BIOWISH: Biometric Recognition Using Wearable Inertial Sensors Detecting Heart Activity

Abstract

Wearable devices have been recently proposed to perform biometric recognition, leveraging on the uniqueness of the collectable physiological traits to generate discriminative identifiers. Most of the studies conducted on this topic have exploited heart-related signals, sensing the cardiac activity either through electrical measurements using electrocardiography, or with optical recordings employing photoplethysmography. In this paper we instead propose a system performing BIOmetric recognition using Wearable Inertial Sensors detecting Heart activity (BIOWISH). In more detail, we investigate the feasibility of exploiting mechanical measurements obtained through seismocardiography and gyrocardiography to verify the identity of a subject. Several feature extractors and classifiers, including deep learning techniques relying on siamese training, are employed to derive distinctive characteristics from the considered signals, so as to differentiate between legitimate users and impostors. A multi-session database, comprising acquisitions taken from subjects performing different activities, is employed to perform experimental tests. The obtained results testify that identifiers derived from measurements of chest vibrations, collected by wearable inertial sensors, could be employed to guarantee high recognition performance, even when considering short-time recordings. Explainability methods have been also employed to derive some insights about the aspects relevant to perform predictions for both people and activity recognition tasks.

Person Identification and Relapse Detection From Continuous Recordings of Biosignals Challenge: Overview and Results

“…The results of the challenge indicate an advancement of the state-of-the-art in both the person identification task and the psychotic relapse detection task. With regards to the person identification task, novel solutions have been developed for the handling of missing features, such as employing a separate encoding for replacing them [39], whereas ensemble models, trained in separate modalities and/or with diverse training settings, yielded better results than the commonly employed end-to-end models [35], [36]. The main advancement, however, concerns the swing at the direction of raw sensorial data instead of aggregated feature representations; the top-performing teams either utilized directly the provided data at the 5-sec resolution [37], [38], or downsampled them, with the final sampling rate remaining less than 1 minute [39].…”

“…The main advancement, however, concerns the swing at the direction of raw sensorial data instead of aggregated feature representations; the top-performing teams either utilized directly the provided data at the 5-sec resolution [37], [38], or downsampled them, with the final sampling rate remaining less than 1 minute [39]. This contrasts with the state-ofthe-art in the task, where mostly spectral features [36] have been employed, and a larger (minute-scale) temporal slice has been used for feature aggregation [27], [31] (with a few exceptions [46]), and implies the potential of achieving further gains with the utilization of raw recordings. On the other hand, for the relapse detection task, the main advancement concerns successful adaptation of Transformerbased architectures [38], in contrast to DNN or CNN ones that reached the best performance in previous works [21], [28].…”

“…More recent approaches utilize deep learning, an approach that effectively generalizes in userdiverse datasets, while being able to disentangle the recorded signals and sensor noise with suitable data augmentation. In this case, CNN-based architectures are mostly being used, in either a discriminative fashion [35] or a representation learning framework [36].…”