Scholarlink

<p>Behavioral biometric signals such as keystroke dynamics and gait are inherently non-stationary due to <i>human adaptation</i>, defined here as intra-user temporal drift across sessions arising from fatigue, cognitive load, device handling, and contextual variation. Because impostor data are typically unavailable at enrollment and genuine behavior evolves over time, biometric authentication is naturally formulated as a <i>one-class anomaly detection</i> problem rather than a fixed-distribution classification task. This work addresses the resulting generalization and stability challenges by representing behavioral consistency using <i>fidelity-based similarity</i>, inspired by quantum state overlap, where gradual fidelity decay provides a bounded and calibrated indicator of anomalous deviation. In the proposed quantum kernel-based anomaly detection (QKAD) framework, biometric feature sequences are mapped to quantum states via hybrid angle–amplitude encoding, behavioral similarity is evaluated using quantum kernel fidelity, and anomaly scores are produced by a classical one-class support vector machine (OC-SVM) operating on the induced kernel geometry. The framework is primarily evaluated using a noise-aware quantum simulator (<Emphasis FontCategory="NonProportional">FakeBrisbaneV2</Emphasis>), with selective real-hardware executions on IBM Marrakesh used exclusively for trend validation. Across keystroke and gait modalities, evaluated on a session-separated multimodal dataset of 160 participants, QKAD achieves authentication performance comparable to strong classical baselines, with EERs in the range of 4.4–7.2% and AUC values above 0.94, while exhibiting stable and predictable behavior under noise and session drift. These results indicate that fidelity-based quantum kernels provide a <i>well-calibrated and physically interpretable similarity representation</i> for anomaly-aware behavioral biometric authentication, without claiming a decisive accuracy advantage over existing methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quantum kernel anomaly detection: a fidelity-based framework for robust behavioral biometric authentication

  • 期刊论文

  • 发表日期：

    2026年2月15日

查看全文
  • Sandip Dutta,
  • Soumen Roy,
  • Utpal Roy

摘要

Behavioral biometric signals such as keystroke dynamics and gait are inherently non-stationary due to human adaptation, defined here as intra-user temporal drift across sessions arising from fatigue, cognitive load, device handling, and contextual variation. Because impostor data are typically unavailable at enrollment and genuine behavior evolves over time, biometric authentication is naturally formulated as a one-class anomaly detection problem rather than a fixed-distribution classification task. This work addresses the resulting generalization and stability challenges by representing behavioral consistency using fidelity-based similarity, inspired by quantum state overlap, where gradual fidelity decay provides a bounded and calibrated indicator of anomalous deviation. In the proposed quantum kernel-based anomaly detection (QKAD) framework, biometric feature sequences are mapped to quantum states via hybrid angle–amplitude encoding, behavioral similarity is evaluated using quantum kernel fidelity, and anomaly scores are produced by a classical one-class support vector machine (OC-SVM) operating on the induced kernel geometry. The framework is primarily evaluated using a noise-aware quantum simulator (FakeBrisbaneV2), with selective real-hardware executions on IBM Marrakesh used exclusively for trend validation. Across keystroke and gait modalities, evaluated on a session-separated multimodal dataset of 160 participants, QKAD achieves authentication performance comparable to strong classical baselines, with EERs in the range of 4.4–7.2% and AUC values above 0.94, while exhibiting stable and predictable behavior under noise and session drift. These results indicate that fidelity-based quantum kernels provide a well-calibrated and physically interpretable similarity representation for anomaly-aware behavioral biometric authentication, without claiming a decisive accuracy advantage over existing methods.