Abstract
<jats:p>This paper addresses the problem of estimating the reliability (sample quality) of the current keystroke-dynamics sample in behavioral user authentication at login time. Unlike risk-based fusion or context–behavior score combination, the focus is: “Can this particular behavioral sample be trusted enough to use the behavioral channel?” We provide a formal definition of sample reliability as a probabilistic estimate of sample usability/utility for biometric comparison, and propose a feature model capturing degradation factors: event completeness, effective sequence length, timing variability, autofill/paste indicators, device-change signals, and timestamp quantization/jitter. An integral reliability estimation method based on a logistic quality model is developed, and the relationship between estimated reliability and the behavioral verifier’s error is analyzed. We also discuss probability calibration (Platt scaling, isotonic regression, Bayesian binning) to convert raw scores into well-interpretable probabilities. Experimental validation is performed on the public DSL-StrongPassword benchmark dataset (51 users, 400 password typings per user) with controlled synthetic degradations (event loss, truncation, jitter/quantization). Results show that reliability-based filtering improves behavioral matching performance (AUC increases from 0.856 to 0.890 for samples with q≥0.8 at ≈53% coverage) and changes the error profile, reducing false rejects for legitimate users in high-quality samples. Practical deployment recommendations for reliability thresholds and “use/do-not-use” gating policies are provided.</jats:p>