ABSTRACT. The lack of floor plans is one of the major obstacles to ubiquitous location-based services indoors. Dedicated mobile robots with high-precision sensors can scan and produce indoor maps, but the deployment remains low. Existing smartphone-based approaches usually adopt computer vision techniques to build the 3D point cloud, at the cost of extensive image collection efforts and the risk of privacy issues. In this paper, we propose BatMapper-Plus which constructs accurate and complete indoor floor plans by acoustic ranging and inertial sensing on smartphones. It employs acoustic signals to measure the distance to a nearby wall segment, and produces the accessible area by surrounding the building during walking. It also refines the constructed floor plan to eliminate scattered segments, and identifies connection areas including stairs and elevators among different floors. Extensive experiments in a teaching building and a residential building have shown our effectiveness compared with the state-of-the-art, without any privacy concerns and environmental limitations.

ABSTRACT. Along with intelligent device popularization, mobile crowd sensing (MCS) has garnered considerable interest as a novel way of sensing data acquisition. Active and continuous worker engagement in tasks is a critical concern for sustainability when selecting workers to accomplish tasks in continuous MCS. Previous worker selection approaches are unsuitable for continuous MCS to ensure a large enough workforce. This paper proposes a framework for dynamic mode-switching-based worker selection called DMWS. DMWS lets temporary low-quality workers at tasks improve their competitiveness through hybrid mode switching based on task completion quality to ensure long-term sustainability. Therefore, they have the opportunity to be selected by the MCS platform again. The ultimate objective is to maximize space coverage at the lowest possible cost by increasing worker participation. As evidenced by experimental results on two real-world data sets, DMWS outperforms other methods in terms of space coverage under budget constraints.

ABSTRACT. Distress signal identification has great significance in saving lives in coal mine rescue. In response to the rescue in the long-distance coal mine tunnel full of dust and dangerous gas, we propose a distress gesture sensing system utilizing LoRa technology, called MineSOS. Inspired by the Morse code ``SOS'' and binary code, we first present a set of distress gestures with an error-check design, only requiring one hand movement. For signal processing, we propose a novel metric to choose the optimal LoRa attribute due to the observation of the complementary relationship between amplitude and phase variation. Finally, a double-check mechanism is presented to recognize and verify the distress information. We conducted extensive experiments to evaluate the Mine SOS's performance, and the results show that it can achieve high accuracy for gesture recognition in a coal mine lab. MineSOS system also has the capability of long-range and though-obstacle sensing, which is believed to benefit emergency coal mine rescue.

ABSTRACT. Wireless sensing is increasingly used in the field of medical rehabilitation because of its advantages of low cost, non-contact and wide coverage. In the rehabilitation of patients, the recovery after upper limb injury is great significant. Nonstandard rehabilitation motions will cause secondary injury to the body. Therefore, how to achieve standardized rehabilitation motions at a low cost in the home environment has become an urgent problem to be solved. In order to settle it, a rehabilitation motion recognition method Wi-KF based on Wi-Fi signal is designed. First, we propose a data segmentation and counting Peak method, which can accurately segment a continuous motion into multiple single motions, and lays a foundation for feature extraction algorithm. The motion segmented by the Peak method is converted into Doppler feature image. Then Bag of Convolutional Feature (BoCF) algorithm is used to extract features and overcomes the difference of image size. Finally, the extracted features are input into Extreme Learning Machine (ELM) algorithm for classification. The Wi-KF method has extensively and fully verified in two real environments. The experimental results show that the average motion recognition rate of Wi-KF method is about 94.9%, hence the method has strong robustness. In sum, the method proposed in the paper provides a low-cost solution for standardizing the rehabilitation motions of patients.

ABSTRACT. Line crossing detection is to check whether people or objects go across a given barrier line, which is quite common and important in our daily life, such as the EAS checkpoint in a retail store or the finish line in track and field. Although existing solutions to line crossing detection have achieved great advancement, they do not function well when multiple objects or people cross the line at the same time. In this paper, we propose a new RFID-based solution called RF-Line to line crossing detection, especially for multi-object scenarios. The biggest challenge is that the RFID reader's coverage zone is invisible and irregular; we cannot roughly take the time when a tag is seen by the reader for the first time as the time when line crossing occurs. In RF-Line, we deploy two antennas opposite to each other and collect the RF phase profiles of two antennas at the same time. By a series of geometric transformations and mathematical derivations, we find that summing up the two phase profiles will get a new phase curve, in which the inflection point of the curve is the time of line crossing. In addition, we address the problem of turning back or long stay on the barrier line. We implement RF-Line with commodity RFID systems. Extensive experiments show that RF-Line can achieve accurate line crossing detection with a small error of 6.1 cm, with no need for any system calibration or complicated deployment.

ABSTRACT. Distracted driving is the act of driving while engaged in other activities, such as using a cell phone, texting, eating, or reading, which takes the driver' attention away from the road. Nowadays, the distracted driving detection models based on deep learning can extract critical information from video data to characterize the driving behavior process. But the distraction driving method based solely on appearance features cannot essentially eliminate the noise impact of the complex environment on the model, and the distracted driving recognition method based solely on skeletal information is unable to recognize the joint action of the human body and the objects. Therefore,the development of an accurate distracted driving detection model has become challenging. In this paper, we propose a distracted driving recognition model MFD-former based on the fusion of posture and appearance. First, a feature extraction module is proposed to extract skeleton data(i.e., posture) and appearance features(i.e., descriptors), which are merged by a graph neural network. Then, the two kinds of information are input into the MFD-former encoder module, and the self-attention mechanism quickly extracts the sparse data. Finally, the classification results of distracted driving are obtained by extracting the classification labels through the MLP Head. The MFD-former model outperforms existing models. It achieved 95.1% accuracy on the State Farm dataset and 90.24% accuracy on the self-built Train Drivers dataset.

ABSTRACT. In the current critical situation of novel coronavirus, the use of contactless gesture recognition method can reduce human contact and decrease the probability of virus transmission. In this context, ultrasound-based sensing has been widely concerned for its slow propagation speed, low sampling rate, and easy access to devices. However, limited by the complexity of ges-tural movements and insufficient training data, the accuracy and robustness of gesture recognition are low. To solve this problem, we propose Ultrason-icG, a system for highly robust gesture recognition on ultrasonic devices. The system first converts a single audio signal into a Doppler shift and subse-quently extracts the feature values using the Residual Neural Network (Res-Net34) and uses Bi-directional Long Short-Term Memory (Bi-LSTM) for ges-ture recognition. The method effectively improves the accuracy of gesture recognition by combining the information of feature dimension with time dimension. To overcome the challenge of insufficient dataset, we use data extension to expand the dataset. We have conducted extensive experiments and evaluations on UltrasonicG in a variety of real scenarios. The experi-mental results show that UltrasonicG can recognize 15 kinds of gestures with a recognition distance of 0.5 meters. And it has a high accuracy and robust-ness with a comprehensive recognition rate of 98.8% under different envi-ronments and influencing factors.

ABSTRACT. When wireless network technology is applied to industrial scenes, the open channel environment makes industrial equipment more vulnerable to attacks and threats from illegal nodes, such as eavesdropping, deception and identity information forgery. The complexity and variety of attack methods make the supervised machine learning algorithm insufficient to establish a low complexity, lightweight and high security authentication mechanism in industrial wireless sensor networks. Moreover, the wireless electromagnetic wave will be affected by additive noise and fading in the propagation process, making the wireless channel in a dynamic state. Based on this, we study a new authentication mechanism based on physical layer security for wireless sensor networks in dynamic industrial scenarios. Using more precise physical layer channel information, and building an authentication model around positive-unlabeled (PU) learning and bootstrap aggregating (bagging) strategy, we can accurately distinguish legal nodes and illegal nodes in the received channel information in the industrial scene where only the channel information of legal nodes is known. Finally, the effectiveness of the scheme is verified by using the public data set collected by the national institute of standards and technology (NIST) in a real industrial scene.

ABSTRACT. A large number of software vulnerabilities and weaknesses existing on the internet result in huge attacks. For spreading knowledge and strengthening defense of software security, researchers record software vulnerabilities, weaknesses and attack patterns through software databases including CVE, CWE, CAPEC and so on. However, software security databases are time-delayed and thus miss unobserved facts. Attackers can take advantage of this problem to execute an attack successfully. Therefore, reasoning task of predicting software security entity relation is critical to supplement software security data. In this paper, we construct a software security knowledge graph and propose a knowledge graph representation learning method combining Sentence-Bert and GAT. The method can implement link prediction and classification tasks for the purpose of knowledge graph completion. We finally design a large number of experiments to evaluate the effectiveness of our model in knowledge graph completion and knowledge graph classification. The experimental results demonstrate that the proposed method can effectively improve the effectiveness of prediction.

ABSTRACT. Traffic monitoring services show great potential for improving the traffic efficiency in vehicular networks. Drivers can obtain the latest information on their upcoming routes by sending location-based queries to the traffic monitoring server. However, it is inefficient to query all events on the route due to the timeliness of traffic events. Besides, sending location-based queries will expose drivers’ privacy. Existing research does not consider both issues under the traffic monitoring scenario. In this paper, we propose an efficient privacy-preserving scheme that ensures privacy is protected in two dimensions. The scheme first segments the whole route into multiple ones. Then drivers send endpoints in each segmented route to satisfy the timeliness requirement. Both location privacy and identity privacy are preserved in every segmented route, such that attackers cannot observe the real location as well as the identity via continuous queries. Specifically, we propose a location obfuscation mechanism based on geo-indistinguishability, and utilize it on every segmented route. We address an issue in geo-indistinguishability where the obfuscated location is unreasonable. Additionally, continuous queries may expose the driver’s identity. We thus define a type of attack called identity linking attack, and propose two possible solutions. We finally conduct experiments on the real dataset. Experimental results demonstrate the efficiency of our proposed scheme.

ABSTRACT. Along with the popularity of the Android operating system, 98% of mobile malware targets Android devices [1], which has become one of the primary source for privacy leakage. Detecting malicious network transmissions in these apps is challenging because the malware hides its behavior and masquerades as benign software to evade detection. In this work, we propose TraceDroid, a framework that can automatically trace abnormally sensitive network transmissions to detect the malware. By leveraging the static and dynamic analysis, the sensitive informations can be firstly inferred from the call graph, and then, the sensitive trans- missions can be detected by analyzing the network traffic per transfer and sensitive information with a machine learning classifier. We validate TraceDroid on 1444 malware and 700 benign applications. And our ex- periments show that TraceDroid can detect 3433 sensitive connections across 2144 apps with an accuracy of 94%.

ABSTRACT. With the rapid development of distributed machine learning and Internet of things, tons of distributed data created by devices are used for model training and what comes along is the concern of security and privacy. Traditional method of distributed machine learning asks devices to upload their raw data to a server, which may cause the privacy leakage. Federated learning mitigates this problem by sharing each devices’ model parameters only. However, it still has the risk of privacy leakage due to the weak security of model parameters. In this paper, we propose a scheme called privacy enhanced federated averaging (PEFedAvg) to enhance the security of model parameters. By the way, our scheme achieves the same training effect as Fedavg do at the cost of extra but acceptable time and has better performances on communication and computation cost compared with Paillier based federated averaging. The scheme uses the CKKS homomorphic encryption to encrypt the model parameters, provided by detailed scheme design and security analysis. To verify the effectiveness of the proposed algorithm, extensive experiments are conducted in two real-life datasets, and shows the advantages on aspects of communication and computation. Finally, we discuss the feasibility of deployment on IoT devices.

ABSTRACT. Wireless Sensor Networks (WSNs) are vulnerable to many security threats from compromised nodes. Trust management system is an effective method to detect the malicious behaviors launched by these compromised nodes in WSNs. In this paper, an effective Comprehensive Trust Evaluation Model (CTEM) for WSNs is proposed and two kinds of trusts are considered, the direct trust and the indirect trust. The direct trust is assessed by monitoring the data collection, the energy consumption and the data forwarding of node. More significantly, the entropy theory is introduced to measure the uncertainty of direct trust. The indirect trust is integrated to evaluate a comprehensive trust when the uncertainty of direct trust is high enough so as to improve the one-sidedness of direct trust. CTEM can not only reduce the computation overhead of node but also prolong the lifetime of network. Simulation results show that the proposed strategy can defend against internal attacks and have better performances compared with some typical trust evaluation mechanisms.

ABSTRACT. Vulnerability identification is crucial for cyber security in the software-related industry. Early identification methods require significant manual efforts in crafting features or annotating vulnerable code. Although the recent pre-trained models alleviate this issue, they overlook the multiple rich structural information contained in the code itself. In this paper, we propose a novel Multi-View Pre-Trained Model (MV-PTM) that encodes both sequential and multi-type structural information of the source code and uses contrastive learning to enhance code representations. The experiments conducted on two public datasets demonstrate the superiority of MV-PTM. In particular, MV-PTM improves GraphCodeBERT by 3.36\% on average in terms of F1 score.

ABSTRACT. Quantum secret sharing (QSS) splits and distributes a secret among N sharers based on quantum resources and quantum properties. While QSS protocols often use shared entangled states, it is also possible to define a notion of QSS based on single states, that is qubits when taking two-level quantum system into account. In such circumstance, local two-level unitary operations are always served to embed secret information and scramble the quantum states so that the secret cannot (easily) be achieved. Can any unitary operations guarantee the security of the protocols? We propose Bell state attack for such protocols, and give the simplified formula of minimum failure probability of such an attack. Thus the quantitative security analysis of different unitary-operation sets is presented under Bell state attack. Finally, some important conclusions about how to select unitary operation are obtained, which are crucial rules in designing and implementing quantum secret sharing protocols based on single qubits. The method and results may also contribute to analyze the security of other multi-party quantum cryptography schemes based on unitary operations.

ABSTRACT. In recent years, the rapidly increasing landscape of industrial control systems (ICS) devices has made the ICS geolocation more important. However, IP-based geolocation cannot provide high accuracy geographical locations for ICS devices. Commercial databases only provide coarse mappings between IP hosts and physical locations. Measured-based geolocation relies on the number of high-quality landmarks. In this paper, we present a novel framework called OSI-Geo for serving high-quality landmarks mining of ICS devices. The main idea is that there are many location-indicating clues in the open-source information exposed by ICS devices, which can be utilized to find their physical locations. The OSI-Geo automatically collects location-indicating clues to generate ICS landmarks at large-scale. We conduct real-world experiments for validating the effectiveness and performance of our method. The results show that OSI-Geo can accurately collect clues with over 99% recall and precision. Based on those clues, 36,872 stable landmarks, covering 162 countries and 5,596 cities, are obtained. Among them, there are 30,290 (82%) fine-grained landmarks accurate to street-level at least. The accuracy of IP geolocation has been improved significantly based on the ICS landmarks. Thus, OSI-Geo achieves effectively landmarks mining for ICS devices.

ABSTRACT. Skin melanoma is one of the most malignant tumors. In recent years, its incidence rate and mortality showed a high growth trend. Early detection and segmentation of skin lesions are vital in timely diagnosis and treatment. As the low contrast of lesion regions and high similarity in terms of appearance, skin lesion segmentation still remains a challenging work. Most of the segmentation methods use single-scale feature fusion, leading to the blur effect on the boundary. In this paper, we propose a new segmentation framework named Intensive Atrous Spatial Transformer Network (IASTrans-Net), which is based on a core module IASPP (Intensive Atrous Spatial Pyramid Pooling). The introduced IASPP can obtain valid features by using multi-scale feature fusion and channel attention. On the one hand, we employ atrous convolution with different dilation rates for multi-scale information extraction, ensuring that the effective information of each channel is obtained. On the other hand, channel attention is used to screen features, which can enable the network to effectively identify targets without increasing the training complexity. The experimental results show that the prposed IASTrans-Net has achieved good results in ISIC2017 and ISIC2018 datasets, surpassing most of the current mainstream methods.

ABSTRACT. Logs are widespread in large and complex software-intensive systems. Log-based anomaly detection is used for system diagnosis and troubleshooting. Existing methods extract log sequences as temporal log vectors, preserving the timing information between logs. However, they lack a reasoning mechanism, which prevents the model from mining the logical relationship between logs and loses the logical association between logs. In this paper, we propose LogLR, a log anomaly detection method based on logical reasoning. LogLR extracts the logical relationship between temporal log vectors and improves detection accuracy by combining Logical Tensor Network (LTN) with LSTM. In order to overcome the problem of ignoring the logical relationship between logs in existing statistical methods for data annotation. LogLR uses LTN to capture the logical relationship between log sequences and obtains weak labels to train an LSTM model through the weak label estimation method, which saves time costs. We evaluate LogLR on two widely used public datasets and the results demonstrate the effectiveness of LogLR.

ABSTRACT. In the era of big data, it has become our research goal to obtain valuable information from massive data simply and efficiently. Text summarization technology has been continuously developed to meet this demand. Recent work has also shown that transformer-based pretrained language models have achieved great success on various tasks in NLP. Aiming at the problem of Chinese news text summary generation and the application of Transformer structure to Chinese text data, this paper proposes a Chinese news headline generation model CNsum based on Transformer structure, and tests it on Chinese data sets such as NLPCC2017 and LCSTS. The experimental results show that CNsum achieves better ROUGE, BLEU and BERTScore scores than the baseline models, which verify the great performance of CNsum.

ABSTRACT. In this paper, a subcarrier index modulation aided code-shifted differential chaos shift keying (SIM-CS-DCSK) system based on orthogonal frequency division multiplexing (OFDM) is proposed. In the proposed system, the transmitted bits are divided into two parts, where one part is used for conventional CS-DCSK modulation and the other part, served as subcarrier mapping bits, is used for the subcarrier index modulation. Benefiting from index modulation, SIM-CS-DCSK can achieve higher data rate and lower peak-to-average power ratio (PAPR) compared to conventional multicarrier CS-DCSK system. Numerical simulations indicate that SIM-CS-DCSK has good BER performance over the time and frequency selective fading channels. Specifically, the proposed system outperforms the multicarrier spread-spectrum system (MCSS) by 0.5 dB to 2 dB. Real-field experiments in water pool and lake also confirm the superiority of the proposed system.

ABSTRACT. Supervised models for network intrusion detection usually rely on many training samples, but the annotation costs are very high. Unlabeled network traffic data is relatively easy to obtain. However, there are only a few methods to utilize these unlabeled data adequately. We propose a novel self-supervised few-shot network intrusion detection method to address the above problems. The method consists of two models: a) network traffic representation model and b) network intrusion detection model. First, the network traffic representation model uses unlabeled network traffic data through self-supervised learning to obtain network traffic representations, which will benefit the training of network intrusion detection model. Then, the shared layers of the network traffic representation model are transferred to the network intrusion detection model and frozen. Finally, a few training samples are used to fine-tune the network intrusion detection model, and we can obtain a model with good generalization. However, self-supervised learning of the network traffic representation model requires a method for generating labels from network traffic. Therefore, we propose a novel method to generate labels based on discrete features of network traffic. Experiments show that our proposed method has better performance than other network intrusion detection models with few-shot. On NSL-KDD, only 200 labeled samples are needed to achieve 95.2% accuracy.

ABSTRACT. Providing truly ubiquitous Internet connectivity requires development of massive satellite constellations, whose interference scenario changes dynamically in time and space, leading to interference analysis becomes more complicated and challenging. Conventional interference analysis is restricted to few number of satellites with deterministic locations which is not capable of evaluating the performance of a massive satellite network consisting of thousands of satellites. In this paper, we propose a Monte Carlo algorithm based on stochastic geometry for simulating massive satellite systems interference. In our algorithm, we first utilize stochastic geometry to model the satellites’ and earth stations’ location as a randomly distributed points process on a sphere, imitating the high spatiotemporal dynamic interference characteristics of the scenario. Then, Monte Carlo is used to randomly sample the interference scenario in an infinite continuous timeline. According to Monte Carlo simulation, we finally calculate the cumulative probability distribution of interference indicator, which can be used as the statistical result of long-term interference of massive satellite constellations.

ABSTRACT. In this paper, we investigate the channel resource allocation problem in device-to-device(D2D) based VANETs. According to the vehicle density, we first mark the urban transportation scenario into intensive and sparse areas, in which we categorize the communication links as "altruistic" and "ego" links respectively in the consequence of marking results and vehicle attributes. Secondly, the altruistic links are further grouped in terms of an improved spectral clustering algorithm proposed hereby. Moreover, channel resources are dedicated to ego links and different clusters of altruistic links in order to alleviate communication interference and achieve better performance. We formulate an optimization problem of power control for channel resource allocation to maximize the total channel throughput. Fortunately, after reshaping the original problem into a D.C (difference of two convex functions) problem, which can be solved by interior point method, the optimal power allocation method is yeilded. Intensive simulations are carried out across various configurations, and the results prove that our scheme has superior performance.

ABSTRACT. PIFuHD can generate high-resolution model in the process of human 3D reconstruction. However, PIFuHD will produce debris outside the human body when the image background is more complex. In order to solve this problem, this paper develops an adaptive BSCO algorithm for the background of human body and image of human body in 3D reconstruction system. The BSCO algorithm is divided into four steps in processing. First, BSCO algorithm uses Go-selfies to separate the background. Second, BSCO algorithm converges the RGB of all pixels of the character into a set. Third, BSCO algorithm finds the greatest difference from the set through HSV conversion. Fourth, BSCO algorithm weighs the set and then calculates the RGB score. To find the highest score of RGB, it is used as the RGB of the background after solid color optimization. The experimental results show that the proposed method improves the reconstruction effect of PIFuHD.

ABSTRACT. In the field of intelligent security, due to the negative effect of complicated influence factors, such as low video quality, different size of the target and occlusion, target detection is hard to be well-applicated in real life. Based on the above problems, this paper proposes a multi-layer feature cascade aggregation pyramid network (MCA-FPN) on the basis of Faster RCNN, which can fully combine the different level of semantic information to generate optimized feature maps and improve the expression ability of different size of features finally. Besides, to remove the negative effect from the imbalance distribution of samples, this paper discusses a new sample balanced weighted loss function SB-Loss to increase convergence speed and make the training process more efficiency. Finally, the method proposed by this paper has been experimented on the Pascal VOC dataset, with a maximum accuracy of 86.0%, which is highly competitive in this research area.

ABSTRACT. When dealing with the real track, the environment is often an unpredictable factor, so filtering is very important. We can use the filter to eliminate the influence of noise as much as possible. Kalman filter is one of them. In this work, we pro-posed a new Particle Swarm Optimization algorithm, called the Sheep Herding Optimization algori thm, which can obtain higher quality solutions with faster convergence speed and better stability. Besides, in order to improve the perfor-mance of Kalman filter, we apply the Sheep Herding Optimization algorithm to the filter. The improved Kalman filter can fuse and predict the track, and has higher computational performance and smaller error.

ABSTRACT. Low-Earth-Orbit (LEO) satellite networks can transmit data globally, but their time-varying nature makes data transmission delayed and inefficient. An important feature of ICN (Information-Centric Network) is in-network caching, which is helpful to increase transmission efficiency. To reduce the time-varying effect on data transmission and to improve transmission efficiency among satellites, this paper introduces ICN into LEO satellite networks and proposes a Caching Strategy based on Spreading Influence (CSSI) for satellite networks. Our strategy portrays satellite nodes and popular contents in terms of spreading influence, and caches popular contents with dynamic probability in satellite nodes with high global spreading influence in the period to obtain a better caching effect. Compared with LCE, LCD, Prob, and Betw strategies, the CSSI effectively improves the cache hit ratio and reduces the average content request delay.

ABSTRACT. With the development of the industrial Internet, IEEE Time-Sensitive Networking (TSN) has attracted more and more attentions due to its capability of providing deterministic network performance. Unlike most existing studies that only considered a single type of traffic, our work addresses the scheduling problem of mixed-criticality traffic in TSN. Time utility function (TUF) is a utility curve that measures the quality of service (QoS) of streams with respect to end-to-end delays. In this paper, we introduce a variety of TUFs for different streams in TSN according to specific timing requirements. To match the transmission protocol of TSN, we first categorize mixed-criticality traffic into periodic and aperiodic streams, and then design a novel scheduling scheme aiming to maximize the total TUF value of all streams. We compare our proposed scheme with two benchmark schemes, and evaluation results show that our proposed one outperforms the counterparts, especially under the worst-case network settings.

ABSTRACT. Federated Learning (FL) is a distributed machine learning approach which is suitable for edge computing environment. While in this environment, how to take full advantage of the computing resources on end devices and edge servers is still a difficult problem. Especially for the synchronous federated learning, computing resources among different participants will lead to extra time cost and cause resource waste. In this paper, we try to reduce the time cost and the computing resource waste by using model splitting and task scheduling. We first establish the mathematical model and find it can not be solved directly. Then we design our algorithm which we name as the Federated Learning Offloading Acceleration (FLOA) algorithm to obtain a sub-optimal solution. The FLOA algorithm first uses the Partition Points Selection method to reduce the size of the solution space, then proposes a task offloading method based on matching theory. Experiments and simulations show that compared to the other three calculation methods, the single iteration time is reduced by 47%, 28%, 14% under our algorithm in turn.

ABSTRACT. With the increasing variety and number of ocean applications, the underwater transmission of heterogeneous ocean data has become a hot spot in the research field of underwater acoustic sensor networks (UASNs). However, due to lack of flexibility in time slot allocation, the existing multiple access control (MAC) protocols for UASNs cannot be effectively applied to the transmission of heterogeneous ocean data and is prone to low time slot utilization and low throughput problems. Therefore, in order to solve the above problems in UASNs with heterogeneous ocean data, we propose a time slot variable MAC protocol (TSV-MAC) based on deep reinforcement learning. In TSV-MAC, the long short term memory (LSTM) deep learning model is constructed and is trained by considering the usage efficiency of time slots and the data collection condition of underwater nodes. Then, the trained LSTM model is applied to predict the generation and transmission of data from each underwater node and a Q-learning model is adopted to allocate a suitable number of time slots to underwater nodes in advance. The TSV-MAC protocol periodically updates the time slot allocation table, to enable UASNs to adapt the different data packets that are dynamically generated. Finally, the effectiveness of the protocol is verified by extensive simulation results.

ABSTRACT. Graph Convolutional Networks (GCNs) have gained much attention and have achieved excellent performance in many graph-based collaborative filtering (CF) tasks in recent years. Its success relies on a fundamental assumption that the original graph structure is reliable and consistent with the properties of GNNs. However, most original graphs can seriously impair model performance due to noise and data sparsity problems. In addition, for large user-item graphs, the explicit message passing in traditional GCNs slows down the convergence speed during training and weakens the training efficiency of the model. Based on this, we propose Constrained Graph Convolution Networks Based on Graph Enhancement for Collaborative Filtering (EL-GCCF). The graph initialization learning module integrates the structural and feature information in the graph by generating two graph structures. It enhances the original graph and effectively mitigates the noise problem. Second, the multi-task constrained graph convolution skips explicit message passing. It effectively mitigates the over-smoothing problem in training and improves the training efficiency of the model by using an auxiliary sampling strategy. Experimental results on two real datasets show that the EL-GCCF model outperforms many mainstream models and has higher training efficiency.

ABSTRACT. Real-time detection and identification of dangerous driving behaviors is an effective measure to reduce traffic accidents. Due to the high network delay, limited communication bandwidth, and weak computing power, lightweight detection models that can run on edge devices have been widely investigated and attracted considerable attention. In recent years, the Graph Convolutional Networks (GCN), which model the human skeleton as a spatiotemporal graph, have advantages over image classification and target detection neural networks in human behavior detection tasks. However, there are disadvantages such as the single way of extracting information, high model complexity, and inability to integrate environmental information. Therefore, we propose a lightweight dangerous driving behavior detection model based on GCN. First, two local information extraction modules are designed to extract skeleton information features. Meanwhile, we propose a multi-information fusion behavior recognition model of "people + objects" by capturing the motion information of related object. Finally, the method based on Singular Value Decomposition (SVD) rank reduction is used to compress the model to improve the speed of recognizing an action sample under sufficient detection accuracy. The proposed model respectively achieves 96% and 87.4% accuracy on the x-view benchmark of NTU-RGBD dataset and the homemade Locomotive Driver Dataset, which attains the state-of-the-art performance.

ABSTRACT. Traffic prediction problem plays a crucial role in the research of intelligent transportation systems. Traffic flow is an important indicator to measure the traffic status. Traffic flow prediction can not only provide a scientific basis for traffic managers but also support other road services. This work proposes a spatial-temporal convolutional neural network model with improved graph representation (IGR-TCN) for predicting urban traffic flow, which solves the limitations of traditional methods considering only a single road section or a single detector. IGR-TCN reduces the computational complexity by using convolutional structure, the temporal convolution layer uses dilated convolution, and causal convolution to optimize the long-term prediction capability. The graph representation proposed in this work improves the existing spatial-temporal correlation model and increases the spatial correlation trend of the data. The IGR-TCN fits better than traditional recurrent neural networks, traditional graph convolution models, and graph spatial-temporal models. It can be more effective for spatial-temporal information prediction.

ABSTRACT. Estimating the traffic incident duration is of great importance to traffic control, traffic navigation, and transportation safety. However, the complex road network topology and dynamic traffic conditions make it challenging. In this paper, we propose a context-aware spatio-temporal graph convolution framework, named TimeBird, to estimate the duration time of traffic incidents. Specifically, we build the dynamic weighted adjacent matrix and traffic incident risk similarity matrix to learn the hidden spatial context correlations based on graph convolution network. Then we employ the historical traffic speed of road segments to learn the temporal dependency. Lastly, we design a context-aware attention mechanism to adaptively learn the heterogeneous traffic features for incident duration prediction. Extensive experiments on two large-scale real-world datasets from DiDi ride-hailing platform demonstrate the effectiveness of TimeBird.

ABSTRACT. With the increasing number of insider threat incidents, insider threat has become one of the most serious network security problems. Currently, the large volume of redundant user behavior data affects the efficiency and accuracy of the detection methods. Meanwhile, existing insider threat detection methods cannot fully learn the important features of user data, resulting in high false alarm rate. To solve these problems, we propose a novel insider threat method based on variational auto-encoder (VAE) and back propagation neural network (BPNN) in the paper. Initially, we use the generative model VAE to construct the normal user behavior model, and obtain the effective feature representation of user behavior. Then, we use the BPNN algorithm to detect abnormal user behavior from a large number of user activity logs. Finally, we conduct experiments to verify the detection performance of the proposed method. Experimental results indicate that the proposed detection method can achieve high accuracy and precision.

ABSTRACT. Communication load in heterogeneous edge networks is becoming heavier because of excessive computation and delay caused by straggler dropout, leading to high electricity cost and serious greenhouse gas emissions. To create a green edge environment, we focus on mitigating computation and straggler dropout to improve the communication efficiency during the distributed training. Therefore, we propose a novel scheme named Dynamic Grouping and Heterogeneity-aware Gradient Coding (DGHGC) to speed up average iteration time. The average iteration time is used as a metric reflecting the effect of mitigating computation and straggler dropout. Specifically, DGHGC firstly uses the static grouping to evenly distribute stragglers in each group. After the static grouping, considering the nonuniform distribution of nodes due to straggler dropout during the training process, a dynamic grouping depending on dropout frequency of stragglers is employed. The dynamic grouping tolerates more stragglers by examining the dropout threshold to improve the rationality of the static grouping for stragglers. In addition, DGHGC applies a heterogeneity-aware gradient coding to allocate reasonable data to stragglers based on their computing capacity and encode gradients to prevent stragglers from dropping out. Numerical results demonstrate that the average iteration time of DGHGC can be reduced largely compared to the state-of-art benchmark schemes.

ABSTRACT. In the field of code clone detection, there are token-based similarity and abstract syntax tree-based detection methods. The former consumes less resources and is faster to detect, while the latter consumes more space and is less efficient. And there are few tools that scale to large-scale databases. To address the challenges, an approach is proposed that can detect code clones using the similarity of tokens and architecture of abstract syntax trees. Architecture of the syntax trees preserves the precision of detecting clone pairs, at the same time, the method also preserves the speed of matching code similarity. In the approach, it first parses the tokens of the code fragments and the features of the syntax trees, eliminates the unqualified parts of them based on the architecture when matching the candidates quickly, and then detects the similarity in detail. Finally, the results are output according to the input threshold range. The experiments confirms that the method substantially improves the precision of code clone detection while keeping the recall rate unabated. Furthermore, it reduces the detection time proportionally to data scale, improving the detection efficiency.

ABSTRACT. Ring signcryption schemes have been proposed on blockchain, but compared with ring signcryption using public key infrastructure (PKI), certificateless can simplify the certificate management process. At the same time, the length of the ciphertext increases with the size of the ring is also a pressing challenge to be solved. In this paper, we propose a new certificateless ring signcryption scheme with constant ciphertext length and prove the security under the random oracle model. Compared with other schemes, the computational overhead of this scheme is smaller and more efficient.

ABSTRACT. Recently, automated machine learning (AutoML) and neural architecture search (NAS), regarded as promising techniques to design deep learning (DL) models automatically, have received increasing attention from both industry and academia. NAS will generate a large number of candidate models, which typically consist of numerous common substructures, providing a vast opportunity for cross-model optimization (e.g., operator batching) to improve training efficiency. However, most of the existing AutoML frameworks do not make use of operator batching and we also lack an efficient batching strategy. In this work, we propose a heuristic scheme named DPBat to guide the operator batching among multiple models in NAS. For most models, the operator batching of DPBat can be finished in just a few seconds, which is negligible compared to the subsequent training. We adopt Microsoft’s open source AutoML framework NNI to implement DPBat to real NAS scenarios. Extensive experiments show that DPBat is highly effective in improving training efficiency and reducing the overhead of operator batching, with a throughput 3.7× higher than the standard practice of running each job without batching.

ABSTRACT. Intelligent transportation is one of the most important applications for improving urban traffic pressure. However, intersections are an important element of urban road network, which makes the complex traffic data face the challenge of security and efficiency in the process of transmission. In this paper, we propose a smart contract-based intelligent traffic adaptive signal control scheme to optimize the traffic efficiency problem at intersections. In the scheme, we use consortium blockchain and smart contracts to ensure secure transmission of traffic data and trusted access permission verification for Internet of Things (IoT) devices. Then, we introduce edge computing into the intelligent transportation system, which can process massive traffic data in real time. In addition, we propose an improved Webster algorithm, aiming at optimizing the dynamic allocation of signal times, so as to reduce the congestion at intersections. The security analysis and evaluation experiments demonstrate that the scheme is feasible and valid, and it can facilitate the adaptive control of traffic signal lights.

ABSTRACT. The simulation of a Mobile Ad Hoc Network (MANET), before deployment or during the system running, provides a priori design validation and insightful observation of the real system. But existing simulation tools mainly enable these by means of centralized instead of distributed deployment, which in some sense, cannot truly replicate the real system settings. In this paper, we present a DIstributively deployable Simulation tool for MANet (DISMAN), to accurately simulate MANET in a fully-distributed fashion thus allowing the emulation to scale with the network nodes without sacrificing accuracy. DISMAN is a fully functional tool that can be integrated with Kubernetes, support link layer (e.g., bandwidth limitation, delay, packet loss) and the multi-path as well as multi-hop transmission simulations. DISMAN is based on a four-layer architecture design, where on the top is a graphical user interface (GUI) layer for presentation and interaction. We further evaluate DISMAN with micro- and macro-benchmarks and show that DISMAN is easy to use and how it can assist MANET design.

ABSTRACT. Unmanned aerial vehicles (UAVs) are adopted as promising platforms to provide aerial wireless communications and networks. However, due to the line-of-sight (LoS) dominant air-ground channels, UAVs cause stronger interference to the terrestrial network devices. In this work, we study a novel interference mitigation method via collaborative beamforming under a UAV-enabled data collection scenario. Specifically, we form a UAV-enabled virtual antenna array (UVAA) to transmit the collected data to the terrestrial base stations (BSs), and formulate an interference mitigation multi-objective optimization problem (IMMOP) to simultaneously enhance the data transmission efficiency, reduce the interference affection and increase the network lifetime. Due to the complexity and NP-hardness of IMMOP, a chaotic multi-objective multi-verse optimizer (CMOMVO) is proposed for solving the problem. Simulation results show that the CMOMVO can effectively solve the IMMOP and has better performance than some benchmark algorithms.

ABSTRACT. With high mobility and adaptability, the Unmanned Aerial Vehicle (UAV) has provided a promising solution for data collection in Wireless Sensor Networks (WSNs). However, few existing works considered that data overwritten would occur if the UAV can not collect data from sensors in time, which will cause data loss in WSNs. Moreover, the importance of data stored in different sensors may vary significantly according to the application scenario. In this paper, we formulate a novelty Loss Minimization Problem (LMP) in a UAV-enabled WSN. The objective is to minimize the volume of weighted data loss in the WSN by jointly considering the UAV hovering locations and hovering durations, subject to the limited energy capacity. We first devise a novelty one-to-many data collection scheme that enables the UAV to collect data from multiple sensors simultaneously. Then we discrete the infinite hovering locations of the UAV into finite to reduce computational complexity. We instead propose efficient heuristic and approximation algorithms for the optimization problem. Finally, we evaluate the performance of the proposed algorithms through extensive experimental simulations. Simulation results demonstrated that the proposed algorithms are promising.

ABSTRACT. In the unmanned aerial vehicle (UAV) edge computing scenario, how to reasonably offload the tasks generated by the user equipment is a research hotspot. Furthermore, assigning a shared global reward to each agent in a multi-agent scenario creates a credit assignment problem. In this work, under the constraints of processing delay and energy consumption, we model the interaction process between the agent and the scene as POMDP, and propose a multi-agent task offloading algorithm CMADDPG based on counterfactual reasoning to efficiently offload tasks . CMADDPG constructs a dynamically adjusted counterfactual action pool to evaluate the influence of each agent on the global reward to address potential credit assignments in scenarios. In addition, CMADDPG changes the output form of the state-action value to have the advantage of learning the actions of each dimension in a targeted manner. Experimental results show that CMADDPG can significantly reduce the processing delay of user tasks, and can effectively solve the problem of credit assignment in deterministic algorithms and avoid the occurrence of lazy agent.

ABSTRACT. In this paper, we study sensory data collection of IoT devices in a wireless sensor network, employing a given number of UAVs. We observe that most existing studies ignored the different importance of data stored in IoT devices and simply minimized the longest data collection latency of IoT devices. Then, it is possible that the data collection latency of a IoT device may be long, the data collection priority of the IoT device is high and its data should be collected faster than IoT devices with low priorities. Considering the data collection priority of each IoT device, we formulate a novel weighted data collection latency minimization problem to collect data stored in IoT devices using the UAVs, by finding a closed tour for each UAV such that the maximum weighted data collection latency of IoT devices is minimized, where the data collection latency of IoT devices is composed of the hovering time of UAVs for data collection and the flying time of UAVs from one hovering location to another hovering location. To deal with the above NP-hard problem, we first propose a simplified data collection latency minimization problem which does not take account of the data collection priorities of IoT devices and base stations. Then, we devise an approximation algorithm for the simplified problem and further employ it to deal with the weighted data collection latency minimization problem. Finally, we evaluate the performance of the proposed algorithms through experimental simulations. Experimental results show that the proposed algorithms are very promising.

ABSTRACT. With the wide application of Internet of Things (IoT) systems in smart ocean, many unmanned surface vehicles (USVs) have been deployed jointly with unmanned aerial vehicles (UA Vs) to monitor the maritime environment. However, conventional means of maritime communications fail to provide high rate services due to the complex mar- itime channel conditions and large transmission distance, which will affect the environmental monitoring performance. In this paper, we propose a USV-UA V collaborative patrol scheme for maritime environment monitoring networks. Considering the characteristic of energy concentration in beamforming, we investigate the joint beamforming and location deployment optimization problem (BLDO) for UA V relay. Specifically, we decompose the BLDO problem into two subproblems. In the first subproblem, the location deployment of UAV and beam gain allocation are optimized via an iterative algorithm based on the approximated beam patterns. The algorithm can effectively reduce the computational complexity of the grid-search method. In the second sub-problem, beamforming optimization is conducted with high-dimensional constant-modulus (CM) constraint. A micro-particle swarm optimization based algorithm with boundary relaxation (BR−µP SO) is proposed to obtain an optimal solution. Finally, the simulation results demonstrate that the proposed algorithms can improve the performance in terms of the achievable sum rate and the beam gain.

ABSTRACT. With the rapid development of information technology, low-cost unmanned aerial vehicles (UAVs) appear. With advanced sensing and actuating technologies, they are being increasingly applied to a variety of scenarios, such as military rescue, communication assistance and civilian monitoring, to provide various kinds of services. However, considering their limited computing resource and restricted battery capability, the computation-intensive tasks or data-intensive tasks will face tough challenges. With the aid of Mobile Edge Computing (MEC), moving computation-intensive tasks from resource-constrained UAVs to edge cloud servers can significantly save energy and finally achieve impressive performance. As multiple service providers are deployed in the UAVs coverage area, this paper proposes an evolutionary game based algorithm to solve the computation offloading problem for UAVs. By replicator dynamics, UAVs select the suitable service provider to offload the computation tasks via achieving a tradeoff between time delay, energy consumption and monetary cost when network externality exists. Simulation results show that the proposed algorithm can rapidly converge to evolutionary equilibrium and achieve desirable performance.

ABSTRACT. As Unmanned Aerial Vehicles (UAVs) are widely utilized for wireless communication services, this has promoted the emergence of promising UAV-assisted vehicle networking. However, due to the ever-increasing data traffic and diversified wireless service requirements of vehicles, there are also privacy issues caused by fraud, which challenges the effective allocation of limited security bandwidth for secure communications. In this article, in order to solve these two problems, we firstly propose a secure bandwidth allocation scheme based on the game theory of smart contracts on the Internet of Vehicles assisted by UAV. Secondly, the proposed blockchain-based system introduces an emerging consensus mechanism that can significantly reduce the delay in exchanging information and protect data privacy. Furthermore, to allocate the limited safe band-width, based on the real-time feedback of each UAV, we designed an optimal decision search algorithm based on gradient descent to achieve Stackelberg equilibrium. Finally, the simulation results show the superiority of improving the utility's security bandwidth allocation scheme.

ABSTRACT. The present research on label-consistent invisible backdoor attacks mainly faces the problem of needing a high poisoning rate to achieve a high attack success rate. To address this problem, this paper proposes a low-poisoning rate invisible backdoor attack based on important neurons (INIB) by enhancing the connection between triggers and target labels with the help of the neural gradient ranking algorithm. The method first identifies the neurons with the most significant influence on the target label with the help of the neural gradient ranking algorithm, secondly establishes a strong link between the important neurons and the trigger using the gradient descent algorithm, and then generates a trigger based on the established strong link by minimizing the difference between the current activation value and the expected activation value of the important neurons, thus causing the important neurons to be strongly activated when images have the trigger, which in turn causes the model to misidentify them as the target label. Finally, detailed experimental results show that INIB is able to achieve a very high attack success rate with a very low poisoning rate while ensuring a high model testing accuracy. Specifically, INIB achieves a 98.7% backdoor attack success rate and a model accuracy of 97.5% with the poisoning rate of only 1.64% on the MNIST dataset.

ABSTRACT. Crowdsensing systems collect multidimensional and heterogeneous data using sensing devices of individuals(called workers). However, there are often conflicts between data from multiple sources. Truth discovery try to resolve the conflicts by evaluate the trustworthiness of each source to find the groud truth. However, recent works have shown that truth discovery are vulnerable to data poisoning attacks. By submitting carefully crafted malicious data, attackers can launch multi-round iterations of hiding and attacking to deceive trustworthiness from truth discovery. This can result in long-term, insidious, and damaging data poisoning attacks. To this end, it is necessary to study how to improve the robustness of truth discovery. In this paper, we study high robust truth discovery methods against multi-round data poisoning attacks in crowdsensing systems, and attempt to detect potentially malicious workers. First, we implement one of the most widely used truth discovery frameworks, and verify its vulnerability to multi-round data poisoning attacks. Second, we propose HRTDMD, a high robust truth discovery method with malicious worker detection. Finally, we conduct experiments on real dataset and verify the effectiveness of the proposed method.

ABSTRACT. In smart home Internet of Things (IoT) systems, interactions between devices are driven in two ways: automation applications (apps), and the physical channels (e.g., temperature, smoke). Meanwhile, device interactions can be maliciously exploited to launch IoT attacks. However, limited efforts explore whether it is feasible to discover potential exploitable device interactions from IoT deployments. This paper proposes a novel framework to detect interactions among devices (D-interact) from eavesdropped network traffic and device function descriptions, and discover all potential exploitable device interactions (i.t., attack paths). First, we use the packet-level patterns to fingerprint IoT device events and then identify all IoT device events from the eavesdropped traffic. Furthermore, we mine temporal and conditional dependencies of IoT events to infer device interactions introduced by IoT apps. Besides, to identify interactions between devices and physical channels, we use the natural language processing (NLP) technique to analyze device function descriptions. Based on the obtained device interactions, D-interact builds a device interaction graph to discover attack paths. To demonstrate the feasibility of our approach, we implement D-interact in a real-world smart home including 24 devices and 29 apps. The experiment results show that 38 device interactions are identified and 26 device interaction paths could be potentially exploited to impact the safety of the IoT environment.

ABSTRACT. False data injection attacks modify the meter measurements to mislead the control center into estimating inaccurate system states and thus affect the reliable operation of smart grids. In this paper, we propose a deep reinforcement learning based vulnerability analysis scheme for smart grids that enables the control center to construct an attack vector from the attacker’s view to identify the vulnerable meters. The control center chooses the attack vector based on power system states, meter measurements, the previous number of analyzed meters and injected errors without knowing the power system topology. This scheme designs an actor-critic architecture that uses an actor network to output the policy probability distribution to handle the continuous and high-dimensional vulnerability analysis policy and a critic network to guide the weights update of the actor network. We also perform simulations to verify the efficacy of the proposed scheme in terms of the vulnerability detection rate, the number of analyzed meters and the utility.

ABSTRACT. Lightweight cryptography is a subfield of cryptography, which is widely used in embedded systems, RFID, sensor networks, and so on. However, the leakage information during the operation of these IoT devices can be exploited by adversaries and subjected to side-channel attacks. Simultaneously, only a small number of previous works show these attacks. In this work, we perform the soft analytical side-channel attack (SASCA) on the encryption of Ascon. Since we construct a unique factor graph for Ascon, we can also use it to attack the masked implementations. The point of attack is the permutation function, one of Ascon's most basic components. Our attack mainly consists of three steps. At the first, we run a side-channel template matching on the initialization phase. Then, we build a factor graph describing the intermediate computations in permutation, including the observed leakage for the intermediate variables. Third, we run a Belief Propagation (BP) algorithm that takes full advantage of these leakages. Through simulations, we show that the entire key can be successfully recovered by only using the leakage information of a few traces, and it also offers low time and memory complexity.

ABSTRACT. Deep learning plays a vital role in malware detection. The Malconv is a well-known deep learning-based open source malware detection framework to prevent damage in advance. Researchers propose adversarial example generation strategies to evade the Malconv by modifying the PE headers or the end of malware. However, these strategies that focus on non-executable portions can be easily pre-processed before classification. Therefore, we propose a new instructions replacement strategy to overcome these flaws. This paper reviews the research progress of adversarial example generation strategies for the Malconv in recent years, analyzes the reason why the Malconv can be evaded by adversarial examples and identifies two layers of the Malconv that can be attacked, and propose the gradient-based instructions replacement strategy named EFGSM that is an enhanced Fast Gradient Sign Method (FGSM) and sheds light on future work in adversarial example defense strategies for the Malconv. The paper assesses the performance of our EFGSM and existing adversarial example generation strategies upon 200 malware. The results of the evaluation show that our strategy improves the success rate from 68% to 81.5% and takes less time to generate malware examples. The paper also assesses the evasion performance of adversarial examples in three antiviruses. The results depict that our strategy generates a higher percentage of samples that can evade the antivirus than existing strategies.

ABSTRACT. A private set intersection (PSI) protocol is a protocol to get the intersection of two sets, each of which belongs to one party, without disclosing extra information of each party's set to the other party. In this paper, we propose a novel semi-honest PSI protocol without using any encryption primitive in a semi-honest security model. Specifically, we first slice all elements in the set and inject dummy slices. Then we utilize the greatest common divisor (GCD) to find the common parts of the two parties' slice set product, which is the product of all slices. Finally, we filter the elements by utilizing the GCD to find out the intersection. Different from most previous PSI protocols, we get the intersection by calculation rather than comparison.

Our protocol has many advantages over other state-of-the-art PSI protocols, such as robust security against quantum attacks, low communication cost, high computation efficiency when the bandwidth is low, etc. Through extensive experiments, we find the optimum parameters in our setting and demonstrate the performance of our protocol. Different from previous PSI protocols, the communication cost of our protocol varies with the cardinality of the intersection. In comparison, the communication cost of our protocol is the lowest, which gains an over $200\%$ improvement than the communication-optimized PSI protocol \emph{spot-low} (Pinkas et al., CRYPTO'19). In addition, the running time of our protocol is the lowest when the communication bandwidth is about $1$ Mbps.

ABSTRACT. Attitude analysis and recognition can be applied in wearable computing for medical assistance, motor-function assessment and dexterous human-robot interaction. The main problems, however, are serious drift and instability during traditional motion measurement fusion methods due to the high dynamic complexity of limb movements. To the best of our knowledge, it is the first attempt to employ an adaptive robust cubature Kalman filtering algorithm in the human attitude analysis based on wearable inertial sensors with time-varying state-process noise. Experiment results show that the adaptive robust CKF algorithm based on quaternion and gyroscope error modeling proposed in this paper can solve motion attitude solution. Lastly, we compare our method with CKF and EKF algorithm, the proposed algorithm can effectively improve the precision of attitude analysis.

ABSTRACT. Light field (LF) images can store multi-view geometry characteristics about the observed scene, which can be helpful in depth estimation. Depth estimation has attracted much attention in recent years for its widely use in the computer vision tasks. Many approaches have been proposed to estimate the depth of LF images, including conventional methods and learning-based methods. But most of them are hard to apply to different complex situations. We propose a robust depth estimation network for LF images with disparity warping (LF-DWNet), which is robust in large disparity pixels, occlusions, and noise areas. To reduce the effect of large disparity pixels, we introduce the disparity warping processing on EPI. To extract the depth feature from warped EPI and reduce the effect of occlusions and noise areas, we design a feature extraction module based on the attention mechanism. To make full use of the depth feature our attention-based module gets, we need to guide the depth estimation by the global structure information. Besides, our LF-DWNet can integrate the depth feature from multi streams of attention-based feature extraction modules and get more credible depth map. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of our method.

ABSTRACT. Head orientation tracking has many potential applications in various fields, e.g.,online courses, online meetings, and somatosensory games. Undoubtedly, with the information of the user’s head orientation, these applications will have more opportunities to enhance performance and provide better user experience. However, reviewing existing works regarding head tracking, the CV-based solutions have limited tracking angle range and privacy issues and the IMU-based solutions have accumulated errors. None of these methods provide accurate and stable user head orientation. In this paper, we propose HeadTracker, a fine-grained 3D head orientation tracking system based on a single headphone. Headtracker achieves high-precision head orientation tracking by installing ultrasonic transmitters on an ordinary headphone and deploying ultrasonic receivers in the environment. We conducted experiments to evaluate the performance of HeadTracker in the real use environment, and the experimental results show that the system can achieve an average error of 6 degrees in the 3D head orientation tracking. To the best of our knowledge, HeadTracker is the first system to use head-mounted ultrasound device to achieve 3D head orientation tracking and achieves the state-of-the-art in this category.

ABSTRACT. Network telemetry provides more accurate and reliable services for intelligent network control by pushing fresh status information actively with help of data plane. However, most existing network telemetry methods are difficult to be deployed effectively in business environment due to the lack of runtime-reconfigurability, huge time-space overhead, and high probability of information loss. In this work, we propose a high versatility approach NT-RP for network telemetry based on FPGA dynamic reconfigurable pipeline to maintain the balance between the accuracy of the measurement and the overhead in different scenarios. NT-RP can change the processing logic in runtime to obtain different network measurement spontaneously desired by users. Benefiting from distributed cyclic storage strategy and telemetry function integration mechanism, NT-RP can greatly reduce the overhead during measurement and mitigate the telemetry information missing problem caused by packet loss. The implementation of NT-RP in FPGA is evaluated in a real network testbed which consists of a few programmable nodes. Experimental results show that the influence of NT-RP in large traffic scenarios is less than 1%. It is not only able to successfully change the telemetry task during operation, but also perform more accurate network measurements with little telemetry information occupancy.

ABSTRACT. The popularity of smart devices and the availability of wireless networks bring considerable attention to Spatial Crowdsourcing (SC). Existing studies mainly focus on solutions to different optimization objectives of the SC platform, ignoring the entitlement of workers. This paper starts from the perspective of workers and investigates how to select suitable tasks for each online worker such that everyone can maximize their individual profit. Since the profit is related to the completion degree of tasks that is determined by the prior unknown parameter, we model the problem as a Multi-Agent Multi-Armed Bandit (MAMAB) problem. We propose a Payment-Estimation-Based Solution (PEBS), allowing workers to sequentially make decisions on task selection based on their observations and estimations. Specifically, the proposed PEBS first utilizes the social network among workers and assists workers in learning the information of tasks from the historical data. Then, it introduces the idea of probability matching in Thompson Sampling (TS) to help estimate the profit of workers and deal with the task selection problem. Finally, extensive simulations show that our proposed mechanism is efficient in optimizing the individual profit of workers.

ABSTRACT. Digital signature technology has been playing a very important role in electronic money transactions while memorizing and storing keys for digital signatures brings a great burden to users. In recent years, certifying authority (CA) has been attacked frequently, and many users have suffered losses. In this article, we dis-cuss the existing mainstream digital signature technologies, analyze their ad-vantages and disadvantages, and propose a novel fuzzy digital signature scheme, which has the characteristics of immediacy and no need for CA. Initially, our scheme uses fuzzy extractors to generate digital signatures seamlessly from bio-metric or other noise data, since its information entropy is sufficient to ensure se-curity. Secondly, we reduce the dependence of digital signatures on third parties and the threats and delays caused by them. Furthermore, We also conducted the computational security analysis of the proposed scheme to illustrate its security level. Finally, the comprehensive experimental evaluation elaborates that our scheme is reliable and practical for real scenarios, ensuring the information meas-ured by the user is untraceable.

ABSTRACT. The problem of limited defense resources owned by the network platform needs to be solved by designing a reasonable defense resource allocation scheme in Industrial Internet of Things (IIoT). However, most of the previously studied defense resource allocation schemes do not consider the impact of network cheat on the defender's total expected utility, resulting in the defender's total expected utility not being optimal. To address this problem, this paper proposes a network defense resource allocation scheme (NDRAS) with multi-armed bandits to maximize the defender's total expected utility. The scheme first proposes a random generation method of node shell configuration based on network cheat, by considering the impact of network cheat on the defender's total expected utility, masking information about the real configuration of nodes, to increase the uncertainty of the attacker's attack on each node and thus reduce the likelihood of the attacker's success. Subsequently, the decomposability and Lipschitz continuity of the defender's total expected utility is exploited to reduce the gap between the cumulative discrete optimal benefit and the continuous optimal benefit, to maximize the defender's total expected utility and thus make the defender's total expected utility optimal. Finally, the detailed experimental results confirm the effectiveness of NDRAS, indicating that the new scheme can give a reasonable defense resource allocation scheme to maximize the defender's total expected utility.

ABSTRACT. Packet classification plays a key role in network security systems such as firewalls and QoS. The so-called packet classification is to classify packets into different categories according to a set of predefined rules. When the traditional classification algorithm is implemented based on FPGA, memory resources are wasted in storing a large number of identical rule subfields, redundant length subfields, and useless wildcards in the rules. At the same time, due to the rough processing of range matching, the rules are extended. These problems seriously waste memory resources and pose a huge challenge to FPGAs with limited hardware resources. Therefore, a field mapping encoding bit vector (MEBV) scheme is proposed, which consists of a field-splitting-recombination architecture that can accurately divide each field into four mapping preparation fields according to the matching method, field reuse rate, and wildcard ratio, and also consists of four mapping encoding algorithms to complete the length compression of the rules, to achieve the purpose of saving resources. Experimental results show that for the 1K OpenFlow 1.0 ruleset, the algorithm can achieve a significant reduction in memory resources while maintaining high throughput and support range matching, and the scheme method can save an average of 38\% in memory consumption.

ABSTRACT. Recently, blockchain was introduced into the cyber-physical systems, which provides services of privacy and trust. However, reliability and system performance issues exist when blockchain and cyber-physical systems are integrated. In this paper, we design a blockchain-enabled cyber-physical system, where a new blockchain consensus is used to solve the problems of reliability and system performance. Firstly, an autonomous consensus mechanism called Proof-of-Weighted-PlannedBehavior is established based on the theory of planned behavior. Then, the behavior of consensus participants gets further explained by introducing credit evaluation and vulnerable node analysis. Moreover, considering the Jain fairness index, a dynamic authorizer group mechanism that coordinates reliability and decentralization is developed. By optimizing the credit threshold of the authorization group, the security and reliability of our designed mechanism are guaranteed. Finally, the experimental simulation results prove that compared with the traditional consensus, our proposed consensus improves the reliability and the system performance of the blockchain-enabled cyber-physical systems.

ABSTRACT. The large scale multiparty computation and private set intersection require a number of oblivious transfer instances as subroutines, but the implementation of oblivious transfer protocols is relatively slow. A feasible way is to use the oblivious transfer variant called random oblivious transfer. In this paper, we propose a 1-2 random oblivious transfer protocol and extend it to a 1-k random oblivious transfer protocol based on the LWE assumption, quantum mechanics. Then, we analysis the stand-alone security of our 1-2 random oblivious transfer protocol under various malicious situations and prove its universally composable security in UC framework. As for the security of our 1-k random oblivious transfer protocol, the similar results can be obtained.

ABSTRACT. With the large scale NGSO constellation systems entering the blowout launch period, new challenges are expected for inter-system satellite coexistence due to the increase in scale of the constellations and the complexity of the interactions, where the problems of frequency compatibility between NGSO system and GSO system are the most prominent. However, the existing methods normally analyze the interference through real-time position acquisition, visible area judgment and single link calculation, as for the large scale constellations under construction, these methods still have great challenges to large amount of data stored in constellation positions, long simulation duration and high requirements for simulation equipment. In this circumstance, we establish a high efficient interference calculation model apply for large scale constellations. Our solution stems from the satellite position probability, by meshing the visible area and deriving the distribution of other interfering satellites in visible area through the initial satellite, we can directly analyze aggregate interference. The proposed model not only has the advantages of low algorithm complexity and high calculation efficiency, but also suitable for large scale constellations with different configurations as well as not depend on computer simulation equipment. The theoretical and simulation results both show that the proposed model has advantages in the accuracy of interference calculation and the duration of interference simulation.

ABSTRACT. Detecting vulnerabilities is essential to maintaining software security. At present, vulnerability detection based on deep learning has achieved remarkable results. The type of vulnerability could tell the vulnerability principles and help the programmer quickly pinpoint the precise location of vulnerabilities. Moreover, the type of vulnerability is very valuable for remediating it. Therefore, it is essential to identify vulnerability types. This paper proposes a new vulnerability type identification framework based on deep learning. The framework is based on syntax and semantics, and the detection granularity is fine to the slice level. To include comprehensive vulnerability types, we use four slicing methods to represent the program. In addition, we model four kinds of code slice features based on BERT. For evaluation, we used 64 three-level CWE-IDs vulnerability types in National Vulnerability Database (NVD) and Software Assurance Reference Dataset (SARD) for vulnerability type identification. The experimental results show that it has significant performance in vulnerability type identification.

ABSTRACT. BBR effectively improves the network bandwidth utilization by its unique minimum delay and maximum bandwidth detection mechanism. However, with the development of 5G communication technology, whether the 10s delay detection interval of BBR can meet the new high throughput and low latency heterogeneous network requirements needs to be studied. Therefore, based on the ns-3, this paper builds some scenarios to simulate the performance of BBR in wired, WiFi, and 5G networks. A spindle-shaped network topology is constructed to simulate the BBR competition. By modifying the delay detection interval of BBR to 5s and 1s, the competition among BBR streams with the same roundtrip time (RTT), the competition among BBR streams with different RTTs, and the competition among BBR and other TCP congestion control algorithms (CCA) are simulated respectively. Then, a formula for calculating the delay detection interval is proposed. According to this formula, we propose a method to dynamically modify the delay detection interval. The method estimates the network state according to the change of RTT, and then calculates and updates the delay detection interval. Simulation results demonstrate that appropriately modifying the delay detection interval of BBR can alleviate the competition among BBR and other algorithms in heterogeneous wireless network, but too short delay detection interval will deteriorate the network performance.

ABSTRACT. As one of the most important basic services of the Internet, the domain name system is abused by attackers for various malicious activities, malicious domain detection is a key technology against attackers. Previous works mainly employ manually selected features to detect malicious domains which are easily evaded by attackers. In this paper, we propose a novel malicious domain detection system with heterogeneous graph propagation network, named HGPNDom, which can jointly consider the global relationship and higher-order features of domains. In HGPNDom, we first model the DNS scene as a heterogeneous information network (HIN) to capture rich information, then we propose a heterogeneous graph propagation network (HGPN) to classify domain nodes in the HIN, including semantic propagation mechanism and semantic fusion mechanism. The semantic propagation mechanism can spread information through more layers and learn higher-order domain features, the semantic fusion mechanism can learn the importance of meta-path and fuse them judiciously. Experimental results on the real DNS dataset show that HGPNDom outperforms other state-of-the-art methods.

ABSTRACT. Image caption is textual explanation automatically generated by a computer according to the content in an image. It involves both image and natural language processing, and thus becomes an important research topic in pattern recognition. Deep learning has been successful in accomplishing this task, and the quality of captions generated by existing methods is already high. However, due to the broadness and variety of image caption applications, the current generated captions are still not sufficiently detailed, and the training efficiency can also be improved. Therefore, under the encoder-decoder framework of deep learning, how to use fewer parameters to improve the training efficiency and retain the quality of the generated image descriptions is a huge challenge. In this work, we introduce an improved method based on the encoder-decoder structure, adding an attention mechanism, and applying the content adaptive recurrent unit (CARU), as the decoder, to generate image captions. Inspired by GRU, CARU is designed to have comparable performance with fewer parameters, and is sensitive to the features in hidden layers. The experimental results show, based on MsCOCO dataset, the proposed method achieved better performance than that using GRU as the decoder, and took less training time, effectively improves the training efficiency.

ABSTRACT. Lane detection is a challenging task in the field of vision detection. The annotation information of lane is very sparse, and it is faced with the interference of occlusion, illumination and other factors, which seriously affects the capture of lane features by neural network. In this paper, we propose the Self-Attention Lane Segmentation Network (SALSN) which allows attention-driven, long-range dependency modeling for lane detection task. Although traditional convolutional neural networks have demonstrated their powerful performance, their ability to capture global relationships in images has not been fully explored. We introduce a self-attentive module to model the long-range dependencies between lane features. Lanes have strong shape constraints but weak coherence. In SALSN, we utilize a dense feature fusion framework to better capture lane context information and use all element information to generate lane segmentation images. Experimental results show that SALSN is not only effective in learning the remote dependencies of lane features, but also significantly improves the lane detection performance. We have validated our approach on two large-scale lane detection datasets, and our method can achieve more competitive results.

ABSTRACT. The main task of Cross-Domain Sentiment Classification is to train a well-performing classification model by using labeled source domain data, and then transfer the model to unlabeled target domain data, thereby solving the expensive labor consumption and domain shift caused by a large number of labels resulting performance degradation. Most mainstream adversarial domain adaptation methods are based on a single discriminator, which ignores the uneven distribution of labels between domains and multiple modalities of data and tends to cause negative transfer and poor generalization performance. We propose a Dropout-based ensemble dual discriminator for Cross-Domain Sentiment Classification. We functionally decouple the single discriminator by using two forms of text data, and replace it with a positive sentiment discriminator and a negative sentiment discriminator. A dynamic set of discriminators will be obtained by random deactivation of the discriminator network neurons, then the feature extractor has to extract richer and more realistic domain-invariant features to fool the discriminator and mitigate the mode collapse phenomenon. To solve the problem of class imbalance in a large number of unlabeled data samples, we use mutual information maximization to train sentiment classifiers to ensure that label predictions are distributed in a reasonably balanced state. We conduct full experiments on the Amazon and Airlines datasets. Experiments showed that our proposed model achieves state-of-the-art cross-domain sentiment classification performance.

ABSTRACT. Low transaction efficiency remains one of the primary constraints to the development of permission blockchain. To enhance the communication performance of blockchain , the majority of research focuses on optimizing the local architecture of blockchain and improving consensus. In practice, increasing the block dissemination capability at the network layer can significantly improve transaction efficiency. We find that the redundancy and instability of the gossip protocol as a broadcast method in Hyperledger Fabric have a significant impact on communication performance. In this work, we introduce the idea of density clustering to propose the DC-Gossip broadcast protocol, constructing a stable network architecture with highly dense connectivity for the blockchain network layer. This architecture can effectively reduce the propagation latency and ensure the integrity of the distributed ledger. In our experiments with Fabric, DC-Gossip reduces latency by more than 19\% after 40 blocks are propagated in a stable network environment with more than 100 nodes. Moreover, the latency decreases by 14\% in a dynamic network under the identical circumstances.

ABSTRACT. Image Semantic Segmentation is an indispensable assignment in computer vision. It has been widely applied in many fields such as medical imaging and robotics. For the sake of meeting the application of real-time scene, how to improve segmentation accuracy while ensuring high segmentation speed has become a new research hotspot. Aims at this problem, a real-time image semantic segmentation method based on dual efficient attention mechanism (DEANet) is proposed. The proposed dual efficient attention module is used to replace the decoder in an encoder-decoder based light-weight network. Pyramid sampling is introduced in the channel dimension to extract multi-scale information, and higher resolution aggregation features are adopted as the input of the spatial dimension to promote the segmentation accuracy and efficiency. The proposed DEANet is tested on two classic datasets. On the Cityscapes dataset, when the input size is 512×1024, the segmentation accuracy reaches 74.90% mIoU and the segmentation speed reaches 99.91FPS. On the Camvid dataset, when the input size is 360×480, the segmentation accuracy reaches 70.07% mIoU and the segmentation speed reaches 142.72FPS.

ABSTRACT. In recent years, the poor generalizability of deep neural networks in multi-model medical images has attracted widespread attention. Domain adaptation is an approach to alleviate the above problem, which transfers the labeled source domain to the target domain. It can reduce the data labeling workload in the target domain and significantly improve the network's generalizability. However, the differences between foreground areas and background areas of medical images are relatively minor, and it is difficult for existing methods to effectively extract domain invariant features. Further optimization of the feature distribution alignment for each category is also lacking. Therefore, a Multi-task Class feature space Fusion Domain Adaptation Network (MCFDAN) is proposed in this paper. Firstly, a reconstruction branch is added to the baseline network to mitigate the feature offset of the target domain during encoding. Secondly, category constraints are added to the fusion of domain feature spaces, improving the generalizability of the source classifier to the target domain. Finally, the network incorporates a recurrent cross-attention module that highlights the feature expression of the lesion region. The evaluation results demonstrate that the proposed network achieves a significant performance improvement, which is important for the application of smart healthcare systems.

ABSTRACT. Due to the drawbacks of federated learning such as vulnerability of a single central server, centralized federated learning is shifting to decentralized federated learning, a paradigm which takes the advantages of blockchain. A key enabler for adoption of blockchain-based federated learning is how to select suitable participants to train models collaboratively. Selecting participants by storing and querying the metadata of data owners on blockchain could ensure the reliability of selected data owners, which is helpful to obtain high-quality models in federated learning. However, querying multi-dimensional metadata on blockchain needs to traverse every transaction in each block, making the query time-consuming. An efficient query method for multi-dimensional metadata in the blockchain for selecting participants in federated learning is absent and challenging. In this paper, we propose a novel data structure to improve the query efficiency within each block named MerkleRB-Tree. In detail, we leverage Minimal Bounding Rectangle(MBR) and bloom-filters for the query process of multi-dimensional continuous-valued attributes and discrete-valued attributes respectively. Furthermore, we migrate the idea of the skip list along with an MBR and a bloom filter at the head of each block to enhance the query efficiency for inter-blocks. The performance analysis and extensive evaluation results on the benchmark dataset demonstrate the superiority of our method in blockchain-based federated learning.

ABSTRACT. Due to complex blockchain programs and numerous blockchain nodes, it takes a huge amount of time and economic cost to conduct blockchain experiments. Existing open source projects do not support modifications to the underlying blockchain, and existing blockchain simulators only focus on a single blockchain system and cannot flexibly extend or replace models. Regarding the issue above, this paper proposes a prototype system for blockchain performance evaluation, including real deployment test and simulation test. In real deployment test, a five-layer architecture for building a lightweight and efficient testing system is proposed. And in simulation test, a general scheme for building blockchain simulator is proposed, which can realize the test of throughput, storage allocation and reputation management. Experiments show that the prototype system proposed in this paper can effectively improve the efficiency of blockchain performance evaluation.

ABSTRACT. Ciphertext-policy attribute-based encryption (CP-ABE) is widely used in access control to achieve secure data sharing over different cloud platforms. However, most of the existing CP-ABE data sharing schemes perform one-time encryption on the shared data, which cannot satisfy the need for data sharing in the public cloud with complex users. In order to realize multi-user data sharing on blockchain and achieve hierarchical decryption of privacy data and shared data, we propose a blockchain-aided hierarchical and searchable attribute-based encryption scheme, named BC-HSABE. In BC-HSABE, we adopt a symmetric encryption algorithm to encrypt data in a hierarchical manner, and use attribute-based encryption technology to encrypt two hierarchical symmetric keys of the encrypted data file, and upload the ciphertext to the cloud server. In addition, data users can decrypt data ciphertext at different levels according to their hierarchical authority. Meanwhile, searching for keyword trapdoor through blockchain ensures the security of keyword ciphertext. The security analysis and experimental evaluations verify the feasibility and effectiveness of BC-HSABE.

ABSTRACT. In recent years, the dynamic environment on the Internet of Vehicles led vehicular communication networks' trust management mechanism to become a research hotspot. Most of the existing trust management in vehicular networks relies on a centralized third party. However, it causes trust management to be limited to a single node and has high requirements for device performance. In order to improve the reliability of information exchanged between vehicles, we propose a scalable blockchain-based trust management strategy, which employs vehicle-related objective factors to evaluate the credibility of information transmitted between vehicles to determine the vehicles' trust level. We also design a consensus mechanism to make all RSUs(Roadside Units) maintain a consistent and reliable distributed ledger as nodes so that vehicles can obtain global trust information more quickly during interaction to improve its reliability. The security and performance analysis shows that our strategy has high reliability and availability.

ABSTRACT. As crowdsourcing continues to evolve, researchers have begun to explore task matching in crowdsourcing extensively. However, the privacy issues such as task content of publishers and ability or interest of worker in task matching are often overlooked. Also, the identity of the task publisher/worker needs to be protected. To address the above issues, we propose a secure task matching scheme in crowdsourcing based on blockchain in this paper. Firstly, we implement multi-publisher/multi-worker task matching in the scheme while protecting task content privacy. Meanwhile, we take advantage of the immutability of the blockchain to ensure the reliability of publishing/matching results. We utilize the smart contract for task publishing/matching without human intervention. Finally, the scheme is shown to be secure and feasible through theoretical and comprehensive performance evaluations.

ABSTRACT. Blockchain-based decentralized federated learning (BCFL) protects data privacy and avoids the single point of failure, which has become a key technology in the Intelligent Internet of Things application. However, the BCFL still be challenged by model attacks from malicious devices. In addition, the BCFL needs to select training devices from a large number of unknown candidate devices. Due to the training resources of devices being quite different, improper selection of unknown devices will lead to inefficient training. To address these issues, we consider two types of Byzantine model attacks and design "Proof of Model Quality"(PoMQ) to protect the security of BCFL. The PoMQ is a three-phase consensus algorithm, which combines the FedAvg and Multi-Krum algorithms to defend against model update attacks. The PoMQ is also based on the model verification process to resist model aggregation attacks. Moreover, we define the unknown device selection problem as a Combinatorial Multi-Armed Bandit problem (CMAB) to improve the training efficiency, and propose an online learning algorithm based on PoMQ consensus (OLAC) to solve CMAB. Finally, through analysis and extensive experiments, this paper proves that PoMQ and OLAC can better improve the robustness and training efficiency of BCFL.

ABSTRACT. Vehicular ad-hoc network (VANET) is emerging as the most essential technology for securing traffic safety as well as enhancing traffic conditions. The confidentiality of messages transmitted in VANETs is of utmost priority, which has attracted extensive scholarly research. However, the security of the existing studied encryption schemes remains to be upgraded with a massy time overhead. Hence, we propose a lightweight encryption scheme for secure message transmission in VANET with a private blockchain. Firstly, the encryption scheme we designed, called SVM, ensures the confidentiality and forward security of messages, while the system is sound and amnesic so that we are not entirely dependent on external servers. Meanwhile, the private blockchain with high privacy-preserving and flexible read/write access assures that merely permitted users to access the message. Furthermore, our scheme achieves excellent lightweight, thus suitable for VANETs with huge messages. Security analysis indicates that our scheme has firmly secured the transmission of messages in VANETs. Performance evaluation demonstrates that our scheme is lightweight and efficient, with encryption and decryption times of 7.00ms and 7.67ms, respectively.