ABSTRACT. In this talk, first, we will introduce the long-term challenges of the development of IoT. Combining AI theory, then we will present some explorations and recent research progresses on intelligent sensing, intelligent transmission, and intelligent service in the IoT environment. In the future, endogenous intelligence will drive the revolution of IoT, we will discuss the open issues on IoT area, such as the theories and key technologies of human-like sensing, concise and intelligent networking for heterogeneous wireless network, and cognitive service. The breakthrough for solving the above problems will promote the development of Internet of Things.

ABSTRACT. In the era of cloud computing and big data, the global data volume presents an explosive growth trend, triggering profound technological and commercial changes worldwide, becoming the focus of competition between countries and enterprises, and being designated as a new factor of production. Data resources are increasingly becoming a production factor and strategic asset of human society. The integration and application of cross domain data will generate immeasurable value and help promote innovative applications, such as intelligent medical treatment, situational cognition, precision marketing, intelligent monitoring, etc., which can significantly improve social and economic benefits. The realization of secure sharing and trading of big data will help break down industry information barriers, optimize and improve production efficiency, and deeply promote industrial innovation. However, due to the ease of data replication, difficulty in value quantification, difficulty in channel control, and user concerns about data security, how to convert the currently closed and hidden data sharing into open and public data circulation and trading has become one of the core challenges that need to be addressed in the era of big data and artificial intelligence, which urgently needs us to conduct corresponding research. In this report, I will share with you some challenges and scientific issues in the big data sharing and trading market, and introduce some of our recent explorations, including data authentication, quality assessment, value assessment and pricing, security and privacy protection, and so on.

ABSTRACT. Wireless technologies have achieved great success in data communication. In the last few years, wireless signals (e.g., WiFi) have been exploited for sensing purposes, enabling exciting applications such as passive localization, contact-free gesture recognition and vital sign monitoring. Although promising progress has been achieved, the potential benefits of wireless sensing are still limited by several fundamental issues including small sensing range and poor performance under device motions. In this talk, I introduce our recent research (i) on the theory side to enable wireless sensing under device motions for the first time and (ii) on the application side to realize fine-grained eye-blink detection with wireless sensing. At the end of the talk, I will briefly discuss several practical challenges associated with wireless sensing and the future directions in this area, pushing wireless sensing one step forward towards real-world adoption.

ABSTRACT. Reliable routing is a classic problem in the field of computer networks. After a network fault occurs, how to choose the recovery path directly determines the performance of network services. This paper introduces service customized techniques into reliable routing. By meeting customized traffic protection re-quirements, network service quality can be ensured after fault recovery. Topol-ogy Independent Loop-free Alternate (TI-LFA) supported by SRv6 is a new re-liable routing technology. In this paper, an SRv6-based service customized reli-able routing mechanism is designed for the single link failure in the case of P-Q space adjacency in TI-LFA. For traffic with QoS requirements, this paper uses fuzzy theory to make the optimal decision for SRv6 candidate protection schemes. Finally, three representative topologies are selected to build an exper-imental network supporting SRv6 based on ONOS, Mininet, and the program-mable data plane. The results show that when responding to a network service customized request, the recovery path selected by the mechanism proposed in this paper is superior to the comparison mechanism of related QoS indicators.

ABSTRACT. Aiming at the problems of high end-to-end transmission delay and low packet delivery rate caused by the high mobility of unmanned aerial vehicle (UAV) nodes, a routing protocol based on the improved equal dimension new information GM(1,1) model (IEDNI-GM) is proposed. By analyzing the motion characteristics of the UAV node, combine the gray prediction model and the Markov chain model to construct IEDNI-GM to predict the location of the UAV node at the next moment. Meanwhile, the paper combines the advantage that clustering structure can optimize network management. We consider the motion state and the communication link state between nodes and use the predicted value of node position to calculate the value of link holding time, motion similarity and expected transmission count. The cluster-head election indicator is constructed by combining these three values, and the UAV nodes in the network are clustered. This clustering structure is adopted to improve the AODV routing protocol. Therefore, the source node can find an effective communication route to the destination node. Experiments under the network simulator NS-3 show that compared with routing protocols such as AODV and AODV-ETX, the routing protocol in this paper can effectively reduce the end-to-end average transmission delay, increase the delivery rate of data packets, and is more suitable for UANET.

ABSTRACT. Unmanned Aerial Vehicles (UAVs) have been widely used in both military and civilian scenarios since they are low in cost and flexible in use. They can adapt to a wide variety of dangerous scenarios and complete many tasks the Manned Aerial Vehicles (MAVs) can not undertake. In order to establish connectivity and collect data in large areas, numerous UAVs often cooperate with each other and set up a UAV wireless network. Many multi-hop routing protocols have been proposed to efficiently deliver messages with high delivery ratio and low energy consumption. However, most of them do not consider that the power level of UAVs is adjustable. In this paper, we propose a Power-Aware Routing (PAR) algorithm for UAV networks. PAR utilizes the pre-planned trajectory information of UAVs to compute the encounters at different power levels, and then constructs a power-aware encounter tree to calculate the transmission path with minimum energy consumption from the source to the destination within the delay constraint. Through extensive simulations, we demonstrate that compared with three classic algorithms, PAR significantly reduces the energy consumption and improves the network performance on the basis of ensuring timely delivery of packets.

ABSTRACT. Nowadays, Time-Sensitive Networking (TSN) has widespread application in many industrial fields, aiming to provide deterministic lowlatency network transmission. Traffic in TSN is roughly divided into three categories: Time-Triggered (TT) traffic, Audio-Video-Bridging (AVB) traffic, and Best-Effort (BE) traffic. These different traffic needs to travel the network satisfying their respective reliability and performance requirements. Existing traffic routing and scheduling mechanisms mainly focus on TT traffic but pay little attention to other traffic types. In this paper, we present a novel Optimization Modulo Theories (OMT) formulation for a comprehensive traffic routing and scheduling problem in TSN. Based on this, we propose a novel reliability-aware routing and scheduling mechanisms for all traffic types, in order to improve their own transmission reliability and performance. We conduct extensive evaluations to validate the effectiveness of the proposed mechanisms, and the results confirm that the proposed mechanism can really guarantee the reliability and latency requirements of TT flows and improve the transmission utility of all flows to a large extent.

ABSTRACT. In opportunistic networks where node clusters exist, it is usually necessary to set up ferry nodes to connect each node cluster to achieve the overall connectivity of the network, and the movement pattern of ferry nodes has an important impact on the overall performance of the network. The existing opportunistic network routing algorithm based on ferry nodes has the problems of insufficient resource utilization and inefficient collaboration of multiple nodes in resource optimization and collaborative work, which often leads to low overall network delivery rate and high network load. Therefore, this paper proposes a cooperative routing algorithm for multiple ferry nodes based on active motion mode, in which ferry nodes actively realize the planning of motion paths according to their own states and network message forwarding requirements, while multiple ferry nodes in the network realize cooperative work. Experimental results show that the proposed algorithm (Opportunistic network routing algorithm based on ferry node cluster active motion and collaborative computing, ORABAC) improves the message delivery rate of the network while reducing the energy consumption of ferry nodes, decreases the message delivery delay, and improves the network performance.

ABSTRACT. Loop Free Alternate (LFA for short) is designed to avoid the interruption of transmission paths during routing convergence when a network suffers from single point failures (e.g. node failure, or link failure). Unfortunately, once more than one associated failure occurs simultaneously, the LFA protocol cannot strictly guarantee lossless packets. Therefore, we propose a LFA-based percolation routing mechanism (P-LFA for short). It establishes the shortest paths and their backup paths by using distributed routing protocol and LFA, and uses a percolation rerouting algorithm to calculate the percolation paths. Once the shortest path and corresponding backup path are interrupted, the network node immediately selects the percolation paths to continue transmitting packets without dropping any packets. And this mechanism allows each node to cache packets. This way can efficiently avoid packet loss when the path to the destination node is unreachable. The experimental results show that P-LFA can effectively guarantee zero-loss packets when some related failures occur. Meanwhile, even though there do not exist reachable paths to the destination node, the P-LFA mechanism can cache unreachable packets into the network node that is the closest to the destination node, and continue to re-forward packets when paths recover.

ABSTRACT. Maritime Ad hoc networks are a type of decentralised wireless network with rapid networking and multi-hop routing, which are independent of fixed base stations. Recently, Ad hoc networks have started to play an increasingly important role in military command, emergency rescue, disaster relief, temporary meetings, and other occasions. However, as the network topology changes rapidly and the node energy and network bandwidth are limited, discovering and maintaining reliable transmission paths have become a highly topical challenge. In order to solve the problem that distributed routing planning of large-scale Ad hoc networks cannot adapt dynamic changes in network topology, and considering the differences of network nodes, this paper proposes federated reinforcement learning to improve the efficiency of distributed routing planning through the joint learning of similar nodes. The routing strategies of different network nodes vary but neighbouring nodes have very similar routing tables. Therefore, our federated reinforcement algorithm learns nodes with similar routing policies. In this study, a communication system simulation software is specially designed to evaluate the performance of the proposed algorithm.

ABSTRACT. The security of medical data sharing (MDS) plays an important role in the area of healthcare. Significantly, achieving its security faces more challenges due to the feature of multiparty holding, higher complexity, and serious data silos. Different from traditional secure schemes, which established model cannot deal with the above three problems due to the low accuracy of the MDS secure model, this paper design a novel secure MDS model and two schemes to increase the accuracy of the model. In detail, to eliminate the issues of data silos and point failure, we combine the federated learning (FL) with blockchain technology into MDS secure model, and the data confidentiality of the exchanged data in the process of FL can be further ensured by differential privacy (DP). Then, to increase the accuracy of the secure MDS model, we design a validation incentive mechanism based on model quality (VIM) and an effective DP method with assigned weights (AWDP), in terms of participants' enthusiasm and noise accumulation, respectively. Simulations show that the established model is effective and correct and the designed VIM and AWDP can achieve higher accuracy than current popular methods, resulting in 30% increment.

ABSTRACT. Radio Frequency Identification (RFID) has been adopted in various applications owning to its many attractive properties such as low cost, no requirement on line-of-sight, and battery-free. This paper studies the problem of RFID-based Handwriting recognition, which is practically important in Human-Computer Interaction (HCI) scenarios. To the best of our knowledge, the state-of-the-art works beget leaking user privacy, because the malicious attacker can eavesdrop on the RFID signals (e.g., tag phase) broadcast in the air and further analyze the user's handwriting activity. To address the privacy leakage issue, we propose a secure RFID handwriting recognition system named SecRFPen to enable privacy-preserving handwriting recognition. In SecRFPen, the legal reader switches the probing frequency and power, the phase angles of RF signals reflected by the tagged pen will change accordingly. both reader-tag hardware characteristics and handwriting movements. We propose an authentication matrix to quantify RFID device hardware characteristics, which can be measured by legal users in advance. Thus, the legal RFID reader can recognize the handwriting activity via analytics on the authentication matrix and tag phase profile. On the contrary, since the malicious attacker knows nothing about the hardware characteristics of legal RFID devices, it cannot understand handwriting even if it can hear the tag signals. We implement the SecRFPen system based on the Commercial-Off-The-Shelf (COTS) RFID devices. Extensive experimental results demonstrate that the recognition accuracy of legal users can reach 94.2%, while the recognition accuracy of the malicious attacker is as low as 35.1%.

ABSTRACT. Federated learning is a novel machine learning framework that effectively satisfies the requirements of multiple organizations for data usage and model training while meeting privacy protection, data security, and government regulations. However, recent research has shown that attackers can infer users' private information from their shared model parameters. To address the issue, in this paper, we propose the smart contract assisted secure aggregation scheme (SCSA). Firstly, we innovatively present a triple layers architecture for secure aggregation, which can adapt to application scenarios where a large amount of devices are involved in model training. Then, with the help of smart contracts, our scheme can efficiently distribute security masks to users in a decentralized form to ensure the security of parameters, and combine with secret sharing to design a double fault tolerance mechanism to effectively improve the robustness of the system. Finally, the theoretical analysis and simulation experiments prove that our scheme has high security and robustness while maintaining efficiency.

ABSTRACT. Smart home applications make our lives more comfortable, more convenient than ever before. However, deploying smart home applications and smart terminals could pose a potential security threat to personal information and home privacy. In order to prevent illegal use of terminals and applications, it is very necessary to establish secure and reliable communication between terminal and edge server. In this paper, we design a two-party authentication and key negotiation protocol for the smart terminal and edge service. The edge-based authentication and key negotiation scheme offloads the terminal's main computational overhead to the edge side, and exploits cryptographic algorithms to achieve user anonymity and untraceability. Security is verified by the BAN logic and AVISPA. We also evaluate the performance by comparing our scheme with other related schemes in terms of computational overhead. The security and performance results show that our proposed scheme is suitable for edge-assisted smart home applications.

ABSTRACT. Encrypted traffic has become the primary carrier of network transmission, and encrypted traffic classification is an essential support for advanced network management and network security protection. The existing studies mainly focus on introducing new classification models to improve the performance under the same features. Some studies also start from the feature engineering of encrypted traffic and try their best to select more expressive features from encrypted traffic for classification. At present, feature engineering uses statistical features or sequence features for classification, both of which have their advantages, but both are at the network or transport layers. Due to the decoupling of each layer in the network protocol stack and the data segmentation, both caused by protocol engineering, the network or transport layer features are more inclined to the factors of network transmission rather than the data attributes of applications or services. As a result, the relevance of the features and application or services is not strong, and the performance is not satisfied. To solve this problem, we introduce the Application Data Unit (ADU) and put forward the application layer feature engineering, which uses the features of the highest protocol level - the application layer to achieve better HTTPS classification. The experimental results show that the proposed ADU features are better than the segment granularity features of the TLS layer and far better than the packet granularity features both in statistical and length sequence features. The average F1-score increase in the application classification scenario of encrypted traffic is more than 10%.

ABSTRACT. With the explosive growth of mobile traffic and the demand for privacy protection and network security, mainstream mobile applications use encryption protocols (mostly TLS), so identifying encrypted traffic has become critical. Traditional machine learning methods are based on hand-designed features that are unreliable in the face of complex traffic data. Although deep learning currently performs well on this task, most of them only describe traffic data from one view, ignoring the heterogeneous nature of traffic. In this paper, we apply multimodal Transformers to mobile encrypted traffic classification and propose a novel model (DF-Net) with a deep fusion mechanism. The key point of deep fusion is that a learnable modal-type embedding enables the model to perform early and unconstrained fusion and interaction of cross-modal information to achieve performance improvements. On the premise of ensuring performance, DF-Net adopts the simplest embedding scheme, and the multi-head self-attention mechanism brings parallel capability. Both the lightweight design and the parallel mechanism can improve the overall efficiency of the model. To verify the performance and efficiency of DF-Net, we implement an automated traffic collection framework to collect a real-world traffic dataset that covers 48 popular apps. Experiments show that DF-Net not only achieves excellent performance but also has better running efficiency compared to state-of-the-art methods.

ABSTRACT. Identity-based proxy re-encryption (IB-PRE) is a type of public key cryptography that allows a proxy to convert a ciphertext under Alice's identity into another ciphertext of the same message under Bob's identity, but the proxy can not access the participants' secret keys or underlying plaintext. As far as practical application is concerned, a key revocation mechanism is an essential feature of an identity-based encryption system. By extending IB-PRE scheme, we propose a new cryptographic primitive revocable identity-based proxy re-encryption with re-encryption verifiability (RIB-VPRE), which allows the RIB-VPRE scheme to support the users' revocation, delegation of decryption rights and re-encryption verifiability at the same time. In this paper, we give the first concrete construction of collusion-resistant unidirectional RIB-VPRE on lattice, which is secure under the standard model based on learning with error (LWE) for both selective and adaptive identities.

ABSTRACT. With the popularity of intelligent vehicles, computation-intensive vehicle tasks rise dramatically. Vehicle edge computing (VEC) is a promising technology that offloads overloaded computation tasks of intelligent vehicles to the edge. However, VEC servers are constrained by their available computation capacity while dealing with numerous tasks. To this end, we propose multi-party cooperation to complete vehicle task offloading. Computation-assisted vehicles (CAVs) with free resources assist VEC servers to offload Computation-required vehicles (CRVs), which enables computation resources of VEC servers and CAVs for CRVs' task execution. To motivate positive participation of VEC servers and CAVs, we design a resource management and pricing mechanism by quantifying their gains and costs. Such design efficiently integrate and leverage the communication mode and computing mode among participants to describe their interactions, which composes two two-stage Stackelberg games. While Nash equilibrium (NE) for each Stackelberg game reaches, none of participants violates unilaterally. Simulation results demonstrate its effectiveness of the proposed model.

ABSTRACT. Compared with traditional mobile edge computing (MEC), heterogeneous MEC (H-MEC), which is assisted by ground vehicles (GVs) and unmanned aerial vehicles (UAVs) simultaneously, is attracting more and more attention from both academia and industry. By deploying base stations (along with edge servers) on GVs or UAVs, H-MEC is more suitable for access-demand dynamically-changing network environments, e.g., sports matches, traffic management, and emergency rescue. However, it is non-trivial to perform real-time user association and resource allocation in large-scale H-MEC environments. Motivated by this, we propose a shared multi-agent proximal policy optimization (SMAPPO) algorithm based on the centralized training and distributed execution framework. Due to the NP-hard difficulty of jointly optimizing user association and resource allocation for H-MEC, we adopt the actor-critic-based online-policy gradient (PG) algorithm to obtain near-optimal solutions with low scheduling complexities. In addition, considering the low sampling efficiency of PG, we introduce importance sampling that can greatly increase the training efficiency of proximal policy optimization. Moreover, we leverage the idea of centralized training and distributed execution to further improve the training efficiency and reducing scheduling complexity, so that each mobile device makes decisions based only on local observation without the need of gathering global states. Extensive simulation results demonstrate that SMAPPO can achieve more satisfactory performances than traditional algorithms do.

ABSTRACT. In vehicular edge computing, the edge servers may be overloaded once too many vehicles request for the task offloading service, which will cause task offloading failure or the high service delay. In this paper, considering the cooperation of cloud center, edge sever and the vehicle, we model the cloud-edge-vehicle cooperation in task offloading as three-party game problem. Then, based on backward induction method, we transform the game problem into a convex optimization problem, and theoretically prove the game has a unique Nash equilibrium. Meanwhile, a cloud-edge-vehicle tripartite collaborative task offloading algorithm based on genetic algorithm is proposed to find the optimal solution of the convex optimization problem. Simulation results show that the proposed strategy can make full use of existing resources to undertake more offloading tasks and has better performance than other solutions in term of task offloading delay.

ABSTRACT. In the field of edge computing, collaborative computing offloading, in which edge users offload tasks to adjacent mobile devices with rich resources in an opportunistic manner, provides a promising example to meet the requirements of low latency. However, most of the previous work has been based on the assumption that these mobile devices are willing to serve edge users, with no incentive strategy. In this paper, an online auction-based strategy is proposed, in which both users and mobile devices can interact dynamically with the system. The auction strategy proposed in this paper is based on an online approach to optimize the long-term utility of the system, such as start time, length and size, resource requirements, and evaluation valuation, without knowing the future. Experiments verify that the proposed online auction strategy achieves the expected attributes such as individual rationality, authenticity and computational ease of handling. In addition, the index of theoretical competitive ratio also indicates that the proposed online mechanism achieves near-offline optimal long-term utility performance.

ABSTRACT. With the rapid development of mobile communications, a large number of latency-sensitive and computation-intensive mobile applications have emerged. There is a huge contradiction between the high resource demands of these applications and the limited resource of mobile devices. In this regard, mobile edge computing (MEC) is a promising technology, where computation tasks can be offloaded from mobile devices onto network edges with strong capability. However, the dependency between different tasks lead to high complexity for offloading decision. In this paper, we investigate the optimal offloading problem for completing dependency-aware tasks by minimizing the time latency and energy cost. We propose an improved NSGA-II algorithm to solve this multi-objective problem subjects to the cost and reliability constraints. Simulation results validate the advantage of the proposed algorithm in terms of the performance of low cost and latency.

ABSTRACT. Mobile edge computing (MEC) is considered as a key technology for addressing computation-intensive and delay-critical applications in the Internet of vehicles (IoV). In MEC-enabled IoV, vehicles lighten their computing load by offloading tasks to edge servers. However, the high speed mobility of vehicles and time-varying network environment bring tough challenges to task offloading. In addition, considering only roadside units (RSUs) or vehicles as offloading objects lead to the waste of computing resources and increase the process delay of task. We express the reduction of task processing delay and improvement of service reliability as an utility maximization problem and propose a distributed vehicle-road cooperative task offloading scheme with task migration. Then we use RSUs and surrounding vehicles as offloading objects and divide offloading tasks into multiple subtasks for offloading objects and local parallel processing, which improves the utilization rate of computing resources. Meanwhile, we reduce the task processing failure by migrating the computing results of offloading subtask. The offloading scheme is formulated as a mixed-integer nonlinear optimization problem, and a Multi-Agent Deep Q-learning Network (MADQN) is proposed to find the near-optimal offloading objects and number of offloading subtasks. Simulation results show that the proposed approach improves the offloading utility and outperforms the baseline approach.

ABSTRACT. Unmanned Aerial Vehicles (UAV) supported by 5G networks can play an important role in providing aerial-aerial/aerial-ground computing services to remote and isolated areas at a low cost. In this paper, we present an aerial-aerial-ground network (AAGN) architecture using High Altitude Unmanned Aerial Vehicles (HAU) and Mini-Drones (MDs) based on aerial computing services where HAU provides computation offloading services for MDs, while MDs can serve as edge computing server that can be equipped with appropriate capabilities to provide computing services for User Equipments (UEs) on demand. This study focuses on the computation offloading services provided by HAU to MDs, where the MD offloads all or a part of the task to the HAU, and the remaining of the task can be executed by MD. The proposed AAGN framework aims to reduce the MDs' energy consumption and minimize the processing delay by optimizing HAU mobility, MDs scheduling, flight speed, flight angle, and tasks offloading, equipping HAU with the required computing resources. We investigate the computation offloading problem using Deep Deterministic Policy Gradient (DDPG) as a computing offloading approach to learning the optimal offloading policy from a dynamic AAGN environment, considering this problem as a non-convex problem. The simulation results show the feasibility and effectiveness of the proposed AAGN environment where DDPG algorithm can achieve an optimal decision offloading policy and obtains a critical optimization in delay and task offloading ratio compared with Deep Q Network (DQN) and Actor-Critic (AC) algorithms.

ABSTRACT. In response to the vast and ever-changing task demands of vehicle terminals, the edge-assistant vehicular network (EAVN) supported by the mobile computation offloading (MCO) technic constituted a new paradigm for improving system performance. The existing edge resource trading mechanisms in EAVN were all centralized processing and suffered from several critical drawbacks of the centralized systems, which inspired the research design of distributed trading mechanisms. In this paper, we proposed an efficient distributed reverse combinatorial auction-based trading mechanism under the anti-manipulation check, namely DRCA, to solve the joint multi-task offloading and multi-resource allocation problem in EAVN with overlapping areas, and prevent the participants from manipulating the auction results. We proved that DRCA has achieved the property of faithfulness and analyzed its network complexity. Besides, compared with existing auction-based mechanisms, DRCA could achieve suboptimal social welfare with relatively low system overhead.

ABSTRACT. Mobile-edge computing (MEC) has become a popular research topic from both academia and industry since it can alleviate the computation and power limitations of mobile devices by offloading computation-intensive and energy-consuming tasks from mobile users to nearby edge servers for remote execution. Existing papers have studied related problems, however, none of them considers the reliability of MEC systems that may suffer soft errors during execution and bit errors during offloading. In this work, we study the task scheduling problem targeting to maximize the quality of experience (QoE) of multi-user MEC systems under a certain reliability requirement. Due to the combinatorial nature of this problem, solving for optimal solution is difficult and impractical for large-scale MEC systems. To overcome this drawback, we propose to decompose the original problem into i) a task offloading optimization problem, ii) a task-to-server assignment problem for ensuring system reliability constraint, and iii) a computing resource allocation problem for maximizing system QoE. To address these sub-problems, we first obtain the optimal offloading decision using the discrete particle swarm optimization method. We then propose a reliability-optimality analysis-based task assignment heuristic and a utility-optimal resource allocation algorithm. Simulation results show that our scheme outperforms two state-of-the-art approaches and two baseline methods. The average improvement on QoE (quantified by offloading utility) achieved by our scheme is up to 63.2% under reliability requirement.

ABSTRACT. Service scenarios under edge-cloud collaboration are becoming more diverse in terms of service performance requirements. For example, smart grids require both intelligent control and long-term optimization, which poses considerable challenges for service providers to meet quality of service (QoS). However, current pioneering work has not yet explored both system utility and QoS guarantees. Therefore, this paper investigates the optimization problem of edge-cloud collaborative scheduling for QoS guarantees. First, we model the edge-cloud collaborative scheduling scenario and derive two sub-problems such as service deployment and request dispatch. Second, we design a near-optimal scheduling algorithm based on a submodular function optimization approach with the objective of maximizing the number of requests that are processed within the edge-cloud cluster under QoS constraints. Finally, our experiments verify the beneficial effects of the proposed algorithm in terms of throughput rate, scheduling time cost, and resource utilization.

ABSTRACT. Malware detection has become a hot research pot as the development of Internet of Things and edge computing have grown in popularity. Specifically, various malware exploits firmware vulnerabilities on hardware platform, resulting in significant financial losses for both IoT users and edge platform providers. In this paper, we propose CodeDiff, a fresh approach for malware vulnerability detection on IoT and edge computing platforms based on the binary file similarity detection. CodeDiff is an unsupervised learning method that employs both semantic and structural information for binary diffing and does not require label data. Through the SkipGram with Negative Sampling, we generate the word vocabulary for instruction data. Then we use the Graph AutoEncoder to generate the representation matrix, which has the function structures, function characteristics and function features of the binary file.

ABSTRACT. Federated learning (FL) has been proposed and applied in edge computing scenarios. However, the complex edge environment of wireless networks, such as limited device computing resource and unstable signal, leads to increase communication overhead and reduce performance for federated learning. Therefore, we propose a hierarchical aggregation mechanism to improve federated learning performance in resource-constrained wireless edge environment. Firstly, we define a cost function to reflect the performance, and design a fuzzy K-means clustering mechanism including three feature model. Then we construct a optimization problem model for the process of hierarchical aggregation, and we design a cluster-based hierarchical federated learning algorithm (CluHFed) which consists of fuzzy clustering, asynchronous aggregation and topology reconstruction. At last, we make an experiment with Pytorch, and experimental results show that the proposed algorithm improves the accuracy by 2.6%-35.8%, and reduces communication costs by 300-500 rounds compared with other popular federated learning aggregation algorithms.

ABSTRACT. In this paper, we study the joint optimization of task allocation and charging scheduling of mobile charging vehicles (MCVs) for parked-vehicle-assisted edge computing networks. In the proposed model, a group of electric vehicles (EVs) that have been parked for a long time must be recharged to their expected energy level within a specified time frame. Meanwhile, an optimal set of parked vehicles (PVs) is selected to compute a machine learning task utilizing their hardware resources and local data while satisfying the task's training performance requirements. Within the calculated time window, an MCV is dispatched to provide power replenishment to the PVs. By jointly deciding the task allocation and MCV charging sequence, the proposed model seeks to minimize the total energy consumption of the parked vehicular network, which includes the PV computation and MCV traveling consumption, subject to the PVs' expected energy level, task target utility and time window. To address this joint optimization problem, a marginal-product-based algorithm is designed, where a deep reinforcement learning method is integrated to solve the MCV scheduling problem. Simulation results demonstrate that the proposed method can efficiently solve the problem and outperform the compared algorithms in terms of energy consumption.

ABSTRACT. Mobile cloud gaming（MCG） can provide users with high-quality gaming services anytime, anywhere, but suffers from long network latency and huge wide-area traffic. In order to reduce latency and provide high quality of ser-vice, mobile edge computing (MEC) is envisioned as a promising approach to enable relevant computing at the edge. Since the quality of experience（QoE） of the game requires high frame rates and low network latency, the placement of service entities can affect the performance of MEC-enabled MCG. In addition, users have a high degree of mobility while enjoying MCG, so service migration is proposed to reduce QoE impairment, and service mi-gration means an increase in system cost. To address these challenges, we investigate the service placement of MEC-enabled MCG. Considering the dynamics of the system, we propose to minimize the QoE impairment ac-cording to the constraint cost of migration. We design the ECP algorithm to solve the problem, and theoretical analysis shows that ECP achieves better performance and outperforms other algorithms.

ABSTRACT. The direct position determination (DPD) method is more accurate than the previous two-step method in passive positioning. Enormous computational complexity in DPD is a severe drawback, which causes both real-time and high-accuracy to be challenging to satisfy. We integrate the formulas of the DPD and the image regression technique in computer vision, offering the unified computational graphs of both to investigate the fundamental reason for large time consumption in DPD. To achieve efficient DPD, we propose a fast DPD with an embedded convolutional neural network (CNNDPD), which is an end-to-end passive positioning network. We use a wavelet transform two-dimensionalization to convert the time domain signal into a time-frequency map and extract the time-frequency attributes effectively for the received data required for positioning. Other information required for positioning is stitched with the findings of time-frequency map processing and sent into fully-connected networks, allowing fuse with time-frequency information effectively. The simulation results show that the CNNDPD has the advantage of fast and highly accurate positioning. In a wide-area localization setting, CNNDPD has 26 times and 46 times faster inference speed than exhaustive search DPD and genetic algorithm DPD, respectively, without reducing accuracy. Furthermore, CNNDPD has a lower false alarm rate than the two benchmarks.

ABSTRACT. Traditional approaches focus on an individual item of most interest to users. However, in most realistic scenarios, the platform needs to recommend a group of items at one time for users' convenience, called bundle recommendation. e.g., a music playlist containing multiple songs. The existing bundle recommendations usually use manual methods to artificially build bundles for different items, ignoring the obtained bundles and the potential relationships among the items in the bundle, especially the relationships between bundles. Therefore, how integrating multiple complex interactions into bundles and obtaining high-quality bundle recommendation is an important problem. To solve the problem, we propose a novel model named IMBR (short for Interactive Multi-Relation Bundle Recommendation with Graph Neural Network). Firstly, we construct a multi-relation interaction graph to capture the interaction relation from the user view. At the same time, we get bundle subordination relation from the item view. They can obtain richer representations of users, bundles, and items. Secondly, we design a bundle frequent term constraint algorithm (BFTC) to constrain the composition of items in a bundle and pay attention to the similarity between bundles. Finally, we leverage a multi-task learning framework to capture user personalized preferences to improve the performance of bundle recommendation. Extensive experiments on two real-world datasets with different scales show that our method can significantly outperform various baseline approaches.

ABSTRACT. The emergence of vertical federated learning (VFL) solves the problem of joint modeling between participants sharing the same ID space and different feature spaces. Privacy-preserving (PP) VFL is challenging because complete sets of labels and features are not owned by the same entity, and more frequent and direct interactions are required between participants. The existing VFL PP schemes are often limited by the communication cost, the model types supported, and the number of participants. We propose FLFHNN, a novel PP framework for heterogeneous neural networks based on CKKS fully homomorphic encryption (FHE). Combining the advantages of FHE in supporting types of ciphertext calculation, FLFHNN eliminates the limitation that the algorithm only supports limited generalized linear models and realizes ``short link'' communication between participants, and adopts the training and inference of encrypted state to ensure confidentiality of the shared information while solving the problem of potential leakage from the aggregated values of federated learning. In addition, FLFHNN supports flexible expansion to multi-party scenarios, and its algorithm adapts according to the number of participants. Our analysis and experiments demonstrate that compared with Paillier based scheme, FLFHNN significantly reduces the communication cost of the system on the premise of retaining the accuracy of the model, and the required interactions and information transmission for training are reduced by almost 2/3 and more than 30% respectively, which is suited to large-scale internet of things scenarios.

ABSTRACT. With the ubiquity of electronic communication devices, detecting the information sources is a critical task in reducing the damage caused by malicious sources. However, in the contemporary research of sources identifications and information propagation identifications are calculated through social network topology structure or mathematics inference. In this paper, we borrow the training tool of neural network and propose a deep convolutional neural network to identify the sources in social networks. Initially, we utilize the 20% of data set to play the role of training set and substitute into the proposed model. Subsequently, we employ a bi-graph to classify the trained sources into truth or rumor vertexes. Finally, we utilize our proposed model to test 80% of data set as evaluation results of our identification mechanism. From the experimental results, our developed method can identify more than 85% of information sources and the classification accuracy can reach 80% in both test and train process. The obtained results further indicate that our model can effectively and accurately identify the information sources with reasonable computation costs.

ABSTRACT. In many scenarios, people have a demand for deploying the artificial intelligence applications on the edge device of IoT. For some special applications, these embedded devices are always required real-time reponse; hence, it is necessary to process machine learning algorithms on microprocessors. However, these devices may be subjected to side-channel attacks(SCA). During the execution, these devices will generate the leakage information can be captured to get the secret data. In this work, we investigate how to reverse engineer the weights of a convolutional neural network(CNN) which is deployed on ARM Cortex-M3 using Chosen Pixel Horizontal Power Analysis(CP-HPA).

We conduct the experiment on ELMO emulating leaks for the ARM Cortex-M3. ARM Cortex-M3 microprocessors are often used to deploy CNNs. Here, we show that it is possible to recover the weights of a CNN using CP-HPA assuming that the adversary only has the knowledge of the architectures. We increase the accuracy of our attack through setting up chosen input pixel to correlate the selected multiplication. We are able to successfully recover the weights of a CMSIS-NN implementation CNN, and accuracy of our attack is 84.625\%.

ABSTRACT. The rapid development of the Internet has brought great changes and convenience to the society and people. With the development of the Internet, its security has been paid more and more attention. Intrusion detection can detect network attacks in real time and respond to them in time, which has become an essential and im-portant security line. With the novel of network attack and the diversification of network traffic, traditional intrusion detection based on attack load matching and the intrusion detection based on machine learning has problems of inaccurate fea-ture extraction and insufficient detection effect. To solve the above problems, this paper designs a hybrid neural network DCT-IDS model, using dense convolution neural network to achieve traffic feature fusion, reducing the number of parame-ters, using Transformer to extract time sequence features, and experimental tests were carried out on the latest dataset CIC-IDS2018. The experimental results show that the accuracy of the proposed DCT-IDS model reaches 98%, and all the indexes are better than the existing excellent models.

ABSTRACT. Blockchain, as an emerging technology, has vulnerabilities that make the blockchain ecosystem rife with many criminal activities. For example, phishing in Ethereum has seriously threatened the security of users’ property and hindered the development of the blockchain ecosystem. However, existing technologies of phishing fraud detection heavily rely on shallow machine learning, leading to low detection precision. To solve this problem, in this paper, we construct a graph classification network model TransDetectionNet. Particularly, we propose a node embedding algorithm named Edge-sampling To Node Vector(Esmp2NVec) that can effectively extract the features hiding in the directed transaction network. Then, we use graph convolutional neural networks (GraphSAGE and GCN) to learn the topological space structure between nodes for further extraction of node features, where the nodes represent Ethereum accounts. To evaluate the method, a series of transaction data from the real Ethereum system is leveraged to train the graph classification model, and several experiments are designed to measure the phishing accounts identification performance comparison between our method and the other related works. The final results of those experiments show that our method can effectively detect phishing accounts from the Ethereum system.

ABSTRACT. This paper designs a rateless low density parity check (LDPC) code for the information transmission of the WiFi backscatter communications. Because WiFi has the characteristics of burst data packages and low anti-jamming capability, the reliability becomes a problem. Therefore, the encoding is significantly crucial when using WiFi signals as the excitation in backscatter communications. Rateless LDPC code can be applied to not only solve these two shortcomings, but also adjust the link state and the bit rate without knowing the channel state information. It ensures that transmission resources are not wasted and the computational resources are saved. We conduct simulation experiments and the obtained results show that the rateless LDPC still performs well under the restriction of the number of retransmissions. Furthermore, the proposed scheme works against the intermittent nature of WiFi excitation signals.

ABSTRACT. Backscatter communications enable the connection of the large scale of the Internet of things (IoT) devices, due to their extremely low power consumption characteristic. As the number of IoT devices is increasing, the effective and reliable communication between devices becomes a key factor to offer services with the desired quality by the IoT. However, due to the impact of noise and the low power of the backscatter signal itself, the system performance is usually unreliable. To this end, in this paper, we propose an integrated cyclic redundancy check code and Polar codes (CRC-Polar) to improve the performance of the ambient backscatter communications. The performance is verified indicating by the bit error rate from the following aspects: excitation source time intervals, excitation source signal-to-noise ratios, coding rates and code lengths. We conduct extensive computer simulations using Matlab platform to verify that the designed method achieves a better enhancement of the excitation source signal transmission process. The experimental results show that our proposed CRC-Polar scheme can effectively improve the communication reliability of backscatter communication with medium and long distances and effectively reduce the influence of environmental factors on the communication quality.

ABSTRACT. Fine Time Measurement (FTM) protocol is included by IEEE 802.11-2016 to address the challenging problem of the high accuracy of the existing system in Wi-Fi positioning. Although FTM promises meterlevel ranging accuracy in line-of-sight (LOS) conditions, non-line-of-sight (NLOS) and multipath effects cause accuracy to decline sharply. In this paper, by diving into fine-grained PHY layer information of higher time resolution, we explore the relationship deeply between FTM error and multipath channel response. On this basis, we propose FSI, a method for calibrating FTM errors using PHY layer information, which can identify environmental characteristics automatically and estimate the length of signal propagation paths. Finally, we design an optimation method based on the mobility of users, to further improve positioning accuracy in actual environments. Experimental results show that FSI improves the ranging accuracy by 24.80% and positioning accuracy by 28.45%.

ABSTRACT. RFID as an important component technology of IoT is rapidly applied in recent years. But it also faces severe security risks like malicious intrusion, most operated by unauthorized reader (UR). There are some researches proposed the unauthorized reader detection algorithms based on commercial off-the-shelf (COTS) devices, but these detection algorithms are often easily affected by moving objects blocking interference, causing false alarms. We adopt a new parameter adjacent signals time interval (ASTI) to improve the UR detection algorithm by reducing the time-delay and propose a new method of eliminating moving object interference, which can reduce the system false alarm rate to less than 7.9% by experimental testing.

ABSTRACT. Radio Frequency Identification (RFID) technology plays an essential role in surveillance scenarios. However, redundant data hinders the efficient processing of data. The redundant processing of RFID data is of great importance to reduce the load of the RFID system and quickly detect the monitored tags. To address the issue, the research community introduced Bloom filtering technology into the RFID system. However, existing methods often use fixed thresholds and cannot adapt to complex environmental conditions. This work presents R-TDBF, a practical solution that enables data redundancy filtering in complex environments by rationally setting filtering thresholds. In addition, a signal strength threshold is also introduced in R-TDBF, which reduces the error caused by signal fluctuation. The experimental results show that the R-TDBF algorithm can filter redundant data well under different threshold conditions. Compared with the existing algorithms, our method has good practicality with an average reduction of 73.7% in the detection error rate.

ABSTRACT. In this paper, we consider a reconfigurable intelligent surface (RIS) assisted rate splitting multiple access (RSMA) transmission system with estimated channel state information (CSI). The RIS is used to artificially construct the transmission environment to achieve more energy efficient transmission. The energy efficiency maximization problem is formulated by satisfying the constraint of power budget, the design principles of RSMA and RIS. To solve this problem, fractional programming is first used to decouple the single-ratio objective function. Then the optimal power allocation coefficients and the phase shift matrix of RIS are obtained by the proposed alternative optimization method, respectively. Numerical simulation results demonstrate that the energy efficiency performance of the RIS-assisted RSMA system can be significantly improved by the iterative joint optimization.

ABSTRACT. Low power and lossy network (LLN) massive terminal deployment has become an inevitable trend. However, traditional routing protocols cannot meet the large-scale data transmission requirements. In this paper, we introduce the multi-channel communication technology into LLN and propose a multi-channel routing protocol based on cross-layer design (MC-RPL), which can increase the data transmission capacity of the network via a parallel data transmission strategy. Specifically, we design a novel super-frame structure to decouple the communication period into a route maintenance phase and a data transmission phase. Nodes can transmit data in parallel during the data transmission phase. Besides, we improve the trickle algorithm to enhance routing maintenance efficiency during the route maintenance phase. Simulation results have demonstrated the effectiveness of the MC-RPL protocol compared to the MRHOF and IRH-OF protocols.

ABSTRACT. In recent years, the demand for urban travel is increasing and the travel modes are diverse. Online car Hailing has become an important way to meet the travel needs of residents. The online car-hailing platform receives tens of thousands of travel requests every day. However, a large portion of the thousands of orders are un-finished, that is, canceled by passengers. This not only reduces the income of drivers but also affects the order dispatching efficiency of the online car-hailing platform. To predict the cancellation probability of online car-hailing or-ders(OCP), the relationship between multi-source heterogeneous data and OCP is first introduced, in which the presence of idle taxis is the main factor for passen-gers to cancel their orders during the waiting period. Secondly, a deep learning model based on the Seq2Seq structure is designed to predict OCP in real-time. The model consists of an attribute fusion module, encoder layer, and decoder lay-er. Finally, a full experiment is carried out using the Didi Chengdu online car-hailing order data set to verify the effectiveness of the algorithm.

ABSTRACT. Autonomous vehicles sense the surrounding environment through various sensors and make behavior decisions based on real-time perception information to change their vehicle's motion state. Most existing studies on behavior use single data, high computational complexity, and single optimization criteria only, which lacks practicality. This work proposes an autonomous vehicle motion behavior decision method. It first extracts the corresponding features according to correlation among adjacent vehicles and predicts driving behavior and trajectory of adjacent vehicles. Then, it abstracts driving states of autonomous vehicles, introduces their state transition process based on a definite state machine, and gives a behavior decision method. Finally, a multi-objective optimization algorithm is used to optimize. Extensive simulation results show that this method can effectively improve the safety, efficiency, and practicability of autonomous vehicle motion behavior decision.

ABSTRACT. The most important thing for Connected and Automated Vehicles (CAVs) is to ensure driving safety and prevent the loss of life and property due to danger. The existence of vehicle blind spots can lead to incomplete or ineffective access to information, which will bring risks. At the same time, the transmission of a large amount of duplicate data will lead to information redundancy and bandwidth waste. In this paper, we design BP-CODS, which uses blind-spot prediction assistance to schedule image data between vehicles with the support of the Edge Server. We model the data scheduling transmission as two processes of uploading and downloading, form the set coverage problem, and propose a heuristic algorithm to solve it. We conduct extensive simulation experiments in CARLA to verify the effectiveness of BP-CODS in reducing a large number of redundant data.

ABSTRACT. Human mobility data play an important role in addressing various urban issues. However, when a new mobility paradigm emerges and continuously evolves with time, it is usually hard to obtain a large-scale and evolving mobility dataset due to various factors such as social and privacy concerns. In this paper, we focus on modeling the evolving mobility of metropolitan-scale electric taxis (ETs), which have different mobility patterns with petroleum vehicles and continuously evolve with the expansion of the ET fleet and the charging station network. To this end, the E2M system is proposed to generate trajectories for large-scale ET fleets by learning the mobility from only a small-scale ET fleet and the corresponding charging station network. First, the ET mobility is decomposed and modeled with transition, charging, and resting patterns. Second, the E2M system generates trajectories with a fleet generation algorithm. Extensive experiments are conducted on a real-world dataset, which has ET trajectories during both the early stage and mature stage in the taxi electrification process in Shenzhen, China, and the results verify the effectiveness of E2M.

ABSTRACT. The high-definition (HD) map is the foundation for autonomous driving, which has a huge data volume and needs to be updated frequently. To ensure low download latency, HD map contents are usually pre-cached at roadside units (RSU) or vehicles. However, the HD map contains a lot of dynamic data, and maintaining its freshness is crucial for ensuring driving safety, which is ignored by the existing HD map caching methods. In this paper, we propose a freshness-aware HD map caching method to minimize both download latency and loss of freshness. First, we introduce a cost function to incorporate both the download latency and the loss of freshness. Next, we formulate the HD map caching problem as an optimization problem to minimize the total cost. To reduce computation complexity, we decompose the original problem into two subproblems. Consequently, we propose a freshness-aware vehicle request algorithm to optimize vehicle request decisions and then leverage a distributed multi-agent multi-armed bandit (MAMAB) algorithm to make optimal caching decisions. Finally, simulation results verify that the proposed freshness-aware HD map caching method outperforms other baseline methods.

ABSTRACT. As an alternative to traditional vehicles, electric vehicles (EVs) have significantly increased their market share in recent years. However, the limited battery capacity of EVs may become a bottleneck in their development. Mobile charging vehicles (MCVs), as emerging charging devices, can provide a more portable charging mode. The MCV deployment and charging strategy plays a decisive role in the effective operation of the whole Internet of Electric Vehicles (IoEV). In this paper, we investigate the joint MCV deployment and charging schedule (JMDCS) in one-to-many mode. An integer linear programming problem is formulated to minimize the completion time, consisting of the deployment time of the MCVs and the time of the charging schedule. Since this problem is NP-hard, an approximate algorithm is proposed, where a special rounding technique is employed to assign jobs to the plugs. The simulation results show that when the number of EVs is small, the proposed algorithm is close to the optimal algorithm. When the number of EVs is large, the proposed algorithm outperforms its counterparts in scheduling performance and shows superiority over the optimal algorithm in scheduling efficiency.

ABSTRACT. Partition-based caching is emerging as an appealing solution to improve the performance of the content caching by increasing the content diversity at the network edge. In this paper, we model and analyze a vehicular network where the vehicles can obtain the requested contents from the roadside units (RSUs) by adopting the random linear network coding in the partition-based caching scheme. Specifically, the geographic distribution of the roads and RSUs are modeled by the stochastic geometry tools. The required content can be obtained from the multiple nearest RSUs and the content can be decoded by using the successive interference cancellation approach. We derive the distance distribution between the typical vehicle and the nearest RSUs, and obtain the analytical expression of the successful transmission probability of the content caching. The numerical simulations verify the analytical results and provide the guidelines for the application of the partition-based caching in vehicular networks.

ABSTRACT. Due to high fairness and allocation efficiency, the task assignment problem of mobile applications via auctions has become a promising approach to motivate bidders to provide their mobile device resources effectively. However, most of existing works focus on the auction mechanism under the plaintexts, and ignore the problems caused by information leakage. In this paper, we study the problem of the privacy-preserving auction for task assignment in outsourced cloud environments without leaking any private information to anyone. Specifically, we use Yao's garbled circuits and homomorphic encryption system as underlying tools. Along with several elaborately designed secure arithmetic subroutines, we propose a privacy-preserving and truthful auction framework for task assignment in outsourced cloud environments. Theoretically, we analyze the complexity of our scheme in detail and prove the security in the presence of semi-honest adversaries. Finally, we evaluate the performance and feasibility of our scheme through a large number of simulation experiments.

ABSTRACT. A smart grid can dynamically adjust the amount of electricity supply by smart meters' personalized needs, reducing energy waste and protecting the environment. However, because the uploaded data could reveal users' sensitive information, and internal adversaries could poison the power statistics, privacy preservation and security supervision in smart grid systems need to be concerned. To solve these, we propose a traceable ring signature scheme based on SM2 with strong security and anonymity, in addition to utilizing this scheme to build a four-layer smart grid model, separating the duty of statistics and regulations. Specifically, the scheme integrates the advantages of a key-insulated linkable ring signature (LRS) for Monero and an SM2-based ring signature: a key derivation mechanism to make the key more secure and a simple SM2-based ring structure. A trapdoor has been introduced in the ``key image'' of the signature, which is often used in LRS for linkability, but in our signature, it's used for traceability. This allows authorized participants to open signatures and reveal the identity of the real signer when exceptions occur. Besides the security and privacy analyses, we also implement the proposed scheme and give serval experiments to evaluate the time and space performance. The results show that our new scheme with space of kilobyte level size and time of linear or constant cost can be effectively adapted to the functional requirement of our smart grid model. In addition, the signature can be ported to wireless mobile devices for privacy protection and security management.

ABSTRACT. Mobile crowdsensing (MCS) is a new paradigm for data collection, data mining and intelligent decision-making using large-scale mobile devices. The efficient task allocation method is the key to the high performance of MCS. The traditional greedy algorithm or ant algorithm assumes that workers and tasks are fixed, which is not suitable for the situation where the location and quantity of workers and tasks change dynamically. Moreover, the existing task allocation methods usually collect the information of workers and tasks by the central server for decision-making, which is easy to lead to leakage of workers' privacy. In this paper, we propose a task allocation method with privacy protection using deep reinforcement learning (DRL). Firstly, the task allocation is modeled as a dynamic programming problem of multi-objective optimization, which aims to maximize the benefits of workers and platform. Secondly, we use DRL for training and learning model parameters. Finally, the local differential privacy method is used to add random noise to the sensitive information, and the central server trains the whole model to obtain the optimal allocation strategy. The experimental results on the simulated data set show that compared with the traditional methods and other DRL based methods, our proposed method has significantly improved in different evaluation metrics, and can protect the privacy of workers.

ABSTRACT. With the widespread deployment of IoT (Internet of Things) devices, WiFi fingerprint-based localization is becoming one of the most promising techniques for indoor localization. A client is able to obtain its location by providing its measured fingerprint (vector of WiFi signal strengths) to the service provider who maps the fingerprint against the database and returns the result back to the client. However, traditional applications of WiFi fingerprint-based localization may disclose the client’s location privacy and often incur high consumption of communication and computing resources. In this paper, we focus on implementing a privacy-preserving framework with high efficiency and accuracy for WiFi fingerprint-based localization. Firstly, to reduce computational overhead at the server side, we introduce a clustering algorithm called k-means++ in offline phase. Besides, we explore the correlation of the fingerprint and propose a Pearson correlation based distance computation method, which achieves better accuracy than traditional Euclidean distance. Finally, we secure the overall computation by adapting a series of secure multi-party computing primitives. Theoretical analysis is carried out to prove the security of our scheme. Experiments on real-world datasets indicate that our scheme achieves better practicality and efficiency compared with existing methods. Compared to existing work PriWFL and PPWFL, our scheme reduces the average distance error by approximately 4.5% and 2.9% under a query time of less than 0.2s.

ABSTRACT. Electronic voting (e-voting) is widely used because of its convenience and efficiency. In response to the security problems in e-voting, such as legality of voters, privacy disclosure, etc., this paper proposes a novel e-voting scheme that can check the format of ballots without disclosing the content based on homomorphic encryption. Firstly, voters encrypt their ballots with the Paillier encryption before sending them to the counter. Then, the counter decomposes the encrypted ballots using the proposed n-ary conversion protocol, and performs the format check of the ballots. Only ballots with correct format are counted. During the whole process of voting, no one except the voter himself can know each ballot content, even the counter, so that the privacy of ballots is preserved. Finally, the counter performs an additive homomorphism operation on the encrypted ballots and the voting manager decrypts it to tally the result. Besides the requirements including legality, privacy, integrity, we furtherly consider the validity of the ballots in e-voting and make the scheme more practical than the existing methods.

ABSTRACT. Differential privacy has become a golden standard for designing privacy-preserving randomized algorithms. However, such algorithms are subtle to design, as many of them are found to have incorrect privacy claim. To help identify this problem, one approach is designing disprovers to search for counterexamples that demonstrate high violation of claimed privacy level. In this paper, we present DP-Opt, a disprover that tries to search for counterexamples whose lower bounds on differential privacy exceed the claimed level of privacy guaranteed by the algorithm. We leverage the insights of counterexample construction proposed by the latest work, meanwhile resolve the limitations of their work. We transform the search task into an improved optimization objective which takes into account the empirical error, then solve it with various off-the-shelf optimizers. An evaluation on a variety of both correct and incorrect algorithms illustrates that DP-Opt almost always produces stronger guarantees than the latest work up to a factor of 9.42, with runtime reduced by an average of 19.2%.

ABSTRACT. The vehicle re-identification (V-ReID) task is critical in urban surveillance and can be used for a variety of purposes. We propose a novel augmentation method to improve the V-ReID performance. Our deep learning framework mainly consists of a local rotation transformation and a target selection module. In particular, we begin by using a random selection method to locate a local region of interest in an image sample. Then, a parameter generator network is in charge of generating parameters for further image rotation transformation. Finally, a target selection module is used to retrieve the augmented image sample and update the parameter generator network. Our method is effective on VeRi-776 and VehicleID datasets, it shows that we achieve considerable competitive results with the current state-of-the-art.

ABSTRACT. In Underwater Sensor Networks (UWSNs), the location information of sensor nodes is essential for making the measured data meaningful. However, UWSNs have a complex node deployment environment. Node mobility caused by ocean currents and other factors would lead to a bigger ranging error and make some nodes cannot receive enough data packets. In this paper, a Localization algorithm based on a Single Mobile Beacon (LSMB) is proposed. LSMB makes use of the attenuation law of signal strength and the geometric relationship between a sensor node and the path of the mobile beacon, reducing the impact of random error on distance measurement. On this basis, by analyzing the overall movement trends of sensor nodes, this paper analyzes and studies the counter-current movement and downstream movement of the mobile beacon respectively, so as to make LSMB suitable for dynamic marine environment. The simulation shows that the algorithm reduces the impact of node mobility on localization and has small average localization error.

ABSTRACT. Reconfigurable intelligent surface (RIS) is a promising technique in the 6G communication system, which effectively improves the wireless propagation channel. Moreover, the RIS also benefits the localization performance since it helps avoid the non-line-of-sight channel when there are obstacles. In this paper, we mainly analyze the 3D localization performance of the millimeter-wave (mmWave) system with a given fixed RIS. The Cram$\acute{\mathrm{e}}$r lower bound (CRLB) is derived with our proposed 3D RIS-based wireless propagation channel. We analyze the localization accuracy of time-of-arrival (TOA) and angle-of-arrival (AOA). The results indicate that the RIS-based localization method can significantly improve localization accuracy, and centimeter-level localization can be attained. In addition, the localization based on TOA outperforms that based on the AOA when the number of the Rx and RIS units is fixed. \keywords{{Reconfigurable intelligent surface (RIS), millimeter-wave, wireless localization, Cramer-Rao lower bound (CRLB)

ABSTRACT. Dense areas of pedestrians in complex crowded scenes tend to disrupt the proper path of the agents. The agents usually avoid gathering areas to find a reasonable pedestrian-sparse path, slow down the speed to walk, and wait for the gathering pedestrians to disperse. The accurate trajectory prediction in gathering areas is a challenging problem. This work introduces a new feature that affects trajectories to address this problem. The area gathering feature that allows agents to plan future paths based on the gathering level of pedestrians. The gathering areas as well as indicate the degree of gathering in the areas by means of a dynamic pedestrian filtering method to generate a trajectory heat map. Besides, the convolutional neural network is used to extract the corresponding area gathering feature. Furthermore, a new approach is proposed for inter-agent interactions that makes full excavation of deep interaction information and takes into account a more comprehensive interaction behavior. This work predicts trajectories by incorporating multiple fac-tors such as area-dense features, social interactions, scene context, and individual intent. The prediction accuracy is significantly enhanced and outperforms state-of-the-art methods.

ABSTRACT. The ride-hailing app must provide users with appropriate pick-up points when they submit their travel demands and their locations are recognized, efficiently reducing users' operation complexity and optimizing the software performance. Most apps currently try to search for locations near users' current GPS locations as the Points of Interest (POIs), which is an efficient method of locating, but seriously ignores personal preferences. In this paper, we deeply analyze the historical ride-hailing orders of users on Didi Chuxing platform. We explore the given dataset, get the general regularity of users' commuting, and propose a Pick-Up Points Recommendation Model (PPRM) based on the clustering algorithm. We cluster users' historical orders using Density-Based Spatial Clustering of Applications with Noise (DBSCAN) according to orders' spatial information. In this way, the candidate outputs closest to the user's current environment/feature can be found in a specific category. The linear addition of the candidate outputs severs as the final pick-up point provided. Therefore, our model can offer recommendations of the best pick-up points. In addition, experimental results based on real-world datasets indicate that our model can efficiently and accurately provide users with optimal points.

ABSTRACT. Instant delivery platforms, equipped with professional couriers to provide convenient delivery services, have emerged rapidly in many cities. For the benefit of platforms, many researchers focus more on maximizing overall efficiency but ignore individual fairness. Current fairness research in mobile systems mainly concentrates on one-sided or two-sided relationships, such as drivers and customers. However, instant delivery services have two new characteristics in fairness: (i) multi-stakeholder involvement, namely couriers, merchants and users should be considered comprehensively; (ii) more complicated matching relationship because of the concurrent dispatch mode, meaning one courier will handle multiple orders simultaneously. To handle this multi-sided fairness problem, our paper proposes a novel order dispatch system to balance the platform revenue and multi-stakeholder fairness. Motivated by the analysis of real-world datasets from Eleme(one of the largest instant delivery companies in China), we formulate the order dispatch problem as a sequential decision-making problem and incorporate multi-sided fairness into the decision criteria. Then, we design multi-sided fairness-aware deep reinforcement learning to solve large-scale decision problem, with the fairness relying on Least Misery Fairness definition for users and Variance Fairness definition for couriers and merchants. Finally, extensive experiments show the effectiveness of our model in balancing multi-sided fairness among stakeholders and long-term profits of the whole platform.

ABSTRACT. With the development of urban transportation networks, the flow of people in cities generally shows the characteristics of concentration, periodicity and irregularity, and a typical example is rush hour. For most existing taxi-hailing apps, users frequently queue up for a relatively long time during rush hour and may even fail to get orders taken due to various factors. To solve this problem, we propose a users' departure time prediction model based on Light Gradient Boosting Machine (TP-LightGBM), which will remind users to book taxis before their journeys. As we know, TP-LightGBM may be the first model for departure time prediction. We uncover that travel behavior patterns vary under different external conditions through statistics and analysis of users' historical orders from multiple perspectives. Furthermore, we extract multiple features from these orders and select the favorable features by calculating their information gain as the input of TP-LightGBM to predict users' departure time. Therefore, our model can provide users with the recommendations of the best departure time if they need them. The final experimental results on our datasets indicate that TP-LightGBM has more excellent performance with great stability in predicting user departure time than other baseline models.

ABSTRACT. Inertial navigation adopts localization base stations to correct cumulative errors for mine rescue robots, while requirements of explosion-proof safety hinder the application of regular powered base stations in harsh coal mine environments. Therefore, we propose MineTag, a novel localization base station for self-positioning of coal mine robots, which is built with low-cost and battery-free optical tags via a differ-neighbor deployment strategy. The main innovation of the tag is to modulate the light retro-reflection with a light absorption mechanism, allowing the tag to reflect a specific light intensity without the need for a power source. According to the topological relationship of tags, we propose a novel tag recognition algorithm based on trajectory matching to determine which tag the robot is under. Finally, we implemented MineTag and evaluated its performance in a real coal mine. Experimental results show that MineTag can achieve the tag recognition accuracy of more than 95%, and the localization accuracy is 98% error of 2.6 m or less.

ABSTRACT. With the development of the Internet of Things (IoT), more and more applications are increasingly demanding latency. Traditional single-task scheduling strategy is difficult to satisfy low-latency demand. This is due to the fact that the task scheduler usually schedules tasks to a closer server, which leads to an increase in task latency when there are more tasks, which in turn leads to an increase in task rejection rate. In this paper, we propose an end-edge cooperative multi-tasks scheduling (MTS) strategy based on improved particle swarm optimization (IPSO) algorithm. At first, we design a Software-Defined Networks controller algorithm to cluster task offload requests. Then, we set the scheduling priority for the multi-task clusters. At last, we minimize the total offloading cost of total tasks as the optimization goal to satisfy its delay. The results demonstrate that the strategy we proposed can effectively reduce the service cost of the system, and the processing delay of tasks, which improves the success rate of task processing.

ABSTRACT. With the growing activities of diverse Maritime Internet of Things (MIoT), mobile edge computing (MEC) becomes a promising paradigm to provision computation and storage for computationintensive tasks of marine users. Although the edge server (ES) deployment and service placement are important issues in the field of MEC, research on joint placement is often overlooked, particularly in the MIoT. In this paper, we propose the buoy-based ES deployment and service placement (BESDSP) problem for MIoT networks, aiming at maximizing the total profit while considering the location constraints of buoys, the different service request rates, the income and delay cost of service provided by ESs, as well as the characteristics of maritime channels. Then, we propose a heuristic approach, the genetic-BESDSP (G-BESDSP) algorithm, to solve the BESDSP problem. Simulation results demonstrate that the proposed G-BESDSP algorithm outperforms existing state of art solutions.

ABSTRACT. Federated Edge Learning (FEEL) is becoming a popular distributed privacy-preserving machine learning (ML) framework where multiple edge devices collaboratively train an ML model with the help of an edge server. However, FEEL usually suffers from a communication bottleneck due to the limited sharing wireless spectrum as well as the large size of training parameters. In this paper, we consider gradient quantization to reduce the communication traffic and aim at minimizing the total training latency. Since the per-round latency is determined by both the bandwidth allocation scheme and gradient quantization scheme (i.e., the quantization levels of edge devices), while the number of training rounds is affected by the latter, we propose a joint optimization of bandwidth allocation and gradient quantization. Based on the analysis of total training latency, we first formulate the joint optimization problem as nonlinear integer programming. To solve this problem, We then consider a variation of this problem where the per-round latency is fixed. Although this variation is proved to be NP-hard, we show that it can be transformed into a multiple-choice knapsack problem which can be solved efficiently by a pseudopolynomial time algorithm based on dynamic programming. We further propose a ternary search based algorithm to find a near-optimal per-round latency, so that the two algorithms together can yield a near-optimal solution to the joint optimization problem. The effectiveness of our proposed approach is validated through simulation experiments.

ABSTRACT. With the development of the sixth generation mobile network (6G), the arrival of the Internet of Everything (IoE) is accelerating. An edge computing network is an important network architecture to realize the IoE. Yet, allocating limited computing resources on the edge nodes is a significant challenge. This paper proposes a collaborative task scheduling framework for the computational resource allocation and task scheduling problems in edge computing. The framework focuses on bandwidth allocation to tasks and the designation of target servers. The problem is described as a Markov decision process (MDP). To minimize the task execution delay and user cost and improve the task success rate, we propose a Deep Reinforcement Learning (DRL) based method. In addition, we explore the problem of the hierarchical hash rate of servers in the network. The simulation results show that our proposed DRL-based task scheduling algorithm outperforms the baseline algorithms in terms of task success rate and system energy consumption. The hierarchical settings of the server's hash rate also show significant benefits in terms of improved task success rate and energy savings.

ABSTRACT. Vehicles request road condition information, traffic information, and various audio-visual entertainment frequently. Repeat Download will burden the core network and seriously affect the user experience. Edge caching is a promising technology that can effectively alleviate the pressure of repeatedly downloading content from the cloud. There are many existing edge cache scheduling methods, but they all have limitations. For example, traditional edge cache scheduling methods can not adapt to dynamic environment changes, and the centralized reinforcement learning algorithm is facing the problem of insufficient training data, while ordinary distributed learning is facing the problem of privacy disclosure. Therefore, this paper proposes an edge cache scheduling method based on the multi-head attention mechanism federal reinforcement learning (FRLMA). Firstly, the problem is modeled as a Markov decision model. The local models are trained through a deep reinforcement learning method. Finally, the federated reinforcement learning framework of edge Cooperative Cache is established. In particular, the multi-head attention mechanism is introduced to weigh the contribution of the local model to the global model from multiple angles. Simulation results show that the FRLMA method has better convergence and is superior to the most current popular methods in terms of hit rate and average delay.

ABSTRACT. With the rapid growth of edge intelligence, a higher level of deep neural network computing efficiency is required. Visual intelligence, as the core component of artificial intelligence, is particularly worth more exploration. As the cornerstone of modern visual modeling, convolutional neural networks (CNNs) have greatly developed in the past decades. Variants of light-weight CNNs have also been proposed to address the challenge of heavy computing in mobile settings. Though CNNs' spatial inductive biases allow them to learn representations with fewer parameters across different vision tasks, these models are spatially local. To acquire a next-level model performance, vision transformer (ViT) is now a viable alternative due to the potential of multi-head attention mechanism. In this work, we introduce EdgeViT, an accelerated deep visual modeling method that incorporates the benefits of CNNs and ViTs in a light-weight and edge-friendly manner. Our proposed method can achieve top-1 accuracy of 77.8\% using only 2.3 million parameters, 79.2\% using 5.6 million parameters on ImageNet-1k dataset. It can achieve mIoU up to 78.3 on PASCAL VOC segmentation while only using 3.1 million parameters which is only half of MobileViT parameter budget.

ABSTRACT. With the assistance of mobile edge computing (MEC), mobile devices (MDs) can optionally offload local computationally heave tasks to edge servers that are generally deployed at the edge of networks. As thus, the latency of task and energy consumption of MDs can be both reduced, significantly improv-ing mobile users’ quality of experience. Although considerable MEC sched-uling algorithms have been designed by researchers, most of them are trained to solve specific tasks, leaving the performance in other MEC envi-ronments remaining dubious. To address the issue, this paper first formu-lates the optimization problem to minimize both task delay and energy con-sumption, and then transforms it into Markov decision problem that is fur-ther solved by using the state-of-the-art multi-agent deep reinforcement learning method, i.e., MADDPG. Furthermore, aiming at improving the overall performance in various MEC environments, we integrate MADDPG with meta-learning and propose Meta-MADDPG which is carefully designed with dedicated reward functions. The evaluation results are given to show-case the more satisfactory performances of Meta-MADDPG over the state-of-the-art algorithms when confronting new environments.

ABSTRACT. In large-scale retrieval of mobile cloud, hash learning is favored by people owing to its fast speed. Nowadays, many hashing methods based on deep learning are proposed, because they have better performance than traditional feature representation methods. According to the latest research, by using the deep hashing method, the speed of mobile cloud retrieval can be significantly improved and the storage space of mobile data can be reduced. In this paper, we propose a cloud guided feature extraction method for mobile image retrieval, which is used to improve the speed of mobile cloud retrieval. In addition, we improve the traditional deep learning hashing method by dividing the multi-label images into "strong similarity" and "weak similarity". In this way, we propose a novel Deep Hash learning method based on Feature-Invariant representation (FIDH). Experiments on common single-label and multi-label data sets show that our method obtains better performance than state-of-the-art methods in large-scale image retrieval.

ABSTRACT. Hybrid beamforming (HBF) is a promising approach for balancing the hardware cost, training overhead and system performance in massive MIMO systems. Optimizing the HBF through deep learning (DL) has gained considerable attention in recent years due to its potential in dealing with the nonconvex problems. However, existing DL-based HBF methods require wider or deeper neural networks to guarantee training performance, which not only leads to higher complexity in training and deploying, but also increases the risk of over-fitting. In this paper, we propose a low-complexity HBF method based on convolutional neural network (CNN) to solve the spectral efficiency (SE) maximization problem with constant modulus constraint for the analog phase shifters over the transmit power budget in a multiple-input single-output (MISO) system. An unsupervised learning strategy is derived for the constructed CNN to learn to generate feasible beamforming solutions adaptively and thus avoiding any label data when training them. Simulations show its advantages in both SE and complexity over other related algorithms.

ABSTRACT. In this paper, we investigate device-free fall detection based on wireless channel state information(CSI). Here, we mainly propose a method that uses continuous wavelet transform (CWT) to generate images and then uses transform learning of convolutional networks for classification. In addition, we add a wavelet scattering network to auto- matically extract features and classify them using a long and short-term memory network (LSTM), which can increase the interpretability and re- duce the computational complexity of the system. After applying these methods to wireless sensing technology, both methods have a higher ac- curacy rate. The first method can cope with the problem of degraded sensing performance when the environment is not exactly the same, and the second method has more stable sensing performance.

ABSTRACT. Personalized federated learning (PFL) is an improved framework that can facilitate the handling of data heterogeneity by learning personalized models. As personalization performance directly depends on the global model, it is desired to acquire a global model with a decent generalization capability under data heterogeneity. This paper proposes a novel PFL scheme, FedALP, integrating the clustering method with an adaptive layer-based fusion algorithm. Experiments are performed using various neural network models on three standard datasets. Experimental results demonstrate that, compared with the FedAvg method, our scheme can significantly improve the local model's performance with a negligible decrease in the generalization capability of the global model. Furthermore, our scheme is customizable for specific PFL applications; hence it may provide a flexible strategy to effectuate a balanced performance for both the global and the local models.

ABSTRACT. Material recognition plays an essential role in areas including industry automation, medical applications, and smart homes. However, existing material recognition systems suffer from low accuracy, inconvenience (e.g., deliberate measuring procedures), or high cost (e.g., specialized instruments required). To tackle the above limitations, we propose a contact-free material recognition system using a millimetre wave (mmWave) radar. Our approach identifies materials such as metal, wood, and ceramic tile, according to their different electromagnetic and surface properties. Specifically, we leverage the following techniques to improve the system robustness and accuracy: (1) spatial information enhancement by exploiting multiple receiver antennas; (2)channel augmentation by applying Frequency Modulated Continuous Wave (FMCW) modulation; and (3) high classification accuracy enabled by Artificial Intelligence (AI) technology. We evaluate our system by applying it to classify five common materials. The experimental results are promising, with 98% classification accuracy, which shows the effectiveness of our mmWave-based material recognition system.

ABSTRACT. The insufficient number of tags is currently the biggest constraint on named entity recognition (NER)technology, with only a small number of  Registers（means the domain of language,which will be explained in Part I）currently having a corpus with sufficient tags. The linguistic features of  different Registers vary greatly, and thus a corpus with sufficient labels cannot be applied to NER in other Registers. In addition, most of the current NER models are more designed for large samples with sufficient labels, and these models do not work well in small samples with a small number of labels. To address the above problems, this paper proposes a model T_NER based on the idea of migration learning and multi-task learning, which learns the common features of language by using the idea of multi-tasking, and passes the model parameters of neurons with common features of language learned from multiple well-labelled source domains to the neurons in the target domain to achieve migration learning based on parameter sharing . In baseline experiments, T_NER's neurons outperformed the original models such as BiLSTM and BiGRU on a small-sample NER task; in formal experiments, the more the Registers in source domains, the better T_NER's recognition of the target domain. The experiments demonstrate that T_NER can achieve NER for small samples and across Registers.

ABSTRACT. Federated learning based on edge computing environment has great potential to facilitate the implementation of artificial intelligence at the edge of the network. However, because of the limited resource at the edge, place the complete Deep Neural Networks (DNN) model on the edge for training may not a good choice. In this paper, we study the time optimization for asynchronous federated learning based on model partition. That is, the DNN model is divided into two parts and deployed separately on the device and the edge server for the model training. First, we give the metric of the relationship between learning accuracy and iteration frequency, and then we build a mathematical model based on this. Because the solution space of mathematical model is too large to be solved directly, we propose an algorithm to minimize the total time by dynamically adjusting the model partition point and bandwidth allocation. Simulation results show that our algorithm can reduce the time by 32% to 60% compared with the other three methods.

ABSTRACT. In smart grid, smart meters record real-time customers' electricity consumption and transmit it to the control center. Leakage of customers' real-time power usage data will impose severe risk on consumers' privacy. To Protect users' privacy, researchers observe that what the control center really needs for providing services is not raw data but aggregate values of customers' real-time usage data. On this basis, many privacy-preserving data aggregation schemes are proposed for smart grid. However, these schemes pay little attention to the problem of collusion. In these schemes, edge nodes store the ciphertexts of customers' raw usage data, and the control center has the decryption key. If edge nodes collude with the control center, customers' usage data will still be leaked. To tackle this problem, a collusion-tolerant and privacy-preserving data aggregation scheme for smart grid is proposed in this paper. By leveraging homomorphic encryption and threshold scheme, the proposed scheme supports typical aggregate operations in smart grid, while keeping the privacy of customers and defending the collusions from edge nodes and the control center in practice. Security analysis and performance evaluation demonstrates that the proposed scheme protects the privacy of customers' usage data effectively even under collusions in practice.

ABSTRACT. In Unattended Wireless Sensor Networks (UWSNs), an itinerant sink does not establish a continue real-time channel with sensors. The sensed data needs to be temporarily stored in the off-line nodes. Due to the unattended nature of sensors, adversaries can easily compromise the sensors and tamper the data by physical method. Therefore, it is necessary to design a data authentication scheme to identify the tampered data and against physical attacks. Existing research mostly rely on a trusted long-term third party or cooperation mechanism, which makes the scheme difficult to be implementation. In this paper, we introduce Physically Unclonable Functions (PUF) to design a practical data authentication scheme for UWSNs. In the scheme, we first design an authentication and key agreement (AKA) protocol using challenge-response pairs (CRPs). After establishing a symmetric channel, we propose a PUF-based Message Authentication Code (MAC) scheme to ensure data security from the source. We also give a security analysis. The result shows the implementable and security of our scheme.

ABSTRACT. Wireless Sensor Networks (WSNs) are composed of many resource-limited nodes which may be laid in an unattended way. As a result, the sensing data the transmission mechanism are sensitive to attacks launched by adversaries. In this paper, we propose a novel Trust Secure Data Aggregation Model (TSDAM) with multiple attributes for WSNs, which regards different index-es of sensor nodes as the attributes to evaluate the trust of WSN in a com-prehensive manner. Firstly, we calculate the direct trust based on the data accuracy, the energy consumption and the forwarding behavior of nodes. Secondly, the indirect trust is evaluated according to the communication be-havior and the recommended credibility of neighbor nodes. Finally, the comprehensive trust is generated depending on various trusts, such as the di-rect and the indirect trust. Different from other mechanisms, TSDAM also selects the trust path according to the self-recommendation which is an at-tribute to indicate the willingness whether a node hope to participate in the communication process or not. The simulations show that TSDAM not only improves the reliability of the relay node, but also promotes the efficiency and accuracy of data aggregation.

ABSTRACT. Although the release and analysis of high-dimensional data bring tremendous value to people, it causes great hidden danger to participants’ privacy in the meantime. Various privacy protection methods based on differential privacy have been proposed at present. However, most of them cannot solve the problems of high computational overhead and privacy threats from untrusted servers caused by the curse of high dimensionality at the same time. Therefore, we propose a safer and more effective high-dimensional data release algorithm based on local differential privacy(LDP), which is referred to as PU_Bpub. This mechanism divides the high-dimensional data attribute set into multiple relatively independent low-dimensional attribute sets through spectral clustering after receiving the data protected by the user’s local privacy. The initial nodes are selected by setting a reasonable weight value, and then a weighted Bayesian network is established to synthesize a new dataset. It effectively preserves the dimensional correlation of the original high-dimensional data and reduces the communication overhead of synthetic data. Extensive experiments on real-world datasets demonstrate that our solution substantially outperforms the state-of-the-art techniques in terms of computational overhead and the synthetic dataset has high utility.

ABSTRACT. Federated Learning (FL) lately has shown much promise in improving sharing model and preserving data privacy. However, these existing methods are only of limited utility in the Internet of Things (IoT) scenarios, as they either heavily depend on high-quality labeled data or only perform well under idealized distribution conditions, which typically cannot be found in practical applications. In this work, we propose FedGAN, a Generative Adversarial Network (GAN) based federated learning method for semi-supervised image classification. In IoT scenarios, a big challenge is that decentralized data among multiple clients are normally non-independent and identically distributed (non-IID), leading to performance degradation. To address this issue, we further propose a dynamic aggregation mechanism that can adaptively adjust client weights in aggregation. Extensive experiments on three benchmarks demonstrate that FedGAN outperforms related federated semi-supervised learning methods, including a 55.36% accuracy on CIFAR-10 with 2k labels and 70.65% accuracy on SVHN with 1k labels - just 100 labels per class. Moreover, we carry out an extensive ablation and robust study to tease apart the experimental factors that are important to FedGAN's improvement.

ABSTRACT. Vehicular ad-hoc networks (VANETs) are core components of the cooperative intelligent transportation system (C-ITS), a new technology that improves road safety and driving experience. Vehicles communicate with each other to obtain traffic conditions on the current road segment by broadcasting authenticated safety messages using their digital certificates. Although this method protects the system against external threats, it is ineffective when faced with internal adversaries who possess legal certificates. Consequently, an increasing number of researchers have focused on intrusion detection (misbehavior detection) technology. VeReMi and its extension version are the only public misbehavior datasets of VANETs in its field, allowing researchers to compare their studies with those of others. We note that denial of service (DoS) attacks in these datasets are insufficiently comprehensive. As a result, we designed a more complete dataset than existing datasets by implementing multiple attacks, including different types of distributed denial of service (DDoS) attacks. We present the detection results of some machine learning algorithms on our proposed dataset. These results indicate that our dataset can be utilized as a reference for future studies to evaluate different detection methods.