In always-on HAR deployments, model accuracy erodes silently as domain shift accumulates over time. Addressing this challenge requires moving beyond one-off updates toward instance-driven adaptation from streaming data. However, continuous adaptation exposes a fundamental tension: systems must selectively learn from informative instances while actively forgetting obsolete ones under long-term, non-stationary drift. To address them, we propose CODA, a continuous online adaptation framework for mobile sensing. CODA introduces two synergistic components: (i) Cache-based Selective Assimilation, which prioritizes informative instances likely to enhance system performance under sparse supervision, and (ii) an Adaptive Temporal Retention Strategy, which enables the system to gradually forget obsolete instances as sensing conditions evolve. By treating adaptation as a principled cache evolution rather than parameter-heavy retraining, CODA maintains high accuracy without model reconfiguration. We conduct extensive evaluations on four heterogeneous datasets spanning phone, watch, and multi-sensor configurations. Results demonstrate that CODA consistently outperforms one-off adaptation under non-stationary drift, remains robust against imperfect feedback, and incurs negligible on-device latency.

Continuous ECG monitoring is vital for early detection of arrhythmias and other cardiac abnormalities—especially during sleep, when symptoms often go unnoticed—yet existing solutions remain expensive, obtrusive, and impractical for long-term daily use. Ballistocardiography (BCG)—a passive, contactless modality that captures cardiac-induced body motion—offers a compelling alternative. However, prior efforts treat ECG reconstruction as waveform regression, leading to overfitting to individual-specific or posture-dependent artifacts. Inspired by the fact that ECG and BCG reflect parallel structures in cardiac event sequences (e.g., P/QRS/T-waves vs. I/J-waves), we design Med2ECG: a structure-aligned system that reconstructs ECG from high-fidelity BCG by explicitly aligning their latent physiological events. Med2ECG incorporates three key designs: (i) Multi-Scale Feature Extractor captures hierarchical temporal dynamics, preserving clinically relevant fine-grained features; (ii) Shared-Personalized Experts employs a Mixture-of-Experts (MoE) to adaptively disentangle signal variations due to individual and environmental factors; (iii) Medical-Informed Strategies introduces a diagnostic-driven multi-objective loss, integrating structural alignment, morphological fidelity, and landmark-aware supervision to preserve clinically critical intervals. Experiments across public and self-collected in-hospital datasets (20 healthy individuals and 10 patients with diverse cardiovascular conditions), Med2ECG achieves > 0.92 Pearson correlation, < 15% amplitude error, and precise PR/QRS/QT/RR interval estimation within 5–20 ms—demonstrating strong generalization across subjects, postures, and environments.

Nasogastric tube (NGT) feeding supports over 10 million patients globally, particularly those with stroke-induced dysphagia, Parkinson's disease, and cognitive impairments. However, current practices face three fundamental limitations. First, caregivers often lack knowledge of the nutritional content of homemade blended meals, leading to high malnutrition rates among long-term NGT patients. Second, commercial solutions struggle to analyze blended meals in large containers due to light path distortion. Third, gastric residual volume (GRV) assessments rely on subjective nurse evaluations, resulting in incomplete records and increased feeding intolerance. To address these challenges, this demo introduces NutriBump, an innovative system designed to enhance NGT feeding through closed-loop control. Utilizing advanced spectral analysis technology, the system accurately assesses food nutrients and employs multi-modal sensors for automatic gastric fluid aspiration and analysis. By integrating long-term nutritional intake, digestion data, and health status, NutriBump leverages large language models and AI agents to generate personalized nutrition reports automatically. This closed-loop nutrition management solution improves the accuracy of blended meal analysis and reduces reliance on subjective digestion assessments.

Hypertension affects 1.28 billion adults, but blood pressure alone is insufficient for accurate hypertension detection. In this paper, we identified a range of multi-dimensional hemodynamics as novel biomarkers that can offer more comprehensive cardiovascular insights for detecting hypertension. We design a novel system that collects ballistocardiogram (BCG) signals from an optic fiber sensor mat. It features a multi-task, multi-branch, unsupervised domain adaptation learning framework, enabling simultaneous prediction of five hemodynamic biomarkers for hypertension detection. In a 4-month trial with 85 subjects, Hyde achieved 97.65% average accuracy for hypertension detection.

Human activity recognition (HAR) has become a research hotspot because of its wide range of application prospects. It has higher requirements for real-time and powerefficient processing. However, a large amount of data transfer between sensors and servers, and computation-intensive recognition models hinder the implementation of real-time HAR systems. Recently, edge computing has been proposed to address this challenge by moving computational and data storage resources to the sensors, rather than depending on a centralized server/cloud. In this paper, we investigated binary neural networks for edge intelligence-enabled HAR using radar point cloud. Point cloud can provide 3-dimensional spatial information, which is helpful to improve recognition accuracy. Time-series point cloud also brings challenges, such as larger data volume, 4-dimensional data processing, and more intensive computation. To tackle these challenges, we adopt the 2-dimensional histograms for point cloud multi-view processing and propose the EdgeActNet, a binary neural network for point cloud-based human activity classification on edge devices. In the evaluation, the EdgeActNet achieved the best results with average accuracies of 97.63% on the MMActivity dataset and 95.03% on the point cloud samples of the DGUHA dataset respectively; and saved 16.9× memory consumption and 11.5× inference time compared to its full-precision version. Our work also is the first to apply 2D histogram-based multi-view representation and BNNs for timeseries point cloud classification.

The use of speakers in electronic devices has become widespread, but the security risks associated with micro-speakers, such as earphones, are often overlooked. Many assume that soundproof barriers can prevent sound leakage and protect privacy. This article presents the prototype MagEar, an eavesdropping system that exploits magnetic side-channel signals leaked by a micro-speaker to restore intelligible human speech. MagEar outperforms some high-precision magnetometers in detecting magnetic fields at the nanotesla level. Even at a distance of 60 cm, it can recover high-quality audio with a 90% similarity to the original audio. Moreover, the MagEar prototype is portable and can be concealed within a headset housing. We have implemented MagEar as a proof-of-concept system and conducted multiple case studies on the eavesdropping of various speaker-embedded devices, including earphones. The recovered speech can be transcribed using automatic speech recognition techniques, even when obstructed by soundproof walls. It is our aspiration that our work can prompt manufacturers to reconsider the security vulnerabilities of speakers.

Multi-user indoor localization is considered to be one of the most useful wireless applications. Low latency and high robustness to dynamic interference from surrounding people are essential requirements for multi-user localization. However, state-of-the-art (SOTA) indoor localization systems cannot satisfy both requirements at the same time. In this paper, we propose RIScan, a Reconfigurable Intelligent Surface (RIS)-aided localization system that can achieve both low latency and high reliability. We leverage RIS to perform Wi-Fi beam scanning so all clients can figure out their direction in a single scan. However, compared with traditional AP-based systems, the introduction of RIS creates a more complicated signal superposition at the receiver, preventing clients from directly obtaining target beams for direction derivation and localization. To overcome this challenge, we fully utilize the reconfigurability of RIS to endow target beams with distinguishing features, so that RIScan can extract stable and accurate direction information from complex and dynamic environments. RIScan is implemented in the real system with our own developed 16 × 16 RIS prototype and COTS Wi-Fi devices. Extensive experiments show that RIScan achieves a median localization error of 47cm and 71cm in static and dynamic environments with only two RIS anchors. Compared to the SOTA methods, RIScan reduces the localization latency by more than an order of magnitude.

Transforming physical surfaces into virtual interfaces can extend the interaction capability of many exciting metaverse applications in the future. Recent advances in vibration-based tap sensing show promise for this vision using passive vibration signals. However, current approaches based on Time-Difference-of-Arrival (TDoA) triangulation suffer the impact of fluctuant wave velocity due to the dispersive and heterogeneous nature of solid mediums, failing to meet the performance requirement for practical use. In this article, we present MM-Tap, a vibration-based tap localization system that can transform ubiquitous surfaces into virtual touch screens with low overhead. A novel localization scheme is proposed based on the finding of spatiotemporal mapping between tap locations and TDoA values, which pushes the accuracy limits of vibration-based tap sensing from unstable cm-level to mm-level. We investigate the geometry of the sensor layout and design a model-based method to synthesize tap data, which enables MM-Tap to adapt to various surface materials and respond to arbitrary sensing scales after a few seconds of calibration. We combine MM-Tap with a COTS projector and facilitate a digitally augmented surface where users can play video games with low latency.

Wearable devices equipped with a variety of sensors facilitate the measurement of physiological and behavioral characteristics. Activity-based person identification is considered an emerging and fast-evolving technology in security and access control fields. Wearables, such as smartphones, Apple Watch, and Google glass can continuously sense and collect activity-related information of users, and activity patterns can be extracted for differentiating different people. Although various human activities have been widely studied, few of them (gaits and keystrokes) have been used for person identification. In this article, we performed person identification using two public benchmark data sets (UCI-HAR and WISDM2019), which are collected from several different activities using multimodal sensors (accelerometer and gyroscope) embedded in wearable devices (smartphone and smartwatch). We implemented eight classifiers, including an multivariate squeeze-and-excitation network (MSENet), time-series transformer (TST), temporal convolutional network (TCN), CNN-LSTM, ConvLSTM, XGBoost, decision tree, and$k$-nearest neighbor. The proposed MSENet can model the relationship between different sensor data. It achieved the best person identification accuracies under different activities of 91.31% and 97.79%, respectively, for the public data sets of UCI-HAR and WISDM2019. We also investigated the effects of sensor modality, human activity, feature fusion, and window size for sensor signal segmentation. Compared to the related work, our approach has achieved the state of the art.

Traditional liquid identification instruments are often unavailable to the general public. This paper shows the feasibility of identifying unknown liquids with commercial lightweight devices, such as a smartphone. The wisdom arises from the fact that different liquid molecules have various viscosity coefficients, so they need to overcome dissimilitude energy barriers during relative motion. With this intuition in mind, we introduce a novel model that measures liquids’ viscosity based on active vibration. The idea sounds straightforward, yet, it is challenging to build up a robust system utilizing the built-in accelerometer in smartphones. Practical issues include under-sampling, self-interference, and volume change impact. Instead of using machine learning techniques, we tackle these issues through multiple signal processing stages to reconstruct the original signals and cancel out the interference. Our approach achieved the liquid viscosity estimates with a mean relative error of 2.3% and distinguish 30 kinds of liquid with an average accuracy of 97.33%.

A wide diversity of sensors has been applied in human activity recognition. These sensors generate enormous amounts of data during human activity monitoring. The long-distance data traveling between sensors and servers increases the costs of bandwidth and latency. However, human activity recognition has a high demand for real-time processing. Recently, edge computing is surging to solve this problem by moving computation and data storage closer to the sensor devices, rather than relying on a central server/cloud. Edge servers are usually designed for low power, low cost, and low computation. They do not support computation-intensive deep learning algorithms or will result in high latency. Fortunately, the development of binarized neural networks enables edge intelligence which supports AI running at the network edge for real-time applications. In this paper, we implement a binarized neural network (BinaryDilatedDenseNet) to enable low-latency and low-memory human activity recognition at the network edge. We applied the BinaryDilatedDenseNet on three sensor-based human activity recognition datasets and evaluated it with four metrics. In comparison, the BinaryDilatedDenseNet outperforms the related work and other three binarized neural networks in accuracy and saves 10 memory and 4.5--8 inference time compared to the FPDilatedDenseNet(the full-precision version of the BinaryDilatedDenseNet).

The inevitable aging trend of the world’s population brings a lot of challenges to the health care for the elderly. For example, it is difficult to guarantee timely rescue for single-resided elders who fall at home. Under this circumstance, a reliable automatic fall detection machine is in great need for emergent rescue. However, the state-of-the-art fall detection systems are suffering from serious privacy concerns, having a high false alarm, or being cumbersome for users. In this article, we propose a device-free fall detection system, namely G-Fall, based on floor vibration collected by geophone sensors. We first decompose the falling mode and characterize it with time-dependent floor vibration features. By leveraging Hidden Markov Model (HMM), our system is able to detect the fall event precisely and achieve user-independent detection. It requires no training from the elderly but only an HMM template learned in advance through a small number of training samples. To reduce the false alarm rate, we propose a novel reconfirmation mechanism using Energy-of-Arrival (EoA) positioning to assist in detecting the human fall. Extensive experiments have been conducted on 24 human subjects. On average, G-Fall achieves a 95.74% detection precision on the anti-static floor and 97.36% on the concrete floor. Furthermore, with the assistance of EoA, the false alarm rate is reduced to nearly 0%.

Speakers have been widely embedded in various electronic devices as a standard configuration. The security vulnerability of microspeakers (such as earphones) is commonly overlooked because it is often assumed that soundproof boundaries, such as walls, can prevent privacy-infringing sound leakage. In this paper, we present the prototype MagEar, an eavesdropping system that leverages magnetic side-channel signals leaked by a microspeaker to recover intelligible human speech. MagEar has sufficiently high sensitivity to detect magnetic fields on the order of nanotesla, exceeding some high-precision magnetometers. It can recover high-quality audio with 90% similarity to the original audio even at a distance of 60 cm. In addition, the MagEar prototype is portable and can be hidden in a headset shell. We have implemented MagEar as a proof-of-concept system and conducted several case studies of eavesdropping on different types of speaker-embedded devices, including earphones, and we have demonstrated the ability to successfully transcribe the recovered speech using automatic speech recognition techniques even when blocked by soundproof walls. We hope that our work can push manufacturers to rethink this security vulnerability of speakers.

Traditional liquid identification instruments are often unavailable to the general public. This paper shows the feasibility of identifying unknown liquids with commercial lightweight devices, such as a smartphone. The wisdom arises from the fact that different liquid molecules have various viscosity coefficients, so they need to overcome dissimilitude energy barriers during relative motion. With this intuition in mind, we introduce a novel model that measures liquids' viscosity based on active vibration. Yet, it is challenging to build up a robust system utilizing the built-in accelerometer in smartphones. Practical issues include under-sampling, self-interference, and volume change impact. Instead of machine learning, we tackle these issues through multiple signal processing stages to reconstruct the original signals and cancel out the interference. Our approach could achieve the liquid viscosity estimates with a mean relative error of 2.9% and distinguish 30 kinds of liquid with an average accuracy of 95.47%.

Smart wristband has become a dominant device in the wearable ecosystem, providing versatile functions such as fitness tracking, mobile payment, and transport ticketing. However, the small form-factor, low-profile hardware interfaces and computational resources limit their capabilities in security checking. Many wristband devices have recently witnessed alarming vulnerabilities, e.g., personal data leakage and payment fraud, due to the lack of authentication and access control. To fill this gap, we propose a secure text pin input system, namely Taprint, which extends a virtual number pad on the back of a user's hand. Taprint builds on the key observation that the hand “landmarks”, especially finger knuckles, bear unique vibration characteristics when being tapped by the user herself. It thus uses the tapping vibrometry as biometrics to authenticate the user, while distinguishing the tapping locations. Taprint reuses the inertial measurement unit in the wristband, “overclocks” its sampling rate with the cubic spline interpolation to extrapolate fine-grained features, and further refines the features to enhance the uniqueness and reliability. Extensive experiments on 128 users demonstrate that Taprint achieves a high accuracy (96 percent) of keystrokes recognition. It can authenticate users, even through a single-tap, at extremely low error rate (2.2 percent), and under various practical usage disturbances.

Nowadays, smart wristbands have become one of the most prevailing wearable devices, as they are small and portable. However, due to the limited size of the touch screens, smart wristbands typically have poor interactive experience. There are a few works appropriating the human body as a surface to type on. Yet, by using multiple sensors at high sampling rates, they are not portable and are energy-consuming in practice. To break this stalemate, we proposed a portable, cost efficient text-entry system, termed ViType, which first leverages a single small form factor sensor to achieve a practical user input with much lower sampling rates. To enhance the input accuracy with less vibration information introduced by lower sampling rates, ViType designs a set of novel mechanisms, including a fine-grained feature extraction to process the vibration signals, and a runtime calibration and adaptation scheme to recover from the error due to temporal instability. Extensive experiments have been conducted on 30 human subjects. The results demonstrate that ViType is robust against various confounding factors. The average recognition accuracy is 95 percent with an initial training sample size of 20 for each key. The accuracy is 1.54 times higher than the state-of-the-art on-body typing system. Furthermore, when turning on the runtime calibration and adaptation system to update and enlarge the training sample size, the accuracy can reach around 98 percent on average during one month.

Smart wristband has become a dominant device in the wearable ecosystem, providing versatile functions such as fitness tracking, mobile payment, and transport ticketing. However, the small form-factor, low-profile hardware interfaces and computational resources limit their capabilities in security checking. Many wristband devices have recently witnessed alarming vulnerabilities, e.g., personal data leakage and payment fraud, due to the lack of authentication and access control. To fill this gap, we propose a secure text pin input system, namely Taprint, which extends a virtual number pad on the back of a user's hand. Taprint builds on the key observation that the hand "landmarks'', especially finger knuckles, bear unique vibration characteristics when being tapped by the user herself. It thus uses the tapping vibrometry as biometrics to authenticate the user, while distinguishing the tapping locations. Taprint reuses the inertial measurement unit in the wristband, "overclocks'' its sampling rate to extrapolate fine-grained features, and further refines the features to enhance the uniqueness and reliability. Extensive experiments on 128 users demonstrate that Taprint achieves a high accuracy (96%) of keystrokes recognition. It can authenticate users, even through a single-tap, at extremely low error rate (2.4%), and under various practical usage disturbances.

Wearable wristbands have become prevailing in the recent days because of their small and portable property. However, the limited size of the touch screen causes the problems of fat fingers and screen occlusion. Furthermore, it is not available for users whose hands are fully occupied with other tasks. To break this bottleneck, we propose a portable, hand-free and secure text-entry system, called FaceInput, which firstly uses a single small form factor sensor to accomplish a practical user input via facial vibrations. To sense the tiny facial vibration signals, we design and implement a double-stage amplifier whose maximum gain is 225. To enhance the input accuracy and robustness, we design a set of novel schemes for FaceInput based on the Mel-frequency cepstral coefficient (MFCC) concept and a hidden Markov model (HMM) to process the vibration signals, and an online calibration and adaptation scheme to recover the error due to temporal instability. Extensive experiments have been conducted on 30 human subjects during the period of one month. The results demonstrate that FaceInput can be successful to sense the tiny facial vibrations and robust to fight against various confounding factors. The average recognition accuracy is 98.2%. Furthermore, by enabling the runtime calibration and adaptation scheme that updates and enlarges the training data set, the accuracy can reach 100%.

The inevitable aging trend of the world's population brings a lot of challenges to the health care for the elderly. For example, it is difficult to guarantee timely rescue for a single-resided elder who falls at home. Under this circumstance, a reliable automatic fall detection machine is in great need for emergent rescue. However, the state-of-the-art fall detection systems are suffering from serious privacy concerns, having a high false alarm or being cumbersome for users. In this paper, we propose a device-free fall detection system, namely G-Fall, based on geophones. We first decompose the falling mode and characterize it with time-dependent floor vibration features. By leveraging Hidden Markov Model (HMM), our system is able to recognize the fall event precisely and achieve training-free recognition. It requires no training from the elderly but only an HMM template learned in advance through a small number of training samples. To reduce the false alarm rate, we propose a novel reconfirmation mechanism, namely Energy-of-Arrival (EoA) positioning to assist in recognizing a human's fall. Extensive experiments have been conducted on 12 human subjects. The results demonstrate that G-Fall achieves a 95.74% recognition precision with a false alarm rate of 5.30% on average. Furthermore, with the assistance of EoA, the false alarm rate is reduced to nearly 0%.

Nowadays, smart wristbands have become one of the most prevailing wearable devices as they are small and portable. However, due to the limited size of the touch screens, smart wristbands typically have poor interactive experience. There are a few works appropriating the human body as a surface to extend the input. Yet by using multiple sensors at high sampling rates, they are not portable and are energy-consuming in practice. To break this stalemate, we proposed a portable, cost efficient text-entry system, termed ViType, which firstly leverages a single small form factor sensor to achieve a practical user input with much lower sampling rates. To enhance the input accuracy with less vibration information introduced by lower sampling rate, ViType designs a set of novel mechanisms, including an artificial neural network to process the vibration signals, and a runtime calibration and adaptation scheme to recover the error due to temporal instability. Extensive experiments have been conducted on 30 human subjects. The results demonstrate that ViType is robust to fight against various confounding factors. The average recognition accuracy is 94.8% with an initial training sample size of 20 for each key, which is 1.52 times higher than the state-of-the-art on-body typing system. Furthermore, when turning on the runtime calibration and adaptation system to update and enlarge the training sample size, the accuracy can reach around 98% on average during one month.