In recent years, sensor-based continuous authentication on mobile devices has emerged as a promising solution for enhancing personal information security. However, many existing approaches depend on both legitimate and illegitimate user data for supervised training, which is time-consuming and often impractical in real-world deployments. To address these limitations, we propose CRAFTAuth, a sensor-based continuous Authentication system built upon Context Reconstruction and Adaptive Fusion Transformer. CRAFTAuth leverages built-in smartphone sensors of the accelerometer and gyroscope to capture behavioral biometrics in a non-intrusive manner. The system employs a self-supervised Transformer-based autoencoder to reconstruct long-term behavioral contexts from historical data, enabling the extraction of stable and discriminative features. To enhance adaptability, an iterative attention adaptive fusion mechanism dynamically integrates real-time behavioral inputs with long-term contextual features through progressive spatial-temporal refinement. In addition, a channel merging strategy is incorporated to compress feature representations and reduce computational complexity while preserving temporal dependencies, facilitating efficient deployment on resource-constrained mobile devices. Extensive experiments on our dataset demonstrate that CRAFTAuth achieves state-of-the-art performance, attaining 99.28% accuracy and 0.79% EER, while significantly reducing model size and inference latency compared with existing methods.

Palmprint recognition has been extensively studied as an effective biometric technique for personal identification. With the rapid development of deep neural networks (DNNs), palmprint recognition methods have achieved remarkable progress. However, their performance often deteriorates significantly under domain shifts. Moreover, existing unsupervised domain adaptation approaches for palmprint recognition typically suffer from unstable training and imprecise feature alignment, thereby limiting their effectiveness. To address these challenges, we propose SPA, a Stable and Precise Alignment framework for cross-domain palmprint recognition. Specifically, we design a lightweight yet robust Style Transformation Module (STM) to mitigate variations in style, color, and illumination. With the aid of STM, we further align joint feature distributions across all high-level layers, achieving more accurate feature alignment and enhancing recognition robustness. We conduct extensive experiments on two public multi-domain palmprint databases encompassing 42 cross-domain scenarios. The results demonstrate that SPA consistently delivers superior performance across both databases, achieving higher recognition accuracy with lower computational overhead compared to existing methods. In particular, SPA improves the average identification accuracies to 94.21% and 81.93%, while reducing the average equal error rates (EER) to 1.36% and 3.62% on the two databases, respectively.

In today’s digital era, ensuring the security of mobile devices is of critical importance. Sensor-based continuous authentication has emerged as an effective approach for protecting personal information on mobile devices. However, most existing systems rely primarily on time-domain features, overlooking valuable information from other domains and leading to incomplete feature representation. In this paper, we propose AMDFAuth, a deep reinforcement learning (DRL)-based Adaptive Multi-Domain Feature Fusion For continuous Authentication on smartphones that integrates a multi-domain feature extraction network with an adaptive feature fusion mechanism based on DRL. During user registration, AMDFAuth implicitly collects standardized behavioral data via built-in accelerometers and gyroscopes, and pre-trains a Diffusion Transformer (DiT) model. Through transfer learning, we integrate two additional feature extraction branches with the pre-trained DiT to construct a multi-domain network that captures time-domain, wavelet-domain, and key latent features. These features are then adaptively fused using DRL, enabling joint optimization of the feature extraction modules, fusion network, and an MLP classifier for user identification. During continuous authentication, real-time sensor data are collected and processed by the trained network and classifier to verify user identification. Extensive evaluations on our dataset demonstrate that AMDFAuth achieves 98.52% accuracy and an Equal Error Rate (EER) of 0.94% using a 2-second time window and 10 unseen users. These results highlight the system’s excellent accuracy, robustness, and generalization capability in real-world mobile authentication scenarios.

Real-world datasets often exhibit long-tailed distributions, raising important questions about how privacy risks evolve when machine learning (ML) models are applied to such data. In this work, we present a comprehensive analysis of membership inference attacks in long-tailed scenarios, revealing significant privacy vulnerabilities in tail data. We begin by examining standard ML models trained on long-tailed datasets and identify three key privacy risk effects: amplification, convergence, and polarization. Building on these insights, we extend our analysis to state-of-the-art long-tailed learning methods, such as foundation model-based approaches, offering new perspectives on how these models respond to membership inference attacks across head to tail classes. Finally, we investigate the privacy risks of ML models trained with differential privacy in long-tailed scenarios. Our findings corroborate that, even when ML models are designed to improve tail class performance to match head classes and are protected by differential privacy, tail class data remain particularly vulnerable to membership inference attacks.

Palm-vein recognition is gaining significant attention as a high-security biometric recognition technology. However, the vein image acquisition process is easily affected by several factors, making vein texture segmentation a challenging task. Recently, foundation models such as Segment Anything Model (SAM) have shown remarkable potential in image segmentation without requiring prior retraining. Nevertheless, due to the large domain discrepancy between the resource and target domains, as well as limited datasets, existing solutions that rely heavily on abundant training images often struggle to extract robust vein texture patterns. To address this challenge, we propose AdVeinSAM, an adversarial learning-based large model for palm-vein texture extraction, which leverages rich knowledge of large models to enhance vein pattern segmentation. Specifically, by alternately optimizing the vein segmentation model and the image generator, AdVeinSAM generates diverse training samples, effectively transferring knowledge from the large model to enhance feature extraction robustness. First, we incorporate the wavelet transform into xLSTM-UNet to design Wavelet-xLSTM-UNet, which generates diverse and realistic vein images for data augmentation. Then, we improve the NOLA model to fine-tune the segmentation anything model (SAM) and develop a specialized vein segmentation model (VeinSAM), which effectively extracts palm-vein texture features. Finally, the image generator (Wavelet-xLSTM-UNet) and the vein segmentation model (VeinSAM) are combined to form AdVeinSAM, where the generator and the VeinSAM are alternatively updated through adversarial training. Concretely, the image generator generates challenging samples to increase the segmentation difficulty for VeinSAM, while VeinSAM learns more robust feature representations from these challenging samples to improve the generalization and segmentation accuracy. We conduct extensive experiments on three public palm-vein databases and experimental results demonstrate that the proposed AdVeinSAM model outperforms state-of-the-art solutions, achieving the lowest equal error rates (EERs) of 1.48%, 4.76%, and 0.72%, respectively. These results confirm the effectiveness and robustness of AdVeinSAM in palm-vein texture extraction.

Vein recognition technology offers high security and privacy as an advanced biometric identification method. While deep learning techniques have achieved state-of-the-art performance in vein recognition due to their powerful pattern recognition capabilities, the Gated Recurrent Unit (GRU), a simplified version of LSTM, still faces limitations: 1) inability to process sequence information in parallel, leading to inefficient training; 2) loss of sensitivity to local features crucial for pattern recognition, despite excelling at modeling long-distance dependencies. To address these issues, we propose WTxGRN, a Wavelet Transform-based extended Gated Recurrent Network, which simultaneously extracts global and local features and supports parallel sequence processing. Specifically, we modify the GRU memory structure to enable parallel training and enhance feature representation through exponential gating and stabilization techniques, resulting in an extended GRU architecture called xGRU. We integrate xGRU into a wavelet transform-based residual backbone to form the xGRU Block. By incorporating a wavelet convolution branch and two Mixer Modules, we facilitate multi-scale feature extraction and fusion, enhancing vein recognition robustness and yielding the WTxGRU Block. Stacking these blocks constructs the WTxGRN. Furthermore, we present Spiking WTxGRN, an energy-efficient spiking version of WTxGRN, pioneering the application of spiking neural networks in vein recognition. Spiking WTxGRN offers high energy efficiency while maintaining excellent recognition performance, making it suitable for real-time vein recognition tasks. Extensive experiments on three public palm vein datasets demonstrate that our methods outperform state-of-the-art models across multiple benchmarks, achieving superior performance.

Nowadays, with the widespread adoption of mobile devices, information security has become particularly important. Existing sensor-based continuous authentication systems ensure the security of mobile devices to some extent, but most have drawbacks, such as lacking end-to-end structure or requiring data from both legitimate users and imposters for training. In this article, we present RPWAEAuth, a sensor-based continuous Authentication system using Reconstruction Probability in Wasserstein AutoEncoder. RPWAEAuth implicitly collects user behavior patterns from the built-in accelerometer, gyroscope, and magnetometer of mobile devices. The Wasserstein autoencoder maps the sensor data into a continuous latent space close to a prior distribution and reconstructs them using reconstruction probability for better authentication. In the registration stage, RPWAEAuth collects and preprocesses the sensor data from a legitimate user for RPWAE training. In the authentication stage, when a user interacts with the device, RPWAEAuth collects and preprocesses the sensor data, and then feeds them into the trained RPWAE to generate a reconstruction probability. This probability is then compared with a predefined threshold for user authentication. We evaluate the performance of RPWAEAuth on our dataset in terms of the effectiveness of RPWAEAuth, impact of sensor numbers, effectiveness of reconstruction probability, authentication time, resilience to mimic attacks, comparison with different AEs, and comparison with state-of-the-art methods. The experimental results demonstrate that RPWAEAuth achieves superior authentication performance compared to other methods, with an accuracy of 99.34% and an EER of 0.66% on 69 unseen users.

With the continual advancement of communication technologies, mobile devices have become indispensable tools in our daily lives. While existing sensor-based continuous authentication systems provide some level of user privacy protection, they often neglect the temporal characteristics of multisensor data and the unique information of each sensor. To further protect the privacy of mobile devices, we present FuMeAuth, a sensor-based continuous Authentication system using a Fused Memory-Augmented transformer Autoencoder. FuMeAuth leverages the built-in accelerometer, gyroscope, and magnetometer of smartphones to implicitly gather user behavior patterns. The Fused global–local Memory network (FuMe) effectively captures and adaptively combines the shared-private features of sensor data in FuMeAuth. During the registration phase, FuMeAuth collects and preprocesses the sensor data and sends the processed data to FuMe, which then records fused shared and private representations across different sensors for legitimate users. In the authentication phase, the trained FuMe reconstructs the current user’s data, computes the reconstruction error between user input data and the corresponding reconstructed data, and compares it against a predefined authentication threshold for authentication. We evaluate the performance of FuMeAuth on our data set in terms of the effectiveness of FuMeAuth, effect of sensor numbers, efficiency of fused memory module, and comparison with state-of-the-art approaches. The experimental results demonstrate that FuMeAuth exhibits superior performance than other approaches by achieving an accuracy of 99.84% and an equal error rate of 0.14% with 69 unseen users.

In recent years, sensor-based continuous authentication on mobile devices has proven highly effective in safeguarding personal information. However, these proposed approaches often require the utilization of both legitimate user and imposters’ data for training authentication models, which is time-consuming and ineffective. In this paper, we present AEGANAuth, a lightweight and effective AutoEncoder GAN-based continuous Authentication system for mobile devices using conditional variational AutoEncoder Generative Adversarial Network. AEGANAuth uses a Conditional Variational AutoEncoder Generative Adversarial Network (CVAEGAN) for data augmentation and utilizes an AutoEncoder Generative Adversarial Network (AEGAN) for user data reconstruction. During the enrollment phase, AEGANAuth employs the accelerometer and gyroscope sensors embedded on mobile devices to implicitly collect user behavioral patterns. Using the normalized sensor data, AEGANAuth selects legitimate user data to train CVAEGAN, which consists of a variational encoder, a conditional generator, a discriminator, and a classifier, for AEGAN training data augmentation. Based on the augmented legitimate user data, AEGAN, comprising an encoder, a decoder, and a discriminator, is trained for user data reconstruction. In the authentication phase, when a user operates the mobile device, AEGANAuth collects and normalizes the current user’s data, and then employs the trained AEGAN to reconstruct this user’s data. The reconstruction error is then computed by comparing the reconstructed data to the normalized data. Finally, AEGANAuth with AEGAN compares the reconstruction error to a predetermined authentication threshold for user authentication. We evaluate the performance of AEGANAuth on our dataset, and the experimental results demonstrate an average equal error rate (EER) of 2.13% and an average accuracy of 97.85% on 10 imposters.

Sensor-based continuous authentication mechanisms have demonstrated promising capabilities in enhancing the security of smart devices. In this article, we present SNNAuth, a novel sensor-based continuous Authentication system on smartphones that utilizes Spiking Neural Networks, leveraging biometric behavioral patterns captured by smartphone sensors. To enhance discriminative feature extraction, we introduce positional encoding into the time slicing of normalized sensor data. We design the artificial neural network (ANN)-SNN model, which transforms the trained ANN into an SNN by converting weights and activations into suitable spike neuron models and synaptic connections. The ANN-SNN model, designed for efficient computation and increased robustness, is specifically trained to extract temporal features of a legitimate user. With the extracted features of a legitimate user, we then train the one-class k-nearest neighbors (OC-kNN), which is employed for conducting the classification for all users. Based on the trained ANN-SNN and one-class kNN, SNNAuth determines whether the current user is legitimate or an imposter. Finally, we evaluate the performance of SNNAuth on two public data sets and our data set, and the experimental results demonstrate that SNNAuth outperforms state-of-the-art solutions by achieving the highest accuracy and the lowest equal error rates (EERs) on all three data sets.

Palm-vein identification is a highly secure pattern biometrics that has become an active research area in recent years. Despite the recent progress in deep neural networks (DNNs) for vein identification, existing solutions for feature representation continue to lack robustness due to the limited training samples. To address this limitation, data augmentation approaches, including Generative Adversarial Networks (GANs), have been investigated, but these schemes suffer from the following issues. First, it is practically unfeasible to use all the generated samples for classifier training due to the limited storage space and computation resources. Further, some of these generated samples may be non-representative or ineffective, seriously compromising models’ generalization capabilities. Second, the augmented dataset is fed to the target classifier repeatedly, resulting in overfitting after substantial training epochs. To tackle the above problems, we propose AdveinAU, an Adversarial vein AUtomatic AUgmentation approach that generates challenging samples to train a more robust vein classifier for palm-vein identification by alternatively optimizing the vein classifier and a set of latent variables. First, we consider a conditional deep convolution generative adversarial net (cDCGAN) to learn the distribution of real data and the generated data, and then a latent variable from the latent variable space is mapped to the sample space. Second, we combine the trained generator with the vein classifier to constitute AdveinAU, where the input sets of the generator and the classifier are alternatively updated by adversarial training. Specifically, a latent variable set is learned to increase the training loss of a target network through generating adversarial samples, while the classifier learns more robust features from harder examples to improve the generalization. To avoid collapsing inherent meanings of images, an exponential moving average (EMA) teacher andcosinesimilarity are employed for regularization to reduce the search space. Unlike previous works where GANs synthesize new realistic images, our model aims to search a latent variable set, based on which the generator can produce challenging samples along with the training process to improve the classifier’s performance. Finally, we conduct extensive experiments on three public palm-vein datasets to evaluate the performance of AdveinAU, and the experimental results demonstrate that the proposed AdveinAU is capable of generating harder samples to improve the performance of the vein classifier.

Finger-vein recognition has attracted extensive attention due to its exceptional level of security and privacy. Recently, deep neural networks (DNNs), such as convolutional neural networks (CNNs) showing robust capacity for feature representation, have been proposed for vein recognition. The architectures of these DNNs, however, have primarily been manually designed based on human prior knowledge, which is both time-consuming and error-prone. To overcome these problems, we propose AG-NAS, an Attention Gated recurrent unit-based Neural Architecture Search to automatically search for the optimal network architecture, thereby improving the recognition performance for different finger-vein recognition tasks. First, we combine the self-attention mechanism and gated recurrent unit (GRU) to propose an attention GRU module employed as a controller to generate the architectural hyperparameters of candidate neural networks automatically. Second, we investigate a parameter-sharing supernet policy to reduce the search space, computation, and time costs. Finally, we conduct rigorous experiments on our finger-vein database and two public finger-vein databases. The experimental results demonstrate that the proposed AG-NAS outperforms the representative approaches and achieves state-of-the-art recognition accuracy.

Finger-vein biometrics has recently gained significant attention due to its robust privacy and high security features. Despite notable advancements, most existing methods focus on extracting features from a 2-dimensional (2D) image projected from 3D vein vessels with a single view. However, recognition based on a single view is prone to errors due to variations in finger positioning, especially those caused by finger roll movements, which can degrade recognition performance. To address this challenge, we propose ABLSTM-TSVT, an Attention Bidirectional LSTM-based Temporal-Spatial Vein Transformer for multi-view finger-vein recognition. First, we enhance LSTM with an attention mechanism to create an attention LSTM for extracting temporal features. We further improve this by introducing a local attention module, which learns temporal dependencies between a patch (token) and its adjacent patches across multiple views, integrating it with the attention LSTM to form a temporal attention module. Second, we develop a spatial attention module that captures the spatial dependencies of patches within an image. Finally, merging the temporal and the spatial attention modules, we create our temporal-spatial transformer model, which effectively represents features from multi-view images. Experimental results on two multi-view datasets demonstrate that our approach outperforms state-of-the-art approaches in enhancing identification accuracy and reducing verification errors in vein classifiers.

The exceptional feature extraction capabilities of deep neural networks (DNNs) have significantly advanced palmprint recognition. However, DNNs typically require training and testing data originate from the same distribution, which limits their practical applications. Moreover, existing unsupervised domain adaptation methods struggle to achieve high accuracy with efficiency. To address these challenges, we propose LSFM, an efficient Light Style and Feature Matching method that enhances palmprint recognition performance in cross-domain scenarios with fewer resources. Specifically, we develop an efficient style transfer model to mitigate domain shifts at the pixel level. We then align features across multiple task-specific layers in high dimensional space to reduce domain discrepancies, further improving cross-domain performance. Finally, we evaluate the effectiveness of the proposed LSFM through extensive experiments on two public multi-domain palmprint databases. The experimental results demonstrate that LSFM achieves superior performance with significantly reduced resource consumption, improving average accuracy to 94.87% and lowering the average equal error rate to 1.46%, while saving over 80% of resources.

Over the last years, sensor-based continuous authentication on mobile devices has achieved great success on personal information protection. These proposed mechanisms, however, require both legal and illegal users’ data for authentication model training, which takes time and is impractical. In this paper, we present MAuGANs, a lightweight and practical Memory-Augmented Autoencoder-based continuous Authentication system on smartphones with conditional transformer Generative Adversarial Networks (GANs), where the conditional transformer GANs (CTGANs) are used for data augmentation and the memory-augmented autoencoder (MAu) is utilized to identify users. Specifically, MAuGANs exploits the smartphone built-in accelerometer and gyroscope sensors to implicitly collect users’ behavioral patterns. With the normalized legitimate user's sensor data, MAuGANs uses a CTGAN composed of a conditional transformer-based generator and a conditional transformer-based discriminator to create additional training data for the MAu. Then, the MAu is trained on the augmented legitimate user's data. The trained MAu reconstructs the current user data and then calculates the reconstruction error between the reconstructed data and current user data. To carry out user authentication, MAuGANs compares the reconstruction error with a predefined authentication threshold. We evaluate the performance of MAuGANs on our dataset, where our extensive experiments demonstrate that MAuGANs reaches the best authentication performance, when comparing with the representative state-of-the-art methods, by 0.33% EER and 99.65% accuracy on 10 unseen users.

This work investigates the truthful auction for dependent task offloading in vehicular edge computing by considering the selfishness and rationality of participating nodes. Specifically, we first illustrate a truthfulness-guaranteed dependent task offloading architecture. Then, we formulate the Truthfulness-Guaranteed Dependent Task Offloading problem, aiming at maximizing the system utility (SU) while ensuring truthfulness and individual rationality in dynamic environments. Further, we design both centralized and distributed auction mechanisms to derive the optimal and approximate solutions, respectively. For centralized auction mechanism, we adopt the branch-and-price algorithm to determine the offloaded nodes, which yields maximum SU. Then, we adopt VCG mechanism to determine the payment of buyers. For distributed auction mechanism, each seller independently chooses the winning bid, and the buyer greedily chooses the offloaded node with maximum utility. Then, a novel payment mechanism regarding the cost of failed buyers is designed to guarantee the truthfulness and individual rationality. Finally, we build the simulation model and conduct the performance evaluation based on realistic vehicular trajectories. The results demonstrate that the proposed distributed auction mechanism achieves performance within approximately 4% of the optimal method, while significantly reducing computational complexity. Additionally, it significantly outperforms other methods in terms of system utility across various task requirements.

With the rapid development of the Internet of Things (IoTs) and mobile communications, mobile devices have become indispensable in our daily lives. Given the substantial amount of private information stored on these devices, the security of mobile devices has emerged as a significant concern for users. Different from conventional methods such as PINs, fingerprints, and face IDs, which authenticate users only during the initial login stage, continuous authentication ensures consistent verification while mobile devices are in use. Current continuous authentication methods require extensive data from a series of users for effective training. Nevertheless, it is challenging to collect sufficient amount of data within a limited time. In this paper, we propose CALL, an unsupervised sensor-based Continuous Authentication system with a Low-rank transformer using Learning-to-rank algorithms. The lightweight CALL is capable of providing both spatial and temporal features for end-to-end authentication. Specifically, CALL utilizes time series data from a legitimate user, collected by the accelerometer, gyroscope, and magnetometer sensors on smartphones, to train a pure one-dimensional autoencoder for spatial features and a shuffle low-rank Transformer (SLRT) for temporal features in the training phase. In the authentication phase, the trained pure one-dimensional autoencoder captures spatial features by reconstructing input data to obtain the reconstruction error, and SLRT captures temporal features by predicting a ranking vector that reveals the order of the shuffled feature sequence. The predicted ranking vector is then used to recover the shuffled sequence and the similarity between the frequency spectrum sequences of the recovered sequence and the original time series data is calculated. The reconstruction error and similarity are compared against pre-defined thresholds, and CALL authenticates a user as legitimate only if both values fall below their respective thresholds. Finally, we evaluate the performance of CALL on UCI_HAR, WISDM_HARB, and our dataset, and the extensive experiments illustrate that CALL reaches the best performance with 96.43%, 95.24% and 96.92% accuracy, and 4.28%, 4.76% and 3.86% EERs on the three datasets, outperforming state-of-the-art continuous authentication methods.

Deep neural networks can be easily fooled by the adversarial example, which is a specially crafted example with subtle and intentional perturbations. A plethora of papers have proposed to use filters to effectively defend against adversarial example attacks. However, we demonstrate that the automatic filter-based defenses may not be reliable. In this article, we present URL2AED, Using a Reinforcement Learning scheme TO escape the automatic filter-based Adversarial Example Defenses. Specifically, URL2AED uses a specially crafted policy gradient reinforcement learning (RL) algorithm to generate adversarial examples (AEs) that can escape automatic filter-based AE defenses. In particular, we properly design reward functions in policy-gradient RL for targeted attacks and non-targeted attacks, respectively. Furthermore, we customize training algorithms to reduce the possible action space in policy-gradient RL to accelerate URL2AED training while still ensuring that URL2AED generates successful AEs. To demonstrate the performance of the proposed URL2AED, we conduct extensive experiments on three public datasets in terms of different perturbation degrees of parameter, different filter parameters, transferability, and time consumption. The experimental results show that URL2AED achieves high attack success rates for automatic filter-based defenses and good cross-model transferability.

Graph embedding converts a graph data into a low dimensional space to preserve the original graph information. However, graph data can be reconstructed by malicious adversaries to train machine learning models from graph embeddings. This paper studies to what extent an adversary (without the original graph data) can recover the original graph data from graph embeddings. To quantify the original graph information leakage from graph embeddings, we develop a deep neural network model InferNet that can be used by adversaries to infer the original graph information from an adversary-accessible graph embedding database. More specifically, we propose the data-free reversed knowledge distillation technique to support the InferNet training even if the original graph dataset is absent. To ensure the performance of InferNet, we design two cycle-consistency loss functions to have an interactive training of InferNet over three series of datasets. To further enhance the performance of InferNet, we provide a joint training algorithm that simultaneously trains the pseudo-sample generator and InferNet, which significantly reduces the storage space. We evaluate the performance of InferNet on three datasets, and the intensive experiments demonstrate that InferNet can infer the original graph information from the graph embedding dataset with high accuracy.

This paper studies to what extent an adversary (without the original graph data) can recover the original raw graph data from graph embeddings. To quantify the original graph data information leakage from graph embeddings, we develop a deep neural network model InferNet that can be used by adversaries to infer the original graph data information from an adversary-accessible graph embedding database. Specifically, we propose the data-free reversed knowledge distillation (KD) technique to support InferNet training even if the original graph dataset is absent. To improve the performance of InferNet, we design two cycle-consistency loss functions to have an interactive training of InferNet over three series of datasets. Our intensive experiments demonstrate that InferNet can infer the original graph data information from the graph embedding dataset with high accuracy.

Deep learning has achieved significant success on intelligent medical treatments, such as automatic diagnosis and analysis of medical data. To train an automatic diagnosis system with high accuracy and strong robustness in healthcare, sufficient training data are required when using deep learning-based methods. However, given that the data collected by sensors that are embedded in medical or mobile devices are inadequate, it is challenging to train an effective and efficient classification model with state-of-the-art performance. Inspired by generative adversarial networks (GANs), we propose TS-GAN, a Time-series GAN architecture based on long short-term memory (LSTM) networks for sensor-based health data augmentation, thereby improving the performance of deep learning-based classification models. TS-GAN aims to learn a generative model that creates time-series data with the same space and time dependence as the real data. Specifically, we design an LSTM-based generator for creating realistic data and an LSTM-based discriminator for determining how similar the generated data are to real data. In particular, we design a sequential-squeeze-and-excitation module in the LSTM-based discriminator to better understand space dependence of real data, and apply the gradient penalty originated from Wasserstein GANs in the training process to stabilize the optimization. We conduct comparative experiments to evaluate the performance of TS-GAN with TimeGAN, C-RNN-GAN and Conditional Wasserstein GANs through discriminator loss, maximum mean discrepancy, visualization methods and classification accuracy on health datasets of ECG_200, NonInvasiveFatalECG_Thorax1, and mHealth, respectively. The experimental results show that TS-GAN exceeds other state-of-the-art time-series GANs in almost all the evaluation metrics, and the classifier trained on synthetic datasets generated by TS-GAN achieves the highest classification accuracy of 97.50% on ECG_200, 94.12% on NonInvasiveFatalECG_Thorax1, and 98.12% on mHealth, respectively.

To address the imbalance problem between supply and demand for taxis and passengers, this paper proposes a distributed ensemble empirical mode decomposition with normalization of spatial attention mechanism based bi-directional gated recurrent unit (EEMDN-SABiGRU) model on Spark for accurate passenger hotspot prediction. It focuses on reducing blind cruising costs, improving carrying efficiency, and maximizing incomes. Specifically, the EEMDN method is put forward to process the passenger hotspot data in the grid to solve the problems of non-smooth sequences and the degradation of prediction accuracy caused by excessive numerical differences, while dealing with the eigenmodal EMD. Next, a spatial attention mechanism is constructed to capture the characteristics of passenger hotspots in each grid, taking passenger boarding and alighting hotspots as weights and emphasizing the spatial regularity of passengers in the grid. Furthermore, the bi-directional GRU algorithm is merged to deal with the problem that GRU can obtain only the forward information but ignores the backward information, to improve the accuracy of feature extraction. Finally, the accurate prediction of passenger hotspots is achieved based on the EEMDN-SABiGRU model using real-world taxi GPS trajectory data in the Spark parallel computing framework. The experimental results demonstrate that based on the four datasets in the 00-grid, compared with LSTM, EMD-LSTM, EEMD-LSTM, GRU, EMD-GRU, EEMD-GRU, EMDN-GRU, CNN, and BP, the mean absolute percentage error, mean absolute error, root mean square error, and maximum error values of EEMDN-SABiGRU decrease by at least 43.18%, 44.91%, 55.04%, and 39.33%, respectively.

Multi-view finger-vein recognition technology has attracted increasing attentions in recent years. Despite recent advances in the multi-view finger-vein identification, existing solutions employ multiple monocular cameras from different views to record two-dimensional (2D) projections of 3D vein vessels, which causes the following problems: 1) 2D images collected from limited views (two or three views) are insufficient for robust 3D vein vessel feature representation. Furthermore, image sequences of the same finger acquired from different views usually show significant differences. As a result, the existing works are still sensitive to positional variations of the fingers, specifically those caused by finger roll movements. 2) Using multiple cameras can lead to increased costs. Moreover, it is impossible to employ several cameras to acquire full-view images because of the limited space on capturing devices. To address the above issues, we present$\mathbb {FV}$-LT, a Full-View Finger-Vein identification system based on a Local attention Transformer, by implementing an image acquisition device with a single camera. First, we design and implement a finger-vein acquisition prototype device that utilizes a single camera and a LED group to rotate along a finger for full-view image collection. This allows capturing all vein patterns concealed beneath human skin to form a complete representation of finger features. Second, given the full-view vein images, we propose a local attention transformer-based approach to extract dependency features of a token (a patch or an image) on its neighborhood’s tokens among image patches and among full-view images, respectively. These dependency features are shown to be robust to positional variations induced by finger rolls. Based on the public database of full-view finger-vein images captured by our designed device and a single-view database, we verify the performance of the proposed$\mathbb {FV}$-LT. The experimental results show that$\mathbb {FV}$-LT significantly outperforms existing 2D/multi-view based approaches with respect to improving the tolerance against finger roll and achieving the state-of-the-art identification accuracy.

Vein biometrics is a high security and privacy preserving identification technology that has attracted increasing attention over the last decade. Deep neural networks (DNNs), such as convolutional neural networks (CNN), have shown strong capabilities for robust feature representation, and have achieved, as a result, state-of-the-art performance on various vision tasks. Inspired by their success, deep learning models have been widely investigated for vein recognition and have shown significant improvement of identification accuracy compared to handcrafted models. Existing deep learning models, however, are vulnerable to adversarial perturbation attacks, where thoughtfully crafted small perturbations can cause misclassification of legitimate images, degrading, thereby, the efficiency of vein recognition systems. To address this problem, we propose, in this paper, VeinGuard, a novel defense framework to defend deep learning classifiers against adversarial palm-vein image attacks, composed of a local transformer-based GAN and a purifier. VeinGuard comprises two components: a local transformer-based GAN (LTGAN) that learns the distribution of unperturbed vein images and generates high-quality palm-vein images, and a purifier consisting of a trainable residual network and of a pre-trained generator from LTGAN that automatically removes a wide variety of adversarial perturbations. The resulting clean images are fed to vein classifiers for identification, thereby avoiding adversarial attacks. We evaluate VeinGuard on three public vein datasets in terms of white-box attacks, black-box attacks, ablation experiments, and computation time. The experimental results show that VeinGuard allows filtering the perturbations and enables the classifiers to achieve state-of-the-art recognition results for different adversarial attacks.

Mobile devices are becoming increasingly popular and are playing significant roles in our daily lives. Insufficient security and weak protection mechanisms, however, cause serious privacy leakage of the unattended devices. To fully protect mobile device privacy, we propose ADFFDA, a novel mobile continuous authentication system using an Adaptive Deep Feature Fusion scheme for effective feature representation, and a transformer-based GAN for Data Augmentation, by leveraging smartphone built-in sensors of the accelerometer, gyroscope and magnetometer. Given the normalized sensor data, ADFFDA utilizes the transformer-based GAN consisting of a transformer-based generator and a CNN-based discriminator to augment the training data for CNN training. With the augmented data and the especially-designed CNN based on the ghost module and ghost bottleneck, ADFFDA extracts deep features from the three sensors by the trained CNN, and exploits an adaptive-weighted concatenation method to adaptively fuse the CNN-extracted features. Based on the fused features, ADFFDA authenticates users by using the one-class SVM (OC-SVM) classifier. We evaluate the authentication performance of ADFFDA in terms of the efficiency of the transformer-based GAN, GAN-based data augmentation, CNN architecture, adaptive-weighted feature fusion, OC-SVM classifier, and security analysis. The experimental results show that ADFFDA obtains the best authentication performance w.r.t representative approaches, by achieving a mean equal error rate of 0.01%.

Mobile devices have been playing significant roles in our daily lives, which has made device security and privacy protection extremely important. These mobile devices storing user sensitive and private information, therefore, need rigorous user authentication mechanisms. In this article, we present SearchAuth, a novel continuous authentication system on smartphones exploiting a neural architecture search (NAS) to find an optimal network architecture and an auto augmentation search (AAS) to more effectively train the optimal network along with the best data augmentation policies, by leveraging the accelerometer, gyroscope, and magnetometer on smartphones to capture users’ behavioral patterns. Specifically, SearchAuth consists of three stages, i.e., the offline stage, registration stage, and authentication stage. In the offline stage, we utilize the NAS on sensor data of the accelerometer, gyroscope, and magnetometer to find an optimal network architecture based on the designed search space. With the optimal network architecture, namely, NAS-based model, the AAS automatically optimizes the augmentation of the input data for more effectively training the model that is for feature extraction. In the registration stage, we use the trained NAS-based model to learn and extract deep features from the legitimate user’s data, and train the LOF classifier with 55 features selected by the PCA. In the authentication stage, with the well-trained NAS-based model and LOF classifier, SearchAuth identifies the current user as a legitimate user or an impostor when the user starts operating a smartphone. Based on our dataset, we evaluate the performance of the proposed SearchAuth, and the experimental results demonstrate that SearchAuth surpasses the representative authentication schemes by achieving the best accuracy of 93.95%, F1-score of 94.30%, and EER of 5.30% on the LOF classifier with dataset size of 100.

With the widespread usage of mobile devices, the authentication mechanisms are urgently needed to identify users for information leakage prevention. In this article, we present CAGANet, a convolutional neural network (CNN)-based continuous authentication on smartphones using a conditional Wasserstein generative adversarial network (CWGAN) for data augmentation, which utilizes smartphone sensors of the accelerometer, gyroscope, and magnetometer to sense phone movements incurred by user operation behaviors. Specifically, based on the preprocessed real data, CAGANet employs CWGAN to generate additional sensor data for data augmentation that are used to train the designed CNN. With the augmented data, CAGANet utilizes the trained CNN to extract deep features and then performs principal component analysis (PCA) to select appropriate representative features for different classifiers. With the CNN-extracted features, CAGANet trains four one-class classifiers of OC-SVM, LOF, isolation forest (IF), and EE in the enrollment phase and authenticates the current user as a legitimate user or an impostor based on the trained classifiers in the authentication phase. To evaluate the performance of CAGANet, we conduct extensive experiments in terms of the efficiency of CWGAN, the effectiveness of CWGAN augmentation and the designed CNN, the accuracy on unseen users, and comparison with traditional augmentation approaches and with representative authentication methods, respectively. The experimental results show that CAGANet with the IF classifier can achieve the lowest equal error rate (EER) of 3.64% on 2-s sampling data.

With the rapid development of data-driven intelligent transportation systems, an efficient route recommendation method for taxis has become a hot topic in smart cities. We present an effective taxi route recommendation approach (called APFD) based on the artificial potential field (APF) method and Dijkstra method with mobile trajectory big data. Specifically, to improve the efficiency of route recommendation, we propose a region extraction method that searches for a region including the optimal route through the origin and destination coordinates. Then, based on the APF method, we put forward an effective approach for removing redundant nodes. Finally, we employ the Dijkstra method to determine the optimal route recommendation. In particular, the APFD approach is applied to a simulation map and the real-world road network on the Fourth Ring Road in Beijing. On the map, we randomly select 20 pairs of origin and destination coordinates and use APFD with the ant colony (AC) algorithm, greedy algorithm (A*), APF, rapid-exploration random tree (RRT), non-dominated sorting genetic algorithm-II (NSGA-II), particle swarm optimization (PSO), and Dijkstra for the shortest route recommendation. Compared with AC, A*, APF, RRT, NSGA-II, and PSO, concerning shortest route planning, APFD improves route planning capability by 1.45%–39.56%, 4.64%–54.75%, 8.59%–37.25%, 5.06%–45.34%, 0.94%–20.40%, and 2.43%–38.31%, respectively. Compared with Dijkstra, the performance of APFD is improved by 1.03–27.75 times in terms of the execution efficiency. In addition, in the real-world road network, on the Fourth Ring Road in Beijing, the ability of APFD to recommend the shortest route is better than those of AC, A*, APF, RRT, NSGA-II, and PSO, and the execution efficiency of APFD is higher than that of the Dijkstra method.

Smartphones have become crucial and important in our daily life, but the security and privacy issues have been major concerns of smartphone users. In this article, we present DeFFusion, a CNN-based continuous authentication system using Deep Feature Fusion for smartphone users by leveraging the accelerometer and gyroscope ubiquitously built into smartphones. With the collected data, DeFFusion first converts the time domain data into frequency domain data using the fast Fourier transform and then inputs both of them into a designed CNN, respectively. With the CNN-extracted features, DeFFusion conducts the feature selection utilizing factor analysis and exploits balanced feature concatenation to fuse these deep features. Based on the one-class SVM classifier, DeFFusion authenticates current users as a legitimate user or an impostor. We evaluate the authentication performance of DeFFusion in terms of impact of training data size and time window size, accuracy comparison on different features over different classifiers and on different classifiers with the same CNN-extracted features, accuracy on unseen users, time efficiency, and comparison with representative authentication methods. The experimental results demonstrate that DeFFusion has the best accuracy by achieving the mean equal error rate of 1.00% in a 5-second time window size.

In recent years, there have been increasing demands for using deep neural networks (DNNs) to provide image processing services for mobile devices. Considering the privacy of users' images, we utilize a two-tiers DNN which deploys the shallow and deep model on mobile devices and the cloud respectively. Then we propose a novel privacy protection mechanism which is deployed on the mobile device to satisfy the differential privacy. Meanwhile, based on the convolution kernel analysis, we also propose a novel method to improve the computation efficiency of mobile devices. The highlight of our mechanism is that it not only provides customized privacy protection which can resist the attack of Generative Adversarial Network (GAN), but also improves the accuracy of the neural network model. The experimental results on the ImageNet dataset show that we have improved the top-5 accuracy of image classification by 2%-3%. Under the premise of ensuring that the accuracy of the network is not degraded, our method reduces the CPU consumption on the VGG16 and ResNet50 networks to 74.6% and 48.9%, respectively, and can reduce 90% of the memory overhead. This improvement makes it possible to enable mobile deep neural network applications.

Despite recent advances of deep neural networks in hand vein identification, the existing solutions assume the availability of a large and rich set of training image samples. These solutions, therefore, still lack the capability to extract robust and discriminative hand-vein features from a single training image sample. To overcome this problem, we propose a single-sample-per-person (SSPP) palm-vein identification approach, where only a single sample per class is enrolled in the gallery set for training. Our approach, named MSMDGAN + CNN, consists of a multi-scale and multi-direction generative adversarial network (MSMDGAN) for data augmentation and a convolutional neural network (CNN) for palm-vein identification. First, a novel data augmentation approach, MSMDGAN, is developed to learn the internal distribution of patches in a single image. The proposed MSMDGAN consists of multiple fully convolutional GANs, each of which is responsible for learning the patch distribution within an image at a different scale and at a different direction. Second, given the resulting augmented data by MSMDGAN, we design a CNN for single sample palm-vein recognition. The experimental results on two public hand-vein databases demonstrate that MSMDGAN is able to generate realistic and diverse samples, which, in turn, improves the stability of the CNN. In terms of accuracy, MSMDGAN + CNN outperforms other representative approaches and achieves state-of-the-art recognition results.

Continuous authentication monitors the security of a system throughout the login session on mobile devices. In this article, we present SCANet, a two-stream convolutional neural network--based continuous authentication system that leverages the accelerometer and gyroscope on smartphones to monitor users’ behavioral patterns. We are among the first to use two streams of data—frequency domain data and temporal difference domain data—from the two sensors as the inputs of the convolutional neural network (CNN). SCANet utilizes the two-stream CNN to learn and extract representative features and then performs the principal component analysis to select the top 25 features with high discriminability. With the CNN-extracted features, SCANet exploits the one-class support vector machine to train the classifier in the enrollment phase. Based on the trained CNN and classifier, SCANet identifies the current user as a legitimate user or an impostor in the continuous authentication phase. We evaluate the effectiveness of the two-stream CNN and the performance of SCANet on our dataset and BrainRun dataset, and the experimental results demonstrate that CNN achieves 90.04% accuracy, and SCANet reaches an average of 5.14% equal error rate on two datasets and takes approximately 3 s for user authentication.

As personal computing platforms, smartphones are commonly used to store private, sensitive, and security information, such as photographs, emails, and Android Pay. To protect such information from adversaries, continuous authentication on smartphone users becomes more and more important. In this paper, we present a novel authentication system, SensorAuth, for continuous authentication of users based on their behavioral patterns, by leveraging the accelerometer and gyroscope ubiquitously built into smartphones. We are among the first to exploit five data augmentation approaches including permutation, sampling, scaling, cropping, and jittering to create additional data by applying them on training data. With the augmented data, SensorAuth extracts sensor-based features in both time and frequency domains within a time window, then utilizes the one-class support vector machine to train the classifier, and finally authenticates users. We evaluate the authentication performance of SensorAuth in terms of the impact of window size, accuracy on each of and combinations of data augmentation approaches, time efficiency, energy consumption, and comparisons with the representative classifiers and with the existing approaches, respectively. The experimental results show that SensorAuth performs highly accurate and time-efficient continuous authentication, by reaching the lowest median equal error rate of 4.66%, and consuming a short authentication time of approximately 5 s.

Garbled circuits are general tools that allow two parties to compute any function without disclosing their respective inputs. Applications of this technique vary from distributed privacy-preserving machine learning tasks to secure outsourced authentication. Unfortunately, the energy cost of garbled circuit evaluation protocols is substantial. This limits the applicability of garbled circuits in scenarios that involve battery-operated devices, such as Internet-of-Things (IoT) devices and smartphones. In this paper, we propose MEG, a Memory- and Energy-efficient Garbled circuit evaluation mechanism. MEG utilizes batch data transmission and multi-threading to reduce memory and energy consumption. We implement MEG on an Android smartphone and compare its performance and energy consumption with state-of-the-art techniques using two garbled circuits of widely different sizes (AES-128 and 256-bit edit distance). Our results show that, compared with “plain” garbled circuit evaluation, MEG decreases memory consumption by more than 90%. When compared with current pipelined garbled circuit evaluation techniques, MEG's energy usage was 42% lower for AES-128 and 23% lower for EDT-256. Furthermore, our multi-thread implementation of MEG decreased circuit evaluation time by up to 56.7% for AES-128, and by up to 13.5% for EDT-256, compared with state-of-the-art pipelining techniques.

The security of smartphones is vital as much privacy information is stored on them. In this paper, we propose a touch-based authentication system by exploiting a novel one-class classification algorithm import vector domain description (IVDD). We record each complete stroke by a 32-dimensional attribute vector and select features with high discriminability by employing the correlation analysis and the conditional mutual information maximization. With the selected features, the trained IVDD classifier generates a probabilistic result for each user, which is further used for authentication based on a predefined threshold. We evaluate the classifier IVDD in terms of the impact of the number of touch behaviors, computation time, individual effect, and accuracy, and the experimental results show that the IVDD reaches the lower BER of 2.5% under 15 touches and lower mean FRR of 2.14%, comparing with the support vector domain description (SVDD).

We present a two-stream convolutional neural network based authentication system, CNNAuth, for continuously monitoring users' behavioral patterns, by leveraging the accelerometer and gyroscope on smartphones. We are among the first to exploit two streams of the time-domain data and frequency-domain data from raw sensor data for learning and extracting universal effective and efficient feature representations as the inputs of the convolutional neural network (CNN), and the extracted features are further selected by the principal component analysis (PCA). With these features, we use the one-class support vector machine (SVM) to train the classifier in the enrollment phase, and with the trained classifier and testing features, CNNAuth classifies the current user as a legitimate user or an impostor in the continuous authentication phase. We evaluate the performance of the two-stream CNN and CNNAuth, respectively, and the experimental results show that the two-stream CNN achieves an accuracy of 87.14%, and CNNAuth reaches the lowest authentication EER of 2.3% and consumes approximately 3 seconds for authentication.

People watching smartphones while walking causes a significant impact to their safety. Pedestrians staring at smartphone screens while walking along the sidewalk are generally more at risk than other pedestrians not engaged in smartphone usage. In this study, the authors propose Safe Walking , an Android smartphone‐based system that detects the walking behaviour of pedestrians by leveraging the sensors and front camera on smartphones, improving the safety of pedestrians staring at smartphone screens. More specifically, Safe Walking first exploits a pedestrian speed calculation algorithm by sampling acceleration data via the accelerometer and calculating gravity components via the gravity sensor. Then, this system utilises a greyscale image detection algorithm to detect the face and eye movement modes based on OpenCV4Android to determine if pedestrians are staring at the screens. Finally, Safe Walking generates a vibration by a vibrator on smartphones to alert pedestrians to pay attention to road conditions. The authors implemented Safe Walking on an Android smartphone and evaluated pedestrian walking speed, the accuracy of eye movement, and system performance. The results show that Safe Walking can prevent the potential danger for pedestrians staring at smartphone screens with a true positive rate of .

Mobile video chat apps offer users an approachable way to communicate with others. As high-speed 4G networks are being deployed worldwide, the number of mobile video chat app users increases. However, video chatting on mobile devices brings users financial concerns, since streaming video demands high bandwidth and can use up a large amount of data in dozens of minutes. Lowering the bandwidth usage of mobile video chats is challenging since video quality may be compromised. In this paper, we attempt to tame this challenge. Technically, we propose a context-aware frame rate adaption framework, named low-bandwidth video chat (LBVC). It follows a sender-receiver cooperative principle that smartly handles the tradeoff between lowering bandwidth usage and maintaining video quality. We implement LBVC by modifying an open-source app–Linphone– and evaluate it with both objective experiments and subjective studies.

In a Body Sensor Network (BSN) activity recognition system, sensor sampling and communication quickly deplete battery reserves. While reducing sampling and communication saves energy, this energy savings usually comes at the cost of reduced recognition accuracy. To address this challenge, we propose AdaSense, a framework that reduces the BSN sensors sampling rate while meeting a user-specified accuracy requirement. AdaSense utilizes a classifier set to do either multi-activity classification that requires a high sampling rate or single activity event detection that demands a very low sampling rate. AdaSense aims to utilize lower power single activity event detection most of the time. It only resorts to higher power multi-activity classification to find out the new activity when it is confident that the activity changes. Furthermore, AdaSense is able to determine the optimal sampling rates using a novel Genetic Programming algorithm. Through this Genetic Programming approach, AdaSense reduces sampling rates for both lower power single activity event detection and higher power multi-activity classification. With an existing BSN dataset and a smartphone dataset we collect from eight subjects, we demonstrate that AdaSense effectively reduces BSN sensors sampling rate and outperforms a state-of-the-art solution in terms of energy savings.

In this paper, we present an optimal packet size solution that optimizes the communication energy consumption in the heterogeneous wireless networks. More specifically, we consider a heterogeneous network system composed of a body sensor network (BSN) and a WiFi network. Then, based on the analysis of data communication in the BSN and WiFi (BSN-WiFi) network, we formulate a communication energy consumption optimization model with the constraints of throughput and time delay. Mathematically, we convert this model into a geometric programming problem, which is then numerically solved. The optimal solution can be applied in both BSN and WiFi network to dynamically select packet payload sizes according to real-time packet delivery ratios (PDRs). Since PDRs are time-varying, we tabulate a packet payload size lookup table for online packet size selection using PDRs as indices. Finally, we collect PDRs from a deployed BSN-WiFi network and evaluate the energy optimization model. The performance evaluation results show that, in comparison with fixed packet size solutions, our optimal solutions achieve up to 70 percent energy savings in a BSN(TDMA)-WiFi network and 68 percent in a BSN(CSMA)-WiFi network.

Body sensor networks (BSNs) and WiFi networks have been widely investigated due to the availability of sensor motes and WiFi devices, but they are commonly deployed separately. In this paper we propose to optimize the total communication energy consumption of BSN and WiFi (BSN-WiFi) networks using joint data rate adaptation. More specifically, we first elaborate the BSN-WiFi network system in four consecutive phases. Then based on the system, we analyze the communication energy consumption, throughput and time delay, and provide a signal-to-noise ratio and packet delivery ratio (SNR-PDR) mappings of BSN and WiFi networks. Next, we build an energy optimization model with constraints of SNR-PDR mappings, throughput, and time delay to minimize the total communication energy consumption in BSN-WiFi networks. With the input of SNR values, we solve this model by cvx to obtain the output of optimal data rates associated with SNR values, which are then tabulated for online data rate adaptation. Finally, we collect 20-minute traces from a specific BSN-WiFi network system for performance evaluation, and the results demonstrate that our optimal data rate solution achieves up to 86% energy savings comparing with the solutions using fixed data rates.

Automatically recognizing human activities in a body sensor network (BSN) enables many human-centric applications. Many current works recognize human activities through collecting and analyzing sensor readings from on-body sensor nodes. These sensing-based solutions face a dilemma. On one hand, to guarantee data availability and recognition accuracy, sensing-based solutions have to either utilize a high transmission power or involve a packet retransmission mechanism. On the other hand, enhancing the transmission power increases a sensor node's energy overheads and communication range. The enlarged communication range in consequence increases privacy risks. A packet retransmission mechanism complicates on-body sensor nodes' MAC layer and hence increases energy overheads. In contrast to the sensing-based solutions, we build Radio Sense, a prototype system that exploits wireless communication patterns for BSN activity recognition. Using Radio Sense, we benchmark three system parameters (transmission (TX) power, packet sending rate, and smoothing window size) to design algorithms for system parameter selection. The algorithms aim to balance accuracy, latency, and energy overheads. In addition, we investigate the minimal amount of training data needed for reliable performance. We evaluate our Radio Sense system with multiple subjects' data collected over a two-week period and demonstrate that Radio Sense achieves reliable performance in terms of accuracy, latency, and battery lifetime.

In this paper, we propose to optimize energy consumption in heterogeneous wireless networks through joint packet size optimization. Specifically, we consider a two-hop data communication system composed of a body sensor network (BSN) and a WiFi network. Within the system, we formulate an energy consumption optimization problem with the constraints of both throughput and time delay. Mathematically, we convert this problem into a geometric programming (GP) problem, which is then numerically solved. The solutions can be used by both the BSN and the WiFi network to dynamically change their packets' payload sizes based on their current packet delivery ratios (PDRs). Since the PDRs are time-varying, we tabulate an offline payload size lookup table for online packet size selection using PDRs as indices. Finally, we collect PDRs from a deployed two-hop BSN-WiFi network and simulate the energy consumption. The performance evaluation results show that our solution achieves up to 70% energy savings compared with solutions that use fixed packet sizes.