dynamics Fundamentals

  • Fixed-text systems use dwell, flight/digraph, and trigraph timing features; no single feature set is universally optimal across passwords or users.A proposed user-specific variability metric complements the study’s participant-level random effect by modelling that variability rather than predicting a per-template score.Source: A. Keystroke Dynamics Fundamentals · p. 2

B. Matching Mechanisms

  • Learned metric-learning embeddings have advanced keystroke matching, but TypeNet and TypeFormer had not been evaluated for longitudinal fixed-text passwords.The evaluated landscape extends beyond these models, while this study uses one recurrent and one attention-based model under a shared triplet objective.Source: B. Matching Mechanisms · p. 2

C. Longitudinal Effects and Password Characteristics

  • Extended multi-week studies combining password length, substitutions, and repetition remain rare; common benchmarks typically use few sessions.Earlier work examined repetition and template generalisability within individual weeks.Source: C. Longitudinal Effects and Password… · p. 2

D. Template Ageing and Adaptive Biometric Systems

  • Adaptation studies mitigate template degradation, whereas raw ageing measurement estimates unmitigated change over time and by mechanism.The paper frames raw measurement as a prerequisite for benchmarking adaptive schemes; its eight-week result concerns short-term behavioural drift, not years-long physiological change.Source: D. Template Ageing and Adaptive… · pp. 2–3

E. Statistical Methodology in Biometric Performance Evaluation

The section motivates methods that account for repeated observations and describes the study’s dataset and input representations. Its design uses participant/password-aware approaches to address non-independence in biometric measurements.

E. Statistical Methodology in Biometric Performance Evaluation

  • Participant-level bootstrap and participant/password random effects offer complementary responses to correlated scores that can make naive confidence intervals overstate precision.Linear mixed-effects models provide standard errors for repeated-measures data, a setting for which reporting standards specify metrics but not the analysis approach.Source: E. Statistical Methodology in… · p. 3

A. Dataset and Input Representation

  • The dataset contains 85 participants, 40 passwords, four repetitions per session, and eight weekly sessions; inputs use raw event sequences or aggregated timing features.After cleaning, 94,894 of 99,607 attempts remained; missing participant-weeks were not imputed. The aggregated representation concatenates timing and trigraph features into a 6n−6-dimensional vector, with clipping fitted on training participants and applied unchanged to validation and test data.Source: A. Dataset and Input Representation · p. 3

B. Subject-Disjoint Experimental Protocol

The protocol evaluates models on participants unseen during training and compares mechanisms and time gaps using shared trials. Recurrent and Transformer embeddings use a common triplet objective, with the gap recorded for each genuine and impostor comparison.

B. Subject-Disjoint Experimental Protocol

  • Five-fold subject-level cross-validation holds out participants for testing, while validation handles early stopping and threshold calibration.All passwords appear in every split, and fold variance is reported rather than averaged away.Source: B. Subject-Disjoint Experimental… · p. 3

C. Embedding Model Architecture

  • The recurrent model maps event sequences to embeddings with triplet loss, while the Transformer uses Gaussian range encoding under the same objective.The embedding gallery is a centroid of enrolment sessions and verification uses negative Euclidean distance; hyperparameters were fixed a priori, not tuned to reported metrics.Source: C. Embedding Model Architecture · p. 3, Table II · p. 4

D. Unified Verification Protocol

  • Table I qualitatively positions the study relative to selected prior work using present, absent, and partially addressed categories.Source: Table I · p. 4
  • Each trial compares a test participant’s enrolment gallery with their later query or other participants’ sessions, recording the gap from zero to seven weeks.The resulting mechanism-by-gap EER surface permits estimating ageing, mechanism, and interaction effects from the same trial set.Source: D. Unified Verification Protocol · pp. 3–4

E. Baselines for Comparison

The study compares four matching mechanisms, reports pooled error metrics with participant-level bootstrap intervals, and models trial-level decision errors using mixed effects. The model tests whether the mechanisms differ in how error changes with the enrolment-to-query gap.

E. Baselines for Comparison

  • Four mechanisms are compared: a scaled-Manhattan matcher, a gradient-boosted classifier, a recurrent embedding model, and a Transformer embedding model.M2 is trained per password, unlike M3 and M4, making its task easier and limiting that comparison.Source: E. Baselines for Comparison · p. 4

F. Evaluation Metrics and Reporting Standards

  • EER and FNMR are reported for each mechanism and gap; EER pools trials, while confidence intervals bootstrap participants within each fold.The pooled convention gives participants or passwords with more trials proportionally more influence; a balanced re-analysis shifted cells by at most 1.7 points without changing mechanism ordering.Source: F. Evaluation Metrics and Reporting… · p. 4

G. Statistical Modelling

  • Trial-level binary errors are modelled with a linear mixed-effects model, using participant and password random intercepts and a fixed fold effect.The central interaction tests whether M3 and M4 degrade at different rates than M1; a logistic sensitivity check changes two of the three interaction conclusions.Source: G. Statistical Modelling · p. 5

H. Training-Seed Robustness Check

The paper checks training-seed variability by retraining M3 and M4 on identical fold data, then examines performance across enrolment-to-query gaps. EER rises with the gap for every mechanism, while the reported low-FMR FNMR is high across mechanisms.

H. Training-Seed Robustness Check

  • M3 and M4 are retrained three times per fold on identical data to estimate within-fold seed variability against across-fold variation.The comparison also estimates how much residual uncertainty three-seed score fusion may leave.Source: H. Training-Seed Robustness Check · p. 5

A. Accuracy and Template Ageing

  • Trial counts fall from 376,387 at gap 0 to 51,729 at gap 7, and bootstrap intervals widen visibly at the larger gap.The full eight-point curve and per-fold breakdowns are in supplementary material.Source: A. Accuracy and Template Ageing · p. 5
  • EER increases monotonically from gap 0 to 7 for every mechanism; M4 has the lowest EER at every gap and M3 the highest.The increases range from 58% for M1 to 75% for M4, and the ordering holds across the five test folds.Source: A. Accuracy and Template Ageing · p. 5, Table III · p. 5, Figure 1 · p. 6
  • FNMR ranges from 69.6% to 98.3%; at 0.1% FMR, all four mechanisms reject most genuine claims.M3 remains worst, but M2 has the lowest FNMR while M4 has the lowest EER, reflecting that fixed-FMR FNMR and EER assess different parts of the DET curves.Source: A. Accuracy and Template Ageing · pp. 5–6, Table IV · p. 6

B. Ablations

Verification error rises with enrolment-to-query gap, while preprocessing and mechanism-specific ageing comparisons show distinct sensitivities. The evidence supports the general ageing effect more firmly than every comparative-rate claim.

B. Ablations

  • Clipping improves classical and shallow-ML results consistently, but M4 changes little and M3 worsens at ∆=0, from 21.4% unclipped to 27.2% clipped.Clipping uses a 1.5×IQR fence fitted on training participants only; the authors give no confident explanation for M3’s sensitivity.Source: B. Ablations · p. 6
  • M2 and M4 are close to saturated by k=3, whereas M1 improves from 33.3% to 27.4% and M3 changes only from 33.5% to 32.3% across k=1 to 5.Values are mean EER pooled across ∆ under the clipped condition.Source: B. Ablations · p. 6

C. Mixed-Effects Model

  • Each additional week increases the probability of decision error by 1.7 points, keeping other covariates constant.The reported effects are on the probability-of-decision-error scale.Source: C. Mixed-Effects Model · p. 6
  • Longer passwords are associated with modestly lower error, while the week-5 hardware-change covariate and substitution category are associated with higher error; the substitution result is considered unstable.The hardware change coincided with participants switching from university to personal equipment during a national COVID-19 lockdown.Source: C. Mixed-Effects Model · pp. 6–7
  • Across three fresh pipeline runs, M3 ages more slowly than M1 under both linear and logistic specifications, while M2 and M4 results vary by specification.The M3 result is the component the authors consider robust; M2 loses logistic significance, and M4 reverses direction under logistic regression.Source: C. Mixed-Effects Model · p. 7, Table VI · p. 8
  • Over seven weeks, M2 and M3 error probabilities rise 1.4–2.8 percentage points less than M1’s; M1 rises 12.1 points versus 9.8 for M2 and 9.9 for M3, while the accuracy ranking is preserved at every value of ∆.The M2 estimate should be read with caution given the logistic results.Source: C. Mixed-Effects Model · p. 8

D. Training-Seed Robustness

Repeated training runs reveal architecture-dependent seed sensitivity: M3 varies substantially more than M4, making single-run estimates less dependable for M3.

D. Training-Seed Robustness

  • 58% of M3’s fold-to-fold EER variance comes from training-run randomness, compared with 19% for M4.Each fold’s models were retrained three times on identical data with different seeds.Source: D. Training-Seed Robustness · p. 8
  • Across all five folds, M3’s independent training runs spread more than M4’s, sometimes as much as or more than the differences between folds.The per-fold comparison uses the three unfused training runs for each model.Source: D. Training-Seed Robustness · p. 8

V. DISCUSSION

The discussion treats template ageing and baseline accuracy as distinct properties, and narrows mechanism-specific ageing-rate conclusions to those robust across replication and outcome-model checks.

V. DISCUSSION

  • Across both linear and logistic mixed-model specifications, error increases monotonically and significantly with gap for every mechanism, and baseline accuracy differs significantly across mechanisms.M4 and M2 materially outperform M1 and M3 at baseline.Source: §V · p. 8
  • Password length and hardware change have small but statistically detectable effects, supporting their inclusion as covariates rather than leaving them implicit.Both effects are among the discussion’s robust findings.Source: §V · p. 8

A. Mechanism-Dependent Ageing Rates

  • Absolute accuracy and ageing robustness are separate properties, so selecting on same-session accuracy alone may miss the relevant trade-off for systems used without re-enrolment.The discussion notes M3 is least accurate while its slower ageing is robust, whereas M4 is most accurate and may age faster.Source: A. Mechanism-Dependent Ageing Rates · pp. 8–9
  • Only M3’s slower ageing than M1 is robust across all checks; M2’s apparent advantage and M4’s apparent parity depend on specification.M4’s logistic-model results suggest it may instead age faster than M1.Source: §V · p. 8, A. Mechanism-Dependent Ageing Rates · p. 9
  • The proposed floor-effect account is explicitly unverified: M3’s limited representational quality might leave less error headroom to grow with gap.The authors identify improved training as a test of this hypothesis, not a conclusion supported by the present data.Source: A. Mechanism-Dependent Ageing Rates · p. 9
  • The authors argue that mechanism comparisons should include seed-variance decomposition, independent replication, and outcome-model sensitivity checks.Two of the three original mechanism-specific sub-claims did not survive the checks performed.Source: A. Mechanism-Dependent Ageing Rates · p. 9

B. Limitations

Fixed-text verification error rises across the eight-week span, while only the recurrent model’s slower ageing rate survives the tested robustness checks. The conclusions are bounded by the participant pool, model comparisons, and outcome-model choices.

B. Limitations

  • The 85-participant pool is far smaller than the thousands TypeNet and TypeFormer are designed for, so M3’s poor accuracy may reflect limited population scale.The study also cannot separate the contributions of population size, training regime, and the difficult cross-week forced-impostor protocol to the gap from prior sub-10% EER results.Source: B. Limitations · p. 9
  • M2 uses separate per-password classifiers, which may benefit it relative to a single generic deployment model; a population-level M2 was not tested.Only one recurrent and one attention-based architecture were evaluated, leaving open whether M3’s ageing-rate advantage generalizes across architectures.Source: B. Limitations · p. 9
  • The logistic robustness check used cluster-robust fixed effects because a full binomial GLMM did not converge, leaving some M2 and M4 conclusions unresolved.A converged GLMM, lme4 fit, or larger dataset might resolve them differently.Source: B. Limitations · p. 9
  • The single-institution sample lacks demographic breakdowns, and shared passwords are not representative of self-selected deployment passwords.Thresholds were recalibrated at each gap, so the curves characterize separability rather than fixed-threshold lifetime error.Source: B. Limitations · p. 9

VI. CONCLUSION

  • Verification error grows significantly over eight weeks across classical, shallow-ML, recurrent, and Transformer mechanisms.The paper identifies this ageing effect as its most robust empirical claim.Source: §VI · pp. 9–10
  • Only M3’s slower ageing rate survives seed fusion, three pipeline replications, and an alternative outcome-model specification.The apparent M2 advantage and M4 parity did not survive the logistic specification.Source: §VI · p. 10

VII. APPENDIX

The appendix provides the password phrases and supplements the main text with trial-count, clipping, and enrolment-size breakdowns across temporal gaps.

B. Password Corpus

C. Full Ablation Results

  • The appendix gives full ablation breakdowns and a qualitative account of a negative-mining ablation resolved during development.Source: C. Full Ablation Results · p. 10

1) Trial Counts, All Eight Gaps:

2) Outlier Clipping, All Eight Gaps:

3) Enrolment Burden, All Eight Gaps:

4) Participant-Balanced EER Sensitivity Check:

The appendix checks participant-balanced EER, reports seed-level results, and documents why semi-hard-negative mining was used for the embedding models.

4) Participant-Balanced EER Sensitivity Check:

  • Pooled-trial and participant-balanced EER agree closely, with a maximum discrepancy of 1.7 percentage points, and produce the same mechanism ordering at every gap.The ordering is M4, M2, M1, then M3; pooled-trial weighting does not materially affect the paper’s comparative claims.Source: 4) Participant-Balanced EER… · pp. 10–11

5) Training-Seed Robustness, Full Per-Fold Breakdown:

6) Negative-Mining Strategy for the Triplet Loss (M3/M4 Training):

  • Hardest-negative mining collapsed both embedding models’ training loss to the margin within one epoch; switching to semi-hard mining resolved the collapse.The reported results use semi-hard mining; the failed configuration produced no model worth quantitative evaluation.Source: 6) Negative-Mining Strategy for the… · p. 11
Loading 2609.29851v1…