XKey: A Unified Deep Learning Framework for Wearable-based Key Generation

Secure key generation from shared biometric signals has emerged as a lightweight alternative to traditional cryptographic key distribution, particularly for resource-constrained wearable devices. However, existing methods are often modality-specific and rely heavily on handcrafted features, which limit their generalizability and efficiency across diverse wearable-based applications. In this paper, we propose XKey, a unified framework that formulates feature extraction from biometric signals as an optimization problem that jointly maximizes inter-device similarity and key entropy. XKey employs deep sequence models to automatically learn robust, information-rich representations, facilitating reliable key generation tailored to biometric signals. To further improve agreement rates, we incorporate a flexible reconciliation mechanism based on compressed sensing. We demonstrate XKey's generalizability by implementing it for distributed key generation using cardiac-based and gait-based signals. Extensive evaluations on cardiac and gait datasets show that XKey achieves high key agreement rates up to 100.0% and faster key generation rate compared to state-of-the-art works (with improvements of 12.5% for cardiac signals and 27.8% for gait signals). These results underscore XKey's promise as a practical and scalable solution for secure communication between current commercial wearable devices. We release the XKey code and instructions at https://github.com/yishuozhao/XKey.