DESIGN OF A MOBILE PLATFORM FOR COLLECTING MULTIMODAL BIOMETRIC KEYSTROKE DYNAMICS DATA WITH REGARD TO THE INTERPRETABILITY REQUIREMENTS OF MACHINE LEARNING MODELS

Abstract

The subject of the study is the design of a mobile platform for collecting multimodal biometric data that combine touchscreen keystroke dynamics with inertial sensor signals of a smartphone, under the condition that the resulting corpus must be suitable not only for training recognition models but also for the subsequent interpretation of their decisions. The purpose of the work is to formulate the requirements for the data schema and the collection protocol that follow specifically from the interpretation task, to develop a platform architecture that satisfies these requirements, and to verify it by direct measurement on devices of different generations. The article solves the following tasks: eight requirements for the data schema and the collection protocol are formulated; the systematic distortions introduced by a naive collector implementation built on a standard text input field are identified and quantified; an architecture based on a custom input method that avoids these distortions is developed; a feature set is justified with regard to the actual availability of every feature on modern devices; the platform is validated on three smartphones. The following methods are used: instrumental measurement of the characteristics of the Android input and sensor subsystems; statistical processing of keystroke event time series using robust estimators; comparative analysis of public behavioral biometrics corpora against a unified set of criteria; analysis of residual information leakage in the de-identified collection mode. The following results were obtained. For the first time, the systematic divergence between the time bases of the Android input subsystem and the sensor subsystem has been quantified: across 12 sessions on three devices the offset ranged from 0.88 ms to 9.24 days, and without its explicit removal, in eleven sessions out of twelve not a single keystroke event falls within the boundaries of the session's own sensor recording, although the data appear complete. It is established that touch pressure and contact area are unavailable to a third-party on-screen keyboard on all three tested devices spanning three Android generations: across 1970 events the pressure value equals exactly one without a single exception, and the contact area is reported in units that are incomparable between devices, which casts doubt on the reproducibility of a part of the published mobile keystroke biometrics methods. It is shown that the requested rate of 200 Hz is sustained by all three devices, and the limiting factor is the application's request rather than the hardware. A collection mode with separable privacy for free-text tasks is proposed and verified, and a quantitative estimate of the residual leakage is provided. Conclusions: the interpretability requirement retrospectively imposes constraints on the data collection stage, while a naive collector implementation neither fails nor signals an error but silently produces files that pass any completeness check and are nevertheless unusable for multimodal analysis.

Keywords: keystroke dynamics; multimodal biometrics; inertial sensors; data collection; model interpretability; behavioral authentication; mobile devices.