EchoType: Identity recognition from typing-induced ultrasonic reflections

Abstract

Securely recognizing a user's identity without disrupting everyday interactions remains a central challenge in modern computing environments. Prior keystroke-dynamics-based approaches primarily rely on timing patterns extracted from discrete key events and often achieve high accuracy through increasingly complex deep learning models. However, because these methods operate on abstracted event-level features, prior studies have shown that they can be vulnerable to spoofing attacks in which adversaries intentionally mimic keystroke timing patterns. In this paper, we introduce ECHOTYPE, an identity recognition approach that departs from event-driven keystroke modeling and instead leverages ultrasonic reflections generated during ordinary typing. As a user types, continuous finger motion and anatomical structure modulate inaudible ultrasonic signals emitted by a commodity smartphone, producing spatiotemporal reflection patterns that are physically coupled to the user's motion. ECHOTYPE operates in a physical signal space shaped by both behavior and anatomy, rather than relying solely on key-event timing. ECHOTYPE requires no dedicated sensors or modifications to input devices and integrates seamlessly into existing computing environments using only the built-in speaker and microphone of a smartphone. To accommodate diverse usage scenarios, we evaluate ECHOTYPE under both single-user verification and multi-user identification settings. Experimental results across multiple devices, keyboard types, and everyday environments demonstrate robust performance: ECHOTYPE achieves an Equal Error Rate (EER) of 0.031 for single-user verification, a macro F1-score of 0.940 for multi-user identification, and 95.5% accuracy in unknown-user rejection. These results indicate that typing-induced ultrasonic reflections can serve as a practical and deployable identity recognition mechanism without compromising usability in modern computing environments. © 2026 Elsevier Ltd