A privacy-preserving framework for continuous mobile authentication using digital twins and multimodal biometrics
Abhishek Singh Dhadwal, Aditya Ranjit KotwalBackground
One-time mobile authentication remains vulnerable to post-login compromise. Biometric templates, once exposed, cannot be revoked. Continuous authentication that models user behavior and physiology over time can improve post-login security, but it must preserve privacy, resist spoofing and replay, and demonstrate measurable performance under realistic data conditions.
Methods
We propose and empirically evaluate a layered Digital Twin framework for continuous mobile authentication using passive smartphone and wearable signals. The pipeline includes sensor ingestion, feature extraction, adaptive Digital Twin modeling, context-aware score fusion, authentication decision logic, and privacy-preserving protection through encrypted templates, hardware-backed storage assumptions, and proxy measurements for zero-knowledge proof and homomorphic-encryption-inspired verification steps. We evaluated the prototype using the Wearable Stress and Affect Detection (WESAD) dataset, Wearable Exam Stress (PhysioNet), the cannabis-use, stress, and physiological-response dataset (CAN-STRESS/CAN-STEESS), and Carnegie Mellon University (CMU) Keystroke data.
Results
The strongest physiological authentication result was obtained on WESAD, with an equal error rate (EER) of 0.2216, area under the receiver operating characteristic curve (AUROC) of 0.8618, and area under the precision-recall curve (AUPRC) of 0.8497. This was followed by CAN-STRESS (EER 0.3203, AUROC 0.7557) and PhysioNet (EER 0.4290, AUROC 0.6003). The secondary CMU keystroke track produced EER 0.5392 and AUROC 0.4152. On WESAD, multimodal fusion improved AUROC by 0.5018 over the best unimodal result, and adaptive updating improved performance on three of four evaluated datasets. Mean feature-extraction latency was 3.281 milliseconds (ms) per window, with low-millisecond proxy timings for encryption and proof/verification steps.
Discussion
The results support the feasibility of an adaptive, privacy-aware Digital Twin architecture, while also showing that performance varies substantially by modality and dataset. The framework remains an offline prototype rather than a deployed mobile system. Remaining limitations include the absence of a live user study, direct battery/energy measurements, confidence intervals, fixed- vs -context-conditioned fusion ablations, replay/mimicry trials, and poisoning-specific adversarial tests.