論文

Blood pressure (BP) estimation methods using photoplethysmogram (PPG) based on deep learning models have been actively studied. These methods are also the basis of non-contact BP estimation methods using a camera or a Doppler radar. However, most previous studies are under data leakage, where subjects are not separated between training and test data. In this paper, we propose a method for BP estimation from PPG under the condition that subjects are separated between training and test data (inter-subject paradigm) using a subject-distinguishable large public dataset. Our BP estimation method estimates an 8-second BP waveform called arterial BP (ABP) from an 8-second PPG segment using U-Net. From the estimated ABP, systolic BP (SBP), diastolic BP (DBP), and mean BP (MBP) are calculated, which are the discrete single values for the 8-second ABP. In addition, we apply domain adversarial training, which facilitates the BP estimation model to extract subject-invariant features to improve the BP estimation accuracy under the inter-subject paradigm. Moreover, we use a considerably larger amount of training data than the previous studies and evaluate our model with smaller model sizes for better generalizability. Our experimental results showed that our method can estimate BP with moderate accuracy under the inter-subject paradigm, particularly MBP. The mean absolute errors for the estimated SBP, DBP, MBP, and ABP were 15.21, 7.12, 8.20, and 10.14 mmHg, respectively. The Pearson's correlation coefficients between the true and estimated values were 0.49, 0.39, 0.54, and 0.84, respectively.

Automatic modulation classification (AMC) is a promising technology to identify the modulation mode of the received signal in non-cooperative communication scenarios. Recently, benefitting from the outstanding classification performance of deep learning (DL), various deep neural networks (DNNs) have been introduced into AMC methods. Most current AMC methods are based on local framework (LocalAMC) where there is only one device, or centralized framework (CentAMC) where multiple local devices (LDs) upload their data to the only one central server (CS). LocalAMC may not get ideal results due to insufficient data and finite computational power. CentAMC carries a significant risk of privacy leakage and the final data for training model in CS is quite massive. In this paper, we propose an AMC method based on decentralized learning with residual network (ResNet). Simulation results show that ResNet-based decentralized AMC (DecentAMC) method achieves similar classification performance to CentAMC while improving training efficiency and protecting data privacy.

Accurate localization of devices in 5G cellular networks is of that utmost importance. This is because location information is a key component of a variety of new emerging applications. In particular, collocation (or co-location) refers to the idea of identifying devices that are located within a certain range from one another. In this paper, we propose a novel technique for inter-user distance estimation that uses low-resolution and high-resolution beam energy-based images as location fingerprints. Our approach uses the beam energy-based images generated by different users to estimate the distance between each pair of them. Nevertheless, we explore the idea of using a deep learning technique referred to as super resolution applied on low-resolution beam energy-based images to enhance their resolution, thus identify collocated users with an accuracy comparable to that of higher resolution ones. More specifically, throughout our experiments, we generate images of resolution 4times 4 and 8times 8 and use these for distance estimation between users. Afterwards, we apply super resolution on images with size 4times 4 to improve their resolution, and compare their results to the ones obtained with the original 8times 8 images. For an area roughly equal to 60times 30 mathrm{m}, our proposed approach reaches an average mean squared error equal to 0.13 m. We also demonstrate how our proposed approach outperforms the conventional ones that rely on user location detection to measure the inter-user distance.

Accurate channel state information (CSI) is necessary for frequency-division duplexing (FDD) massive multi-input multi-output (MIMO) systems. Existing deep learning-based CSI feedback methods, e.g., CSI sensing and recovery neural network (CsiNet), designed based on an autoencoder architecture, achieves higher feedback accuracy and reconstruction speed. However, this network needs to be retrained due to different communication scenarios and channel conditions, which is costly in practical deployment. To solve this problem, this paper proposes a deep learning-based modular adaptive multiple-rate (MAMR) compression CSI feedback framework. Extra padding modules are added at the base station to pad compressed CSI into different compression rates into the same dimensions, thereby realizing a general autoencoder performing variable-rate compression. Simulation results are given to confirm the effectiveness of the proposed method in terms of normalized mean square error.

Monitoring elderly people living alone is of the utmost importance given the amount of risk they are exposed to. Being aware of the activities of the elderly person in real time could help prevent/detect dangerous event that might occur such as falling. In this paper, we propose a method for activity detection using a 2D LIght Detection and Ranging (LIDAR) and deep learning. Unlike conventional work, where an activity refers to moving from one position to another, we use the term 'activity' to refer to a set of movements including walking, standing, falling and sitting. Not only does our approach detect these activities, but it also identifies a given person from his gait, and identifies unsteady gait (i.e., when he is about to fall or feeling dizzy). Throughout our experiments, we show that the proposed approach could reach an accuracy equal to 92.3% and 91.3% in activity and unsteady gait detection, respectively. It is also capable of identifying up to 3 people's gait with an accuracy equal to 92.4% using 10 seconds of walking data.

Human skeleton estimation using Frequency-Modulated Continuous Wave (FMCW) radar is a promising approach for privacy-preserving motion analysis. However, the existing methods struggle with sparse radar point cloud data, leading to inaccuracies in joint localization. To address this challenge, we propose a novel deep learning framework integrating convolutional neural networks (CNNs), multi-head transformers, and Bi-LSTM networks to enhance spatiotemporal feature representations. Our approach introduces a frame concatenation strategy that improves data quality before processing through the neural network pipeline. Experimental evaluations on the MARS dataset demonstrate that our model outperforms conventional methods by significantly reducing estimation errors, achieving a mean absolute error (MAE) of 1.77 cm and a root mean squared error (RMSE) of 2.92 cm while maintaining computational efficiency.

This paper proposes a heart rate estimation method using Variational Mode Decomposition (VMD) employing Multiple-Input Multiple-Output (MIMO) Frequency Modulated Continuous Wave (FMCW) radar. The proposed method first estimates the human position within the radar's coverage area, then reduces noise by focusing on the signal from these positions. The signal is decomposed into Intrinsic Mode Function (IMF) signals using VMD, and only the IMF signal related to the heartbeat is retained. Heart rate signal is reconstructed by weighting IMF signals based on their energy within the specific spatial area in which the human is located. The reconstructed signals are utilized for heart rate estimation through peak detection. The estimation is done over consecutive time windows. Until the fourth time window, among the estimated heart rates from the different cells, the selection is based on energy and periodicity. From the fifth time window onwards, the heart rate closest to the average of the previous four estimates is chosen. This approach considers the gradual transition of heart rate estimation over time to mitigate extraneous variations. Validation experiments involved 4 subjects being seated and stationary within the radar coverage area, with heart rate observed via MIMO FMCW radar. Results showed an average Mean Absolute Error (MAE) of 2.54 BPM, with an exclusion rate of 2.12%.

With the increasing demand for secure communication in 5G and beyond, authentication of wireless devices has become a crucial task for communication security. Radio frequency fingerprint identification (RFFI) leverages the hardware-specific features in radio frequency (RF) signals, known as radio frequency fingerprints (RFF), to achieve highprecision device identification. However, the dependence of RFFI on the physical characteristics of devices makes it vulnerable to physical spoofing attacks. This paper proposes an innovative physical spoofing attack framework that combines spoofed transmitter and legitimate transmitter models. It performs RFF modeling, RFF concealment (RFFC), and RFF spoofing (RFFS) sequentially to achieve precise spoofing of the original baseband signal. We validate the effectiveness of the proposed physical spoofing mechanism through simulations of seven types of transmitters using MATLAB Simulink. The performance is further evaluated on an RFFI model based on complexvalued convolutional neural networks (CVCNN). Experimental results show that neural networks (NN) significantly outperform the generalized memory polynomial (GMP) model in nonlinear data fitting and temporal relationship modeling. Consequently, NN-based physical spoofing methods exhibit superior attack effectiveness. Specifically, under the signal-to-noise ratio (SNR) of 15 dB, the NN-based physical spoofing method achieves a target attack success rate (TSR) as high as 98%, which is superior to adversarial attack methods. NN-based methods also enhanced performance in terms of stealthiness metrics.

Deploying deep neural network (DNN) models to the Internet of Things (IoT) for distributed inference (DI) is becoming an increasingly common requirement for AI applications. When IoT transmissions occur over lossy networks, data loss can significantly affect the accuracy of inference tasks. Therefore, we propose a plug-and-play module called Loss-Adapter, which can improve the accuracy of DI on lossy networks. To simulate network packet loss, we design a Gaussian distribution sampling dropout. This idea is utilized during the training of Loss-Adapter to acquire fault-tolerant inference capability on unreliable communication links. Moreover, we use the Tree-structured Parzen Estimator algorithm to optimize the parameters of the Gaussian dropout layer. To avoiding catastrophic forgetting, we implement freeze training. In addition, as a plug-and-play module, it is very easy to train and use. Our experiments show that by applying the Loss-Adapter, the inference accuracy of DNN models on lossy IoT is significantly improved.

Millimeter-Wave (mmWave) radar-based skeleton estimation has gained significant attention in the field of human motion analysis and sensing. It offers distinct advantages over RGB camera-based, depth camera-based, and inertial sensor-based approaches. It operates independently of lighting conditions. Additionally, it overcomes limitations of depth camera-based methods such as reflections and occlusions, and provides precise real-time tracking, enhancing its overall performance. However, existing human skeleton estimation methods utilizing point cloud data mostly rely on single-frame inputs or incorporate voxelization, which introduces drawbacks. This paper proposes a novel Convolutional Neural Network (CNN) and Bi-directional Long Short-Term Memory (BiLSTM)-based model for multi-frame point cloud data without voxelization. Furthermore, an Long Short-Term Memory (LSTM)-based neural network is introduced to assess point cloud reliability, enhancing robustness. Experimental results demonstrate improved accuracy and robustness in human skeleton estimation compared with the conventional single frame-based methods.

This study introduces a radar-based model for estimating blood pressure (BP) in a touch-free manner. The model accurately detects cardiac activity, allowing for contactless and continuous BP monitoring. Cardiac motions are considered crucial components for estimating blood pressure. Unfortunately, because these movements are extremely subtle and can be readily obscured by breathing and background noise, accurately detecting these motions with a radar system remains challenging. Our approach to radar-based blood pressure monitoring in this research primarily focuses on cardiac feature extraction. Initially, an integrated-spectrum waveform is implemented. The method is derived from the short-time Fourier transform (STFT) and has the ability to capture and maintain minute cardiac activities. The integrated spectrum concentrates on energy changes brought about by short and high-frequency vibrations, in contrast to the pulse-wave signals used in previous works. Hence, the interference caused by respiration, random noise, and heart contractile activity can be effectively eliminated. Additionally, we present two approaches for estimating cardiac characteristics. These methods involve the application of a hidden semi-Markov model (HSMM) and a U-net model to extract features from the integrated spectrum. In our approach, the accuracy of extracted cardiac features is highlighted by the notable decreases in the root mean square error (RMSE) for the estimated interbeat intervals (IBIs), systolic time, and diastolic time, which were reduced by 87.5%, 88.7%, and 73.1%. We reached a comparable prediction accuracy even while our subject was breathing normally, despite previous studies requiring the subject to hold their breath. The diastolic BP (DBP) error of our model is (Formula presented.) mmHg (mean absolute difference ± standard deviation), and the systolic BP (SBP) error is (Formula presented.) mmHg.

Traditional task offloading in Internet of Energy (IoE) edge environments struggles to accommodate heterogeneous user tasks and dynamic system conditions. Although Large Language Models (LLMs) offer potential solutions, their deployment is hindered by the dynamic resource constraints and stringent Quality of Service (QoS) requirements of IoE. Retrieval-augmented generation (RAG) enables LLM to overcome these limitations. In this paper, we propose an RAG-empowered adaptive task offloading framework. First, the framework transforms raw IoE-collected data into LLM-comprehensible textual inputs. Second, we design an intelligent preference-matching module that guides RAG to retrieve relevant knowledge and cases using key QoS features. Third, we build a composite knowledge vector database with both positive and negative experiences. We introduce a failure severity value to quantify the importance of counterexamples and enhance decision-making robustness. Finally, we design a hybrid knowledge prompt generation module. It utilizes a hierarchical filtering strategy and optimizes the ranking of retrieved knowledge for comprehensive prompts. Experimental results demonstrate that our proposed framework significantly outperforms baselines, increasing cumulative reward by 48.4% while reducing latency and energy consumption by 44.8% and 50.9%, respectively.

With the increasing number of new malicious software and network attacks, anomalous network traffic detection technology needs to continuously identify unknown attacks in order to adapt to complex network environments. In this paper, we propose a novel adversarial learning aided malicious network traffic detection (AMD) method based on encoder-decoder-encoder (EDE) framework. The proposed AMD method detects unknown anomalies in a single type of semi-supervised form and converts network traffic data into pictures for later processing. First, an anomalous network traffic detection model consisting of two encoders, a decoder and a discriminator is introduced. Second, by using the visibility of normal samples and the unrecognizability of unknown anomaly samples, the difference between the two samples is enlarged to facilitate effective sample detection. Finally, experiments show that the proposed model detects anomalies in targeted space and is robust to complex and noisy network environments. At the same time, compared to traditional machine learning methods, AMD's shows performance improvement of up to 40%.

We develop an accurate real-time scheduling framework for federated learning (FL) in wireless caching networks to guarantee the successful delivery of files at a low cost and with a short delay. The following persisting challenges motivated our work: i) Enforcing an excessive number of FL model update per communication round is infeasible due to the limited backhaul spectrum; ii) Naive scheduling policy accounting for FL model update renders service backlogs, thus leading to network parameter staleness and in-network caching utility (ICU) deterioration. Optimal scheduling in FL is challenging, as the mobile users' preferences for content, request patterns, and network traffic are dynamic and unknown. To tackle that challenge, we first formulate an instantaneous ICU optimization problem against the stale FL models. Afterward, based on the concept of age-of-update (AoU), we propose a federated learning with an unsatisfactory set selection (FedUSS) approach capable of executing the multiple-tasks of short-term predictions and making cache replacement decisions at low cost. Theoretical and numerical analyses manifest the effectiveness of our approach.

With the ubiquitous network applications and the continuous development of network attack technology, all social circles have paid close attention to the cyberspace security. Intrusion detection systems (IDS) plays a very important role in ensuring computer and communication systems security. Recently, deep learning has achieved a great success in the field of intrusion detection. However, the high computational complexity poses a major hurdle for the practical deployment of DL-based models. In this paper, we propose a novel approach based on a lightweight deep neural network (LNN) for IDS. We design a lightweight unit that can fully extract data features while reducing the computational burden by expanding and compressing feature maps. In addition, we use inverse residual structure and channel shuffle operation to achieve more effective training. Experiment results show that our proposed model for intrusion detection not only reduces the computational cost by 61.99% and the model size by 58.84%, but also achieves satisfactory accuracy and detection rate.

The rapid proliferation of the Internet of Things (IoT) has intensified the need for strong authentication mechanisms to ensure the integrity and reliability of connected devices. Recent advancements in Deep Learning (DL)-based Specific Emitter Identification (SEI) have demonstrated significant potential in leveraging unique Radio Frequency Fingerprints (RFF) for accurate device identification and authentication. However, the efficacy of these DL-based SEI methods is critically dependent on the availability of extensive labeled datasets, which are often scarce and expensive to obtain in practical applications. To address this limitation, Self-Supervised Learning (SSL) becomes a promising solution, capable of harnessing unlabeled data to learn effective representations. Furthermore, current surveys and reviews on SEI are generally summarized from a high-level perspective, lacking a detailed discussion of SEI methods under label-limited scenarios. This article comprehensively surveys SSL-based SEI, including its motivation, definition, paradigms, related work, challenges, and future direction combined with large models. To help readers quickly engage with this field, this paper also undertakes two specific efforts: collecting and organizing currently available open-source datasets with download links and comparing various SSL-based SEI methods with related codes.

Currently, the field of wireless sensing is moving towards non-contact and easy-to-deploy passive sensing technologies. Among these, sensing techniques based on WiFi channel state information (CSI) have emerged as highly promising due to their excellent performance and suitability for current sensing needs. However, despite the potential of WiFi CSI-based sensing technologies, there are still pressing issues that need to be addressed. Traditional WiFi CSI-based methods face challenges such as weak generalization ability and poor robustness, especially when the sensing target and environment change, which hinders the further development. To address this problem, this paper introduces the transfer learning method and combines it with the deep neural network to propose a novel cross-subject WiFi sensing method. Specifically, this method utilizes the source domain-target domain partitioning approach in transfer learning. In the source domain, deep neural networks are trained to obtain the source domain feature model, which is then transferred to the target domain. Fine-tuning is performed using techniques such as network freezing, aiming to meet the faster and more accurate sensing task demands of cross-subject. From the final experimental results, the novel cross-subject transfer learning method proposed in this paper achieved higher recognition accuracy and shorter training time. Moreover, the implementation of network freezing further enhanced the performance and efficiency of cross-subject, achieving up to 86% performance on the dataset used in this paper.

Existing transfer learning-based beam prediction approaches primarily rely on simple fine-tuning. When there is a significant difference in data distribution between the target domain and the source domain, simple fine-tuning limits the model’s performance in the target domain. To tackle this problem, we propose a transfer learning-based beam prediction method that combines fine-tuning with domain adaptation. We integrate a domain classifier into fine-tuning the pre-trained model. The model extracts domain-invariant features in adversarial training with domain classifier, which can enhance model performance in the target domain. Simulation results demonstrate that the proposed transfer learning-based beam prediction method achieves better achievable rate performance than the pure fine-tuning method in the target domain, and close to those when the training is done from scratch on the target domain.

Malware traffic classification (MTC) has become one of the important tasks to protect network and system security. However, it is a difficult task to obtain large-scale labeled malware traffic data in practice due to the wide variety and ever-changing malware. Existing MTC techniques require a large number of samples for training, the classification performance is poor when the training samples are few. Therefore, classification techniques based on few-shot learning (FSL) have attracted much attention in MTC. In this paper, we propose a few-shot MTC method based on knowledge transfer learning. Specifically, generic malware feature representation can be learned by pre-training the deep learning model on source tasks. Then, the parameters of the pre-trained model are used as initial parameters to fine-tune on the target task to adapt to feature distribution of the target task. The simulation results show that the few-shot MTC method based on knowledge transfer learning proposed in this paper has a 20-shot average accuracy of 95.23% when using GoogLeNet network as the backbone network, compared with CNN, VGG, and ResNet models, GoogLeNet has the best performance.

An Electrocardiograms (ECG) is a recording of the heart's electrical activity. It can help detect problems with one's heart rate or heart rhythm. Traditional methods to measure and collect ECG are not practical for daily use due to their invasive nature and inaccessibility. Thus, a less intrusive and more accessible method is needed. In this paper, we propose a method that uses attention-based Wasserstein Generative Adversarial Networks-Gradient Penalty (aWGAN-GP) to reconstruct ECG signals from ones collected using a Doppler sensor. GANs can capture pertinent features in the Doppler signals to enable such reconstruction. Our approach uses the integrated spectrum of the Doppler signal to identify R-peaks, which are then employed to train the aWGAN-GP to reconstruct the ECG signal. We evaluated this method on 19 healthy subjects and obtained a correlation coefficient of 0.88 between the reconstructed and actual ECG signals, outperforming the conventional CNN-based method which reached only 0.35. Our results demonstrate that it is possible to reconstruct an ECG signal from a heartbeat signal collected via a Doppler sensor using aWGAN-GP.