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
cs.CR · cs.AI
Quick verdict
Keystroke template ageing is real (~1.7pp error/week over 8 weeks, 4 matchers), but which matcher ages slowest is partly specification-dependent.
Summary
The paper measures how fixed-text keystroke verification degrades as the enrolment-to-query gap grows, using an eight-week study of 85 participants typing 40 assigned passwords four times per weekly session (94,894 clean samples from 99,607 raw). Under 5-fold subject-disjoint cross-validation, four matchers — scaled Manhattan (M1), per-password gradient-boosted classifiers (M2), a TypeNet-style LSTM (M3), and a TypeFormer-style Transformer (M4) — are scored on a shared trial grid spanning gaps of 0–7 weeks. EER rises monotonically for every mechanism (M4: 14.6%→25.5%; M1: 21.0%→33.1%), roughly 1.7 percentage points of decision error per week in a mixed-effects model (p<0.001), with Δ=7 exceeding Δ=0 in all 20 fold-by-mechanism comparisons. Baseline accuracy differences (12.6-point spread) dwarf accumulated degradation (2.3 points), and ageing never reorders the mechanisms. The only robust mechanism-specific effect is that M3, the least accurate model, ages more slowly than M1 across linear and logistic specifications and three pipeline replicates; M2's advantage and M4's parity do not survive the logistic check. Seed-noise decomposition attributes 58% of M3's fold variance to training randomness versus 19% for M4. Main limitations: absolute errors far above benchmark levels (FNMR at 0.1% FMR reaches 93.9–99.3%), no hyperparameter search, a per-password M2 asymmetry, an 85-participant single-institution pool, and per-gap threshold recalibration rather than fixed-threshold deployment modelling.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fixed-text keystroke verification error rises monotonically and significantly with the enrolment-to-query gap: EER moves from 14.6–27.2% at Δ=0 to 25.5–37.1% at Δ=7, about 1.7pp of decision error per week (p<0.001).
- Baseline accuracy differences (12.6pp) far exceed accumulated degradation (2.3pp over seven weeks), and ageing never reorders the four mechanisms, so matchers can be chosen on same-session accuracy with ageing managed by re-enrolment scheduling.
- The least accurate mechanism (M3, TypeNet-style LSTM) ages significantly more slowly than the Manhattan baseline under every specification tested — the only mechanism-specific ageing claim the authors consider robust.
- Claims that M2 ages more slowly and that M4 ages at parity are artifacts of the linear-probability model; the logistic specification removes M2's advantage and tentatively suggests M4 may age faster.
- Training-seed randomness accounts for 58% of M3's fold-to-fold EER variance versus 19% for M4, so comparative ageing-rate claims require seed fusion, independent replication, and outcome-model sensitivity checks.
Strengths
Single unified trial design
All four mechanisms score the identical (gallery, query) grid across Δ=0–7 (Section III-D), with trial counts identical per cell across mechanisms (Table VIII), so accuracy, ageing-rate and interaction estimates are mutually comparable rather than stitched from separate experiments.
Rare statistical rigor for the domain
Crossed participant/password random effects with fold as fixed effect (Eq. 1, N≈300k trials), participant-level bootstrap CIs following Bolle et al., ISO/IEC 19795-1 metrics, and a pooled-vs-participant-balanced estimator sensitivity check showing a maximum 1.7pp shift with unchanged ordering (Table XI).
Exemplary self-audit of its own claims
Highly robust core ageing effect
Useful documented negative results
Weaknesses
Absolute accuracy far below cited benchmarks
Mechanism comparison confounded by training adequacy
Thin surviving comparative-ageing evidence
Comparability asymmetries between mechanisms
Internal numeric inconsistency and scope limits
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a deceptively simple question: does fixed-text keystroke verification degrade as the enrolment-to-query gap grows, and does the degradation rate depend on the matching mechanism? Using an eight-week dataset (85 participants, 40 assigned passwords, four repetitions per weekly session, 94,894 clean samples from 99,607 raw; Section III-A, based on [12]), the authors compare four matchers — scaled Manhattan (M1), per-password gradient boosting (M2), a TypeNet-style BiLSTM (M3), and a TypeFormer-style Transformer (M4) — on a single subject-disjoint trial grid spanning Δ=0–7 weeks (Section III-D). Three headline claims emerge: (1) ageing is large, monotonic, and systematic (~1.7pp decision error per week, p<0.001); (2) baseline accuracy differences (12.6pp spread) dwarf accumulated degradation (2.3pp spread), so ageing never reorders mechanisms; (3) accuracy and ageing rate are dissociable — M3, the least accurate model, is the only mechanism whose slower ageing survives every statistical check. Claim 3 is presented with unusual candor as a narrowing from three original sub-claims to one (Sections V-A, VI).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is template ageing in keystroke dynamics?
Template ageing is the decline of biometric verification accuracy as the time between enrolment and verification grows. This paper measures it directly for typed passwords: over eight weeks, decision error rose by roughly 1.7 percentage points per week for all four matching algorithms tested, with equal error rates climbing from 14.6–27.2% at the same week to 25.5–37.1% at a seven-week gap.
Which keystroke verification model performed best in this study?
The TypeFormer-style Transformer (M4) had the lowest equal error rate at every time gap (14.6% at gap zero), followed by gradient-boosted classifiers (M2, 17.2%); the TypeNet-style LSTM (M3) was worst (27.2%). Differences in starting accuracy (12.6 points) far exceeded differences in how fast the models degraded (2.3 points over seven weeks).
Does any keystroke model degrade more slowly over time?
Only the least accurate model (the TypeNet-style LSTM, M3) showed a robustly slower degradation rate than the Manhattan-distance baseline, and this held under every statistical check performed. The corresponding claims for the other two models depended on the choice of statistical model and did not survive a logistic re-analysis, which even hinted the Transformer may age faster.
Is the keystroke dataset and code from this paper available?
The paper states that all collected data and code are available at github.com/sparkins01/keystroke (Appendix VII-A). The underlying dataset was previously described in a 2023 IET Biometrics paper by the same group: 85 participants typing 40 fixed passwords in weekly sessions across eight weeks.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–2 (ageing curves, DET curves) and the supplementary per-fold material are referenced but not analyzable from the extracted text.
- Internal inconsistency in the source: Section IV-A prose cites FNMR 69.6–98.3% while Table IV shows 79.7–99.3%; table values are used in this review.
- Code/data availability is asserted (Appendix VII-A, github.com/sparkins01/keystroke) but the repository contents were not verified.
- The abstract is truncated mid-sentence in the extraction ('with 58%'), but the figure is restated fully in the body (Table VII).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.