Interaction-anchored behavioral biometrics for continuous authentication on smartphones

Cariello, Nicholas

Interaction-anchored behavioral biometrics for continuous authentication on smartphones

2026

Old Westbury, N.Y. : New York Institute of Technology, [2026]

Biometric identification
Authentication
Smartphones
Human-computer interaction

English

Smartphones have become central to personal and professional life, yet their security remains highly dependent on single-point verification mechanisms that cannot guarantee protection once a device is unlocked. Continuous authentication offers persistent verification, but current systems rely on long decision windows or fixed sampling that introduce unacceptable latency, creating vulnerable windows for compromise. These limitations represent a critical gap between the speed at which an adversary can take control of a device and the rate at which the system can detect misuse. This dissertation proposes an interaction-anchored framework for continuous authentication that re-centers the problem around natural user events such as keystrokes, swipes, and possession changes. By extracting and aggregating features within short event windows, it is possible to reliably authenticate a user in 1-2 seconds, sharply reducing vulnerability without sacrificing accuracy. Across five empirical studies, this work demonstrates that smartphone dynamics (swiping, typing, and motion features), including when combined with body motion features, yield competitive performance at sub-five-second windows, including the first evidence that keystroke-based continuous authentication can achieve error rates below two percent at one-second intervals. Further analysis within these studies establishes the value of event alignment, multimodal fusion, and carefully selected feature families. Finally, to assess the robustness of these features and methods, statistical forgery attacks under zero-shot, few-shot, and whitebox threat models are systematically evaluated across keystroke and swiping modalities; generative adversarial networks (GANs) are additionally explored to synthesize behavioral traces and probe cross-modal transfer. Together, these analyses provide one of the first systematic assessments of generative threats to short-window authentication and establish empirical bounds on robustness, highlighting avenues for defensive strategies such as feature hardening and multimodal liveness checks. Overall, this work reframes continuous authentication as a problem of event-driven, low-latency verification. The findings show that accurate and efficient authentication can be achieved with lightweight models suitable for on-device inference, while also charting a path toward systems resilient to emerging adversarial threats. This positions interaction-anchored biometrics as both a practical and forward-looking foundation for smartphone security.

PDF

Thesis

School of Engineering and Computing Sciences

Department of Computer Science

Doctor of Philosophy (Ph.D.) in Computer Science