Mechanistic explanation for M3’s sensitivity to outlier clipping

Establish why the TypeNet-style bidirectional LSTM embedding model’s verification performance is substantially more sensitive to outlier clipping than the other evaluated mechanisms, particularly why clipping worsens its same-session EER despite leaving its raw sequence representation unchanged.

Background

The study evaluates both clipped and unclipped timing data. At zero enrolment-to-query gap, clipping improves or has little effect on most mechanisms, but the TypeNet-style LSTM model’s EER worsens from 21.4% without clipping to 27.2% with clipping.

The authors connect this result to M3’s broader instability but explicitly state that they do not have a confident explanation. Thus, the mechanism by which preprocessing affects the recurrent embedding model remains unresolved.

References

We do not have a confident explanation for M3's sensitivity to clipping, beyond noting that it is consistent with the broader finding, discussed in Section~\ref{sec:seedresults}, that M3 is the least stable of the four mechanisms.

— Template Ageing and Longitudinal Verification in Fixed-Text Keystroke Dynamics: A Subject-Disjoint Study Across Eight Weeks  (2609.29851 - Parkinson et al., 24 Sep 2026) in Section 4.2, “Ablations”