論文

Due to the implementation and performance limitations of centralized learning automatic modulation classification (CentAMC) method, this paper proposes a decentralized learning AMC (DecentAMC) method using model consolidation and lightweight design. Specifically, the model consolidation is realized by a central device (CD) for edge device (ED) model averaging (MA) and multiple EDs for ED model training. The lightweight is designed by separable convolutional neural network (S-CNN), in which the separable convolutional layer is utilized to replace the standard convolution layer and most of fully connected layers are cut off. Simulation results show that the proposed method substantially reduces the storage and computational capacity requirements of the EDs and communication overhead. The training efficiency also shows remarkable improvement. Compared with convolutional neural network (CNN), the space complexity (i.e., model parameters and output feature map) is decreased by about 94% and the time complexity (i.e., floating point operations) of S-CNN is decreased by about 96% while degrading the average correct classification probability by less than 1%. Compared with S-CNN-based CentAMC, without considering model weights uploading and downloading, the training efficiency of our proposed method is about N times of it, where N is the number of EDs. Considering the model weights uploading and downloading, the training efficiency of our proposed method can still be maintained at a high level (e.g., when the number of EDs is 12, the training efficency of the proposed AMC method is about 4 times that of S-CNN-based CentAMC in dataset D1 = 2 {FSK, 4FSK, 8FSK, BPSK, QPSK, 8PSK, 16QAM} and about 5 times that of S-CNN-based CentAMC in dataset D2 = 2 {FSK, 4FSK, 8FSK, BPSK, QPSK, 8PSK, PAM2, PAM4, PAM8, 16QAM}), while the communication overhead is reduced more than 35%.

The problem of unknown modulation signal recognition has been received intensely attentions in next-generational intelligent wireless communications. The deep learning (DL) has been widely used in unknown modulation signal recognition due to its excellent performance in solving classification problems and the DL-based automatic modulation classification (AMC) had been proposed. However, DL-based AMC method usually has high space complexity and computational complexity, which limits DL-based AMC to miniaturized devices with limited storage and computing capability. Therefore, a lightweight residual neural network (LResNet) for AMC is proposed in this paper. The simulation results show that the model parameters of LResNet is about 4.8% of the traditional CNN network, and about 14.9% of the ResNet and the classification performance of LResNet improves more than 3% compared with the traditional CNN network and decreases less than 1.5% compared to the ResNet.

Mobile communication systems are not only the core of the Information and Communication Technology (ICT) infrastructure but also that of our social infrastructure. The 5th generation mobile communication system (5G) has already started and is in use. 5G is expected for various use cases in industry and society. Thus, many companies and research institutes are now trying to improve the performance of 5G, that is, 5G Enhancement and the next generation of mobile communication systems (Beyond 5G (6G)). 6G is expected to meet various highly demanding requirements even compared with 5G, such as extremely high data rate, extremely large coverage, extremely low latency, extremely low energy, extremely high reliability, extreme massive connectivity, and so on. Artificial intelligence (AI) and machine learning (ML), AI/ML, will have more important roles than ever in 6G wireless communications with the above extreme high requirements for a diversity of applications, including new combinations of the requirements for new use cases. We can say that AI/ML will be essential for 6G wireless communications. This paper introduces some ML techniques and applications in 6G wireless communications, mainly focusing on the physical layer.

Device-to-Device (D2D) communication enabling direct data transmission between two mobile users has emerged as a vital component for 5G cellular networks to improve spectrum utilization and enhance system capacity. A critical issue for realizing these benefits in D2D-enabled networks is to properly allocate radio resources while coordinating the co-channel interference in a time-varying communication environment. In this paper, we propose a Stackelberg game (SG) guided multi-agent deep reinforcement learning (MADRL) approach, which allows D2D users to make smart power control and channel allocation decisions in a distributed manner. In particular, we define a crucial Stackelberg Q-value (ST-Q) to guide the learning direction, which can be calculated based on the equilibrium achieved in the Stackelberg game. With the guidance of the Stackelberg equilibrium, our approach converges faster with fewer iterations than the general MADRL method and thereby exhibits better performance in handling the network dynamics. After the initial training, each agent can infer timely D2D resource allocation strategies with distributed execution. Extensive simulations are conducted to validate the efficacy of our proposed scheme in developing timely resource allocation strategies. The results also show that our method outperforms the general MADRL based approach in terms of the average utility, channel capacity, and training time.

This paper presents a new three-dimensional (3D) trajectory design approach for a solar-powered, fixed-wing unmanned aerial vehicle (UAV) to harvest solar energy and collect data from multiple smart devices (SDs). The trajectory is optimized based on max-min fairness to balance the total amount of uploaded data and the fairness among the SDs. The key idea is that we develop non-trivial variable substitution and successive convex approximation (SCA) techniques to convexify data transmission, UAV energy consumption and mobility, and energy harvesting constraints under a persistent round-robin transmission schedule of the SDs. The resulting algorithm guarantees a locally optimal trajectory satisfying the Karush-Kuhn-Tucker (KKT) conditions. Another important aspect is that we further jointly optimize the transmission schedule along with the trajectory, and prove that absolute fairness in terms of uploaded data can be achieved among the SDs under the max-min fairness. The new algorithms apply to both line-of-sight (LoS)-dominant and probabilistic SD-UAV channels. Numerical results show that a 3D trajectory increases the uploaded data by 111%, compared to a two-dimensional (2D) trajectory. The proposed algorithms can balance the energy harvesting and data collection, and achieve fairness in both LoS-dominant and probabilistic SD-UAV channels.

We develop mobility-aware routing and caching strategies to solve the network cost minimization problem for dense small-cell networks. The challenge mainly stems from the insufficient backhaul capacity of small-cell networks and the limited storing capacity of small-cell base stations (SBSs). The optimization problem is NP-hard since both the mobility patterns of the mobilized users (MUs), as well as the MUs' preference for contents, are unknown. To tackle this problem, we start by dividing the entire geographical area into small sections, each of which containing one SBS and several MUs. Based on the concept of one-stop-shop (OSS), we propose a federated routing and popularity learning (FRPL) approach in which the SBSs cooperatively learn the routing and preference of their respective MUs, and make caching decision. Notably, FRPL enables the completion of the multi-tasks in one shot, thereby reducing the average processing time per global aggregation.1 Theoretical and numerical analyses show the effectiveness of our proposed approach.

By using Multiple-Input Multiple-Output (MIMO) Frequency Modulated Continuous Wave (FMCW) radar, a range-Angle map can be obtained that displays the signal strength in two dimensions with respect to distance and angular directions. In this process, a threshold algorithm such as Constant False Alarm Rate (CFAR) is used to detect positions where the signal strength exceeds a certain threshold value, enabling position estimation of targets. Furthermore, a method has been proposed to use the curve length of the trajectory of the I/Q signal to emphasize the signal of the person's position on the range angle map. However, the conventional CFAR may cause false alarms when noise and clutter signals are strong. In this paper, we propose a new method for human location estimation using the MIMO FMCW radar. Comparing the reflected signals from people and clutter, we see that I/Q signal variations associated with human motion have little correlation with clutter from stationary objects or noise, and I/Q signal variations associated with clutter often have high correlation with signals at other positions. We focus on the fact that there are many positions where the clutter component signal has a high correlation, compared to the signal caused by a person. Using the correlation map that expresses the correlation between the received I/Q signals on the range-Angle plane, we evaluated the characteristics of noise and clutter components from walls on the correlation map by standard deviation. As a result, we can remove strong noise and clutter components caused from near walls, thus improving the accuracy of object location estimation.

Accurate downlink channel state information (CSI) is one of the essential requirements for harnessing the potential advantages of frequency-division duplexing (FDD) massive multi-input multi-output (MIMO) systems. The current state-of-art in this vibrant research area include the use of deep learning to compress and feedback downlink CSI at the user equipments (UEs). These approaches focus mainly on achieving CSI feedback with high reconstruction performance and low complexity, but at the expense of inflexible compression rate (CR). High training overheads and limited storage capacity requirements are some of the challenges associated with the design of dynamic CR, which instantaneously adapt to propagation environment. This paper applies transfer learning (TL) to develop a multi-rate CSI compression and recovery neural network (TL-MRNet) with reduced training overheads. Simulation results are presented to validate the superiority of the proposed TL-MRNet over traditional methods in terms of normalized mean square error and cosine similarity.

Automatic modulation classification (AMC) is a critical algorithm for the identification of modulation types so as to enable more accurate demodulation in the non-cooperative scenarios. Deep learning (DL)-based AMC is believed as one of the most promising methods with great classification accuracy. However, the conventional CNN-based methods are lack of generality capabilities under time-varying signal-to-noise ratio (SNR) conditions, because these methods are merely trained on specific datasets and can only work under the corresponding condition. In this paper, a novel multi-task learning (MTL)-based generalized AMC method is proposed, and a more realistic scenario is considered, including white non-Gaussian noise and synchronization error. Its generalization capability stems from knowledge-sharing-based MTL in varying noise scenarios. In detail, multiple CNN models with the same structure are trained for multiple SNR conditions, but they share their knowledge (e.g. model weight) with each other. Thus, MTL can extract the general features from datasets in different noise scenarios. Simulation results show that our proposed architecture can achieve higher robustness and generalization than the conventional ones.

In this paper, we propose a deep learning-based technique for activity detection that uses wide-angle low-resolution infrared (IR) array sensors. Alongside with the main challenge which is how to further improve the performance of IR array sensor-based methods for activity detection, throughout this work, we address the following challenges: we employ a wide-angle infrared array sensor with peripheral vision in comparison to a standard IR array sensor. This makes activities at different positions have different patterns of temperature distribution, making it challenging to learn these different patterns. In addition, unlike previous works, our goal is to perform the activity detection using the least possible amount of information. While the conventional works use a time window equal to 10 seconds where a single event occurs, we aim to identify the activity using a time window of less than 1 second. Nevertheless, we aim to improve over the accuracy obtained in previous work by employing deep learning, while keeping the approach light for it to run on devices with low computational power. Therefore, we use a hybrid deep learning model well suited for the classification of distorted images because the neural network learns the features automatically. In our work, we use two IR sensors ( 32\times24 pixels), one placed on the wall and one on the ceiling. We collect data simultaneously from both the IR sensors and apply hybrid deep learning classification techniques to classify various activities including 'walking', 'standing', 'sitting', 'lying', 'falling', and the transition between the activities which is referred to as 'action change'. This is done in two steps. In the first step, we classify ceiling data and wall data separately as well as the combination of both (ceiling and wall) using a Convolutional Neural Network (CNN). In the second step, the output of the CNN is fed to a Long Short Term Memory (LSTM) with a window size equal to 5 frames to classify the sequence of activities. Through experiments, we show that the classification accuracy of the ceiling data, wall data, and combined data with the LSTM reach 0.96, 0.95, and 0.97, respectively.

The purpose of a network intrusion detection (NID) is to detect intrusions in the network, which plays a critical role in ensuring the security of the Internet of Things (IoT). Recently, deep learning (DL) has achieved a great success in the field of intrusion detection. However, the limited computing capabilities and storage of IoT devices hinder the actual deployment of DL-based high-complexity models. In this article, we propose a novel NID method for IoT based on the lightweight deep neural network (LNN). In the data preprocessing stage, to avoid high-dimensional raw traffic features leading to high model complexity, we use the principal component analysis (PCA) algorithm to achieve feature dimensionality reduction. Besides, our classifier uses the expansion and compression structure, the inverse residual structure, and the channel shuffle operation to achieve effective feature extraction with low computational cost. For the multiclassification task, we adopt the NID loss that acts as a better loss function to replace the standard cross-entropy loss for dealing with the problem of uneven distribution of samples. The results of experiments on two real-world NID data sets demonstrate that our method has excellent classification performance with low model complexity and small model size, and it is suitable for classifying the IoT traffic of normal and attack scenarios.

This article discusses the use of unmanned aerial vehicles (UAVs) to detect, warn, and suppress forest fires. The proposed system consists of: 1) an energy harvesting-assisted surveillance UAV (SUAV) that detects and reports fires and 2) multiple ground-based firefighting UAVs (FUAVs) that travel to fire spots to suppress the fire. First, the SUAV carries out the forest fire detection using the onboard color and infrared cameras. The SUAV then uses fire alarm signals to warn/inform the ground-based FUAVs to begin the firefighting mission. The goal is to inform as many FUAVs as possible as the required task (extinguishing the fire) can be completed more effectively and faster with more FUAVs. To do this, the SUAV with onboard processing capabilities makes use of the advantages of reinforcement learning. In particular, we propose a double Q-learning-based trajectory design problem that enables the energy-constrained SUAV to find the optimal sequence of FUAVs to visit (optimal flying trajectory) in order to maximize the number of informed FUAVs while considering the limited execution time, as the crucial issue is to control fire from the beginning to prevent uncontrolled spread of fire.

Federated learning (FL) has become an increasingly popular solution for intrusion detection to avoid data privacy leakage in Internet of Things (IoT) edge devices. However, most of the current FL-based intrusion detection methods still suffer from three limitations: (1) model parameters transmitted in each round may be used to recover private data which leads to security risks, (2) not independent and identically distributed (non-IID) private data seriously adversely affects the training of FL (especially distillation-based FL), and (3) high communication overhead caused by the large model size greatly hinders the actual deployment of the solution. To address these problems, this paper develops an intrusion detection method based on semi-supervised FL scheme via knowledge distillation. First, our proposed method leverages unlabeled data via distillation method to enhance the classifier performance. Second, we build a CNN-based model for extracting deep features of the traffic packets, and take this model as both the classifier network and discriminator network. Third, discriminator is designed to improve the quality of each client's predicted labels, to avoid the failure of distillation training caused by a large number of incorrect predictions under private non-IID data. Moreover, the combination of hard-label strategy and voting mechanism further reduces communication overhead. Experimental results on the real-world traffic dataset show that our proposed method can achieve better classification performance as well as lower communication overhead than state-of-the-art methods.

Accessing the accurate downlink channel state information (CSI) is essential to take full advantage of frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) systems due to its weak channel reciprocity. Meanwhile, great computational burdens will happen, which is accompanied by continuous CSI feedback. The existing compressive sensing (CS)-based and deep learning (DL)-based methods try to solve such problems, but do not achieve desired effect to get ideal CSI feedback or decrease the overhead. An adaptive deep neural network (DNN)-based CSI feedback method is proposed in this paper to address this. A classification block of the compression ratio is adopted and modified to apply to a more complex channel model named Clustered-Delay-Line (CDL), which helps decrease the computational overhead of the network. Besides, the reconstruction accuracy of the CSI feedback is further improved by proposing a new structure of the encoder. Quantization and dequantization modules are also applied to make the whole network more robust and effectively minimize the quantization distortion in the real communication scenario, respectively. The simulation results show that the proposed method performs better than the conventional ones on the CSI reconstruction accuracy in terms of normalized mean square error (NMSE), even though the quantization module is added.

The detection of Hate speech has been an increasingly active research topic. The results reported by the state-of-the-art systems to automatically detect hateful contents achieved almost perfect performance on common data sets. However, "hate speech"is a very subjective term, and people with different backgrounds have different levels of tolerance to what constitutes hate. In this paper, we show the limitations of having a single classifier handling the problem of hate speech detection. We then propose to build classifiers customized for different people, instead of a single classifier. The main obstacle towards achieving such a goal is the scarcity of data. Therefore, we use transfer learning to overcome this issue and use very limited amount of annotated data to build these customized classifiers. In a first stage, we build a classifier on a large data set which classifies tweets into 3 classes: hate, offensive, clean, and which we refer to as the general classifier. In the second stage, we asked 3 annotators with different backgrounds to re-annotate a small sub-set of tweets (600 tweets) from the original one. We refer to this newly created data set as "the customized data set."We then fine-tune the general classifier on the customized data set and build the customized classifier for each annotator. The accuracy of classification of corresponding customized data set got 0.08, 0.06 and 0.11 higher than the general classifier. The result shows that it is possible to start with a general classifier, and adjusted it to each individual despite the very limited amount of the training data for him/her.

In this paper, we propose an adaptive vehicle clustering algorithm based on fuzzy C-means algorithm, which aims at minimizing power consumption of the vehicles. Specifically, the proposed algorithm firstly dynamically allocates the computing resources of each virtual machine in the vehicle, according to the popularity of different virtualized network functions. The optimal clustering number to minimize the total energy consumption of vehicles is determined using the fuzzy C-means algorithm and the clustering head is selected based on vehicles moving direction, weighted mobility, and entropy. Simulation results are provided to confirm that the proposed algorithm can decrease the power consumption of vehicles while satisfying the vehicle delay requirement.

Deep learning-based radar signal recognition is considered one of the important technologies in the field of electronic countermeasure (ECM). However, existing deep learning-based methods require much time to design a specific neural network by experts for recognizing radar signals. It is difficult to employ these methods in real application scenarios. To solve this problem, we proposed an effective radar signal recognition method using neural architecture search (NAS) to automatically design convolutional neural networks (CNN). Experiments are given to validate the proposed method via comparing with both machine learning and deep learning-based methods. Experimental results show that the proposed method can achieve the optimal accuracy with low parameters and floating-point operations.

With the increasing popularity of Internet of things (IoT), the emergence of many IoT devices has led to security vulnerabilities. The classification of wireless signals is very important for secure communications. Most of existing signal classification tasks only focus on single signal classification task, while ignoring the relationship between radio frequency fingerprinting identification (RFFI) and automatic modulation classification (AMC). To solve the multi-task classification problem, this paper designs a multi-task learning convolutional neural networks (MTL-CNN). Real-radio datasets are generated by Signal Hound VSG60A and collected by Signal Hound BB60C to solve the lack of RFF samples with numerous modulation types. Experimental results confirm that the MTL-CNN method can work well by using the generated dataset. The MTL network designed in this paper improves the accuracy of RFFI by 1xs% relative to the single-task learning (STL) network. The keras code is released at https://github.comLiuK1288/1hw-000.

The proliferation of wireless sensor networks (WSNs) and their applications has attracted remarkable growth in unsolicited intrusions and security threats, which disrupt the normal operations of the WSNs. Deep learning (DL)-based network intrusion detection (NID) methods have been widely investigated and developed. However, the high computational complexity of DL seriously hinders the actual deployment of the DL-based model, particularly in the devices of WSNs that do not have powerful processing performance due to power limitation. In this letter, we propose a lightweight dynamic autoencoder network (LDAN) method for NID, which realizes efficient feature extraction through lightweight structure design. Experimental results show that our proposed model achieves high accuracy and robustness while greatly reducing computational cost and model size.

Specific emitter identification (SEI) is a promising technology to discriminate the individual emitter and enhance the security of various wireless communication systems. SEI is generally based on radio frequency fingerprinting (RFF) originated from the imperfection of emitter's hardware, which is difficult to forge. SEI is generally modeled as a classification task and deep learning (DL), which exhibits powerful classification capability, has been introduced into SEI for better identification performance. In the recent years, a novel DL model, named as complex-valued neural network (CVNN), has been applied into SEI methods for directly processing complex baseband signal and improving identification performance, but it also brings high model complexity and large model size, which is not conducive to the deployment of SEI, especially in Internet-of-things (IoT) scenarios. Thus, we propose an efficient SEI method based on CVNN and network compression, and the former is for performance improvement, while the latter is to reduce model complexity and size with ensuring satisfactory identification performance. Simulation results demonstrated that our proposed CVNN-based SEI method is superior to the existing DL-based methods in both identification performance and convergence speed, and the identification accuracy of CVNN can reach up to nearly 100% at high signal-to-noise ratios (SNRs). In addition, SlimCVNN just has 10% sim 30 % model sizes of the basic CVNN, and its computing complexity has different degrees of decline at different SNRs; there is almost no performance gap between SlimCVNN and CVNN. These results demonstrated the feasibility and potential of CVNN and model compression.