Abstract:With the rapid development of e-commerce, the number of small and medium-sized express delivery has increased sharply. The last-mile delivery in the express industry faces many challenges, and a more efficient and environmentally friendly delivery mode is urgently needed. Small drone featuring high delivery efficiency, low cost, environmental protection, and flexible deployment attracts widespread attention from the logistics industry and academic community. Aiming at the innovative application and routing of drones in last-mile delivery, this study systematically summarizes the current four typical delivery modes. The modes include drone-independent delivery (one drone with multi-parcel delivery, drone relay docking and delivery), parallel drone-vehicle delivery (parallel drone-truck delivery and parallel drone-truck delivery with transit points), coordinated drone-vehicle delivery (drone take-off and landing delivery in the same truck, drone take-off and landing delivery in different trucks, drone-public transport coordinated delivery, pick-up and delivery with drones, and route truck-drone delivery), integrated drone-vehicle delivery (multi-mode integrated drone-truck delivery). Then, routing algorithms of the last-mile drone delivery are classified and summarized. The relevant test data sets of drones participating in the last-mile delivery are sorted out. Finally, the future application and research direction of drones in the last-mile delivery are discussed.

Abstract:The operating system kernel is the most fundamental software component in a computer system. It controls and manages computer hardware resources and provides interfaces and services necessary for accessing and managing other applications. The security of the operating system kernel directly affects the stability and reliability of the entire computer system. Kernel fuzzing is an efficient and accurate security vulnerability detection method. However, in current kernel fuzzing work, the overhead of calculating the relationship between system calls is too high, or it is easy to misjudge the relationship between system calls. In addition, the existing method for constructing system call sequences lacks reasonable energy allocation, making it difficult to explore problems of low-frequency system calls. This study proposes to learn the relationship between system calls by using an N-gram model and prioritize the expansion of system calls with low frequency or high TF-IDF values based on the frequency and TF-IDF information of system call occurrences. With minimal overhead, this study achieves a coverage increase of 15.8% and 14.7% in 24-hour experiments on Linux versions 4.19 and 5.19, respectively. Besides, one known CVE (CVE-2022-3524) and eight new crashes are discovered, one of which is numbered CNNVD (CNNVD-2023-84723975).

Abstract:Fuzzing plays a significant role in discovering security vulnerabilities and improving security in protocol software. In recent years, the introduction of the state into server program fuzzing has received widespread attention. This study addresses the problem of low efficiency of fuzzing due to the insufficient utilization of information in the protocol fuzzing process and the inability to continuously focus on key states. The study also proposes a protocol fuzzing method based on the cooperation of double cover information. Firstly, the state selection algorithm proposed in this study sets weights for each state by mapping the state space to the program space and using heuristic calculation methods to guide the fuzzing to continuously focus on states that are more likely to have defects. Secondly, the study detects a seed position that will not affect the state but can change the program coverage and restricts the mutation position to adequately test the code area corresponding to the focus state. The study also verifies the effectiveness of the improved algorithm on the baseline tools AFLNet and SnapFuzz and integrates them into a protocol fuzzing tool, namely C2SFuzz. Experiments are carried out on the latest version of protocol server programs such as LightFTP and Live555, and five unknown vulnerabilities are detected.

Abstract:Support vector machine (SVM) is a machine learning method based on structural risk minimization and can solve classification problems. However, with the complexity of research problems, the real classification problems are often multi-classification ones, whereas SVM can only be adopted to deal with binary classification tasks. To this end, the multiple birth support vector machine (MBSVM) combined with the one-against-all strategy can realize multi-classification with low complexity, but the classification accuracy is low. This study improves MBSVM and proposes a new SVM multi-classification algorithm which is a multiple birth support vector machine based on the hypersphere and fruit fly optimization algorithm with adaptive step size reduction (ASSRFOA). The algorithm is referred to as HA-MBSVM. Through the information obtained from hypersphere fitting, firstly all classes are divided into several blocks and then classifiers are constructed for each class. The constraint distance regulation factor is introduced to properly improve the difference of the classifiers. At the same time, ASSRFOA is employed to solve the quadratic programming problems and HA-MBSVM can better solve the multi-classification problems. Six datasets are utilized to evaluate the performance of HA-MBSVM. The experimental results show that the overall performance of HA-MBSVM is better than that of the comparison algorithms.

Abstract:The continuous development of information technology and the popularization of intelligent terminal devices have led to the continuous growth of the total amount of global data storage, and the threats and challenges faced by data have increased with the prominence of their importance. However, currently, some computing and storage devices still lack data protection modules or have weak data protection capabilities. Existing data security storage technologies generally protect data through encryption, but data encryption and decryption operations, or data protection processes, are usually performed on the applied devices, resulting in threats to the security of stored data when the applied devices are subjected to various attacks. In response to the above issues, this study proposes a DICE-based Internet of Things (IoT) device attestation storage scheme, which utilizes trusted IoT devices built based on the lightweight root of trust DICE to provide secure storage services for general-purpose computing devices (collectively referred to as hosts), moves data encryption and decryption operations to trusted IoT devices, and eliminates threats to stored data caused by risks such as host memory attacks. This study mainly includes the following three aspects: (1) building a trusted IoT device by using the root of trust DICE to provide a security prerequisite for providing trusted services; (2) establishing a DICE-based remote attestation mechanism and access control mechanism to achieve secure authentication and establish a secure communication channel; (3) using the trusted IoT device to provide trusted and secure storage services for legitimate host users, which achieves secure data storage and takes into account isolation and flexibility in the use process. The experimental results show that the secure storage service provided by this scheme has a high file transfer rate and high security, which can meet the requirements for secure data storage in general scenarios.

Abstract:Due to the complexity of the Java Web application business scenario and the high requirement for the structural validity of the input data, the existing test methods and tools have the problems of low efficiency of test cases when testing Java Web. To solve the above problems, this study presents a gray-box fuzzing method for Java Web applications based on parse trees. First, the study models the syntax of the input packets of Java Web applications, creates a parse tree, distinguishes between delimiters and data blocks, and hooks up a seed pool for each leaf node in the parse tree. In addition, the study isolates the single data block of the test case and generates the input in line with the Java Web application business format by packet splicing, so as to improve the efficiency of test cases. In addition, in order to retain high-quality data blocks, each data block seed is assigned a separate weight during the test according to the execution feedback information of the test program. In order to break through the deep path, the seed features of data blocks are extracted based on conditional probability learning in the corresponding seed pool. This study implements a gray-box fuzzing test system for Java Web applications based on parse trees, namely PTreeFuzz, and the test results show that the system achieves better test accuracy compared with existing tools.

Abstract:Source code migration techniques are designed to convert source code from one programming language to another, which helps reduce developers’ burden in migrating software projects. Existing studies mainly use neural machine translation (NMT) models to convert source code to target code. However, these studies ignore the code structure features, resulting in poor source code migration performance. Therefore, this study proposes a source code migration model based on a code-statement masked attention Transformer (CSMAT). The model uses Transformer’s masked attention mechanism to guide the model to understand the syntax and semantics of source code statements and inter-statement contextual features when encoding and make the model focus on and align the source code statements when decoding, so as to improve migration performance of source code. Empirical studies are conducted on the real project dataset, namely CodeTrans, and model performance is evaluated by using four metrics. The experimental results have validated the effectiveness of CSMAT and the applicability of the code-statement masked attention mechanism to pre-trained models.

Abstract:At present, convolutional neural networks (CNNs) based on local attention mechanism have yielded sound results in feature extraction of kinship recognition. However, the improvement of backbone models based on CNNs is not obvious, and few researchers employ self-attention mechanisms with global information capture ability. Therefore, an S-ViT model based on a convolution-free backbone feature extraction network is proposed, which is to adopt Vision Transformer with a self-global attention mechanism as the basic backbone feature extraction network. By constructing a twin network and a CNN with a local attention mechanism, the traditional classification network is expanded for research on related issues of kinship recognition. The final experimental results show that compared with the leading method of the RFIW2020 Challenge, the proposed method has performed well in the three kinship recognition tasks. The first task ranks second with verification accuracy of 76.8%, and the second and third tasks rank third. As a result, the feasibility and effectiveness of the method are improved to propose a new solution to kinship recognition.

Abstract:In the field of automatic driving, existing target detection algorithms are haunted by low detection precision due to complicated road backgrounds and insufficient information about small targets. Since the onboard camera has fixed viewing angles, and targets on the road are somewhat regularly distributed in the image space, richer information can be provided to autonomous vehicles for target detection. Therefore, a spatial feature augmentation network (SE-YOLOv5s) to improve YOLOv5s is proposed. A location attention module (LAM) is added to the neck network of YOLOv5s, which can be weighted according to the distribution characteristics of road targets in the image and enhance the network’s perception and localization ability for the target category location distribution. A small target enhancement module (STEM) is designed to fuse shallow features and deep ones, so as to obtain richer semantic information and detailed space information about small targets, thereby improving the detection effect of small targets. The results of the experiment show that the improved model witnesses an increase in detection precision against targets of different scales, with AP_S increased by 2.8%, AP_M increased by 2.5%, and AP_L increased by 2%.

Abstract:In software-defined wide area networks (SD-WANs), link failures can result in substantial packet loss, leading to partial network paralysis in severe cases. The existing traffic engineering approaches can expedite failure recovery by installing backup paths in advance on the data plane. However, it is difficult to adapt to various network failures with limited resources, which degrades the network performance after recovery. To maintain the network performance after failure recovery and reduce the consumption of backup resources, this study proposes a proactive failure recovery scheme based on congestion and memory awareness (CAMA), which can not only redirect the affected data flows quickly but also realize the load balancing to avoid the potential link congestion after recovery. Experimental results demonstrate that compared with existing schemes, CAMA can effectively utilize backup resources, performs well in load balancing, and requires only a small number of backup rules to cover all single-link failure scenarios.

Abstract:Optic neuritis is an eye nerve disease that causes acute optic nerve injury in children and adults, and there is a risk of blindness in severe cases. Therefore, early detection and diagnosis of optic neuritis is of great help to the recovery of patients. Based on the fact that the characteristics of retinal image lesions of optic neuritis are not obvious and the classification of artificial diagnosis is difficult, with low accuracy, an improved hybrid attention mechanism CS-CBAM module is designed in this study, and it is integrated into the improved AlexNet network to form a deeper AlexNet2_att optic neuritis classification model, so as to realize the automatic classification of optic neuritis images. First, the retinal images in the dataset are preprocessed through image size adjustment, removal of image redundancy information, histogram equalization, and data enhancement. Then, based on the AlexNet network, the batch normalization layer is introduced to improve the training speed, and then the proposed hybrid attention mechanism CS-CBAM is integrated into the improved AlexNet network to form an AlexNet2_att model. Finally, the clinical data from the Third People’s Hospital of Dalian are used to evaluate the performance of the network, and the experimental results show that the classification accuracy of the model can reach 99.19%, which proves that the model has excellent practicability and robustness and high practical value, so it can assist doctors in the classification and diagnosis of optic neuritis.

Abstract:When the traditional fault diagnosis method is used to judge the bearing fault, there are some problems such as insufficient feature extraction, incomplete use of time sequence characteristics, and complicated calculation. In order to solve these problems, a combined diagnosis method based on a multi-scale convolutional neural network (MCNN), bidirectional gated recurrent unit (BiGRU), and attention mechanism is proposed in this study. Firstly, MCNN is used to extract multi-scale features from signal data, which realizes further extraction of features in terms of spatial level. Secondly, the BiGRU network is used to obtain sequential relations from positive and negative directions in terms of time level. Next, an attention mechanism is introduced to ignore some information that is not relevant to the results and reduce the risk of information loss to improve accuracy. After mapping is created through the full connection layer, the bearing fault diagnosis is finally completed through the Softmax classification method. In this study, LSTM model, GRU model, SVM model, CNN-BiGRU model, and other models are compared through experiments, and the experimental results show that the proposed model has higher accuracy than the previous models. The accuracy of fault diagnosis in a single working condition reaches 98.1%, and the average accuracy of fault diagnosis in multiple working conditions reaches 97.8%.

Abstract:Mobile target detection algorithms require fewer model parameters, less computation, faster reasoning speed, and better detection effects. The target detection algorithms are prone to false detection of small targets and missing detection and have insufficient ability for feature extraction. To this end, this study proposes a lightweight small target detection algorithm based on YOLOv5. In this algorithm, the lightweight network MobileNetV2 is used as the backbone network of the target detection algorithm to reduce the number of parameters and calculation amount of the model. The deep separable convolution combined with a large convolution kernel is applied to decline the number of parameters and calculation amount, and improve the detection accuracy of small targets. GhostConv is adopted to replace part of common convolution to further decrease the number of parameters and computation amount. Multiple comparison experiments are carried out on VOC competition data sets and COCO competition data sets. The results show that compared with other models, the proposed algorithm has fewer parameters, less computation, faster reasoning speed, and higher detection accuracy.

Abstract:Motor imagery is a promising brain-computer interface paradigm. In the motor imagery classification tasks based on EEG, the equipment and the subjects will lead to the phenomenon of data distribution drift related to the subjects and time. This data distribution drift will reduce the classification accuracy of the classifier. Transfer learning can solve this distribution drift phenomenon very well. In this study, a new single source domain selection algorithm, multi-subdomain transferability estimation (MSTE) and a new transfer method, task-oriented subdomain adversarial transfer network (ToSAN), for the classification tasks of EEG signals are proposed. MSTE can evaluate the similarity in time and category between the source domain and the target domain. ToSAN can decompose features for classification tasks and perform multiple subdomain alignments on task-related features to overcome distribution differences. The experimental results on BCI Competition IV 2a and BCI Competition IV 2b show that compared with other methods, ToSAN improves the classification accuracy by at least 2.67% and 8.6%, respectively. The combination of MSTE and ToSAN achieve a classification accuracy of 81.73% and 88.73% on the BCI Competition IV 2a and BCI Competition IV 2b datasets, which is significantly better than all comparison methods.

Abstract:Taking the scheduling problem in the flexible manufacturing of a large furniture enterprise as the research object, this study proposes a multi-strategy whale optimization algorithm (MWOA), which is mainly used to solve the flexible job shop scheduling problem. First, in order to improve the diversity of the initial population, chaos theory is introduced to initialize the population; at the same time, the nonlinear convergence factor and adaptive inertia weight coefficient are designed to balance the global exploration and local development capabilities; then the differential evolution (DE) operator is used to improve the utilization and search ability of WOA. Finally, the optimal individual chaotic search strategy is adopted to reduce the probability of premature convergence of WOA. With the objective of minimizing the maximum completion time, the benchmark test problem and the scheduling optimization problem of the manufacturing process of a furniture enterprise are solved. The results show that MWOA overcomes the shortcomings of the basic WOA, such as low optimization accuracy, slow convergence speed, and easy falling into local optimization. Compared with the comparison algorithm, MWOA achieves better optimization results.

Abstract:Deep reinforcement learning can be used to extract effective information from high-dimensional images and thus automatically generate effective strategies for solving complex tasks such as game AI, robot control, and autonomous driving. However, due to the complexity of the task environment and the low exploration efficiency of the agent, it is still necessary for the agent to interact with the environment frequently even for relatively simple tasks. Therefore, this study proposes a CCLF algorithm (Bootstrapped CCLF), which combines Bootstrapped exploration method to generate more different potential actions through multiple heads in the actor network, so that more different states can be accessed to improve the exploration efficiency of the agent, and thus the convergence process can be accelerated. The experimental results show that the algorithm has better performance and stability than the original algorithm in the DeepMind Control environment, which proves the effectiveness of the algorithm.

Abstract:Wind speed prediction is an important factor affecting the efficiency and stability of wind farms. Based on the spatio-temporal features of wind speed, the VMD-based hybrid spatio-temporal network (VHSTN) integrates variational modal decomposition (VMD) and hybrid deep learning framework to predict the short-term wind speed. The hybrid deep learning framework is composed of convolutional neural network (CNN), long and short-term memory (LSTM), and self-attention mechanism (SAM). After the cleaning of raw data, the VMD is employed to decompose the spatio-temporal data of wind speed for multiple sites into intrinsic mode functions (IMF) components, eliminating the instability of the wind speed data. For each IMF component, the spatial features are extracted by the CNN at the bottom of the model. Next, the temporal features are captured by the top-level LSTM. Then, SAM is applied to strengthen the extraction of key hidden features through adaptive weighting and obtain the prediction results of each component. Finally, the results are amalgamated to determine the final predicted wind speed. Experiments are conducted on the commonly used dataset WIND in this study. The experimental results prove the effectiveness and superiority of the proposed algorithm compared with related typical algorithms.

Abstract:With the gradual development of smart factories, mobile robots are applied more and more widely in the factory. However, as there are many obstacles in the factory, the traditional artificial potential field method is easy to produce unreachable targets and local minimum values and other problems. This study improves the unreachable target and the local optimal solution of the traditional artificial potential field method in path planning. Firstly, a new repulsive potential field function is adopted to solve the problem of unreachable targets by adding an influence function to the repulsive potential field function in the original artificial potential field method. Secondly, for the local optimal solution, the artificial potential field method is combined with the simulated annealing method, and the additional subpoints in the simulated annealing method are applied to break the equilibrium state, so as to get out of the obstacles. Finally, through Matlab comparison, the travel time of the proposed algorithm in 10 obstacles is improved by 6.70% and the path length is reduced by 9.20% compared with algorithms in other literature. In 20 obstacles, the travel time of the proposed algorithm is improved by 9.10% and the path length is reduced by 12.10% compared with algorithms in other literature.

Abstract:A subpixel edge detection algorithm based on Zernike moment is proposed to detect the size of the valve steam seal. Firstly, the preprocessing effect of the Gaussian filter and bilateral filter is compared and analyzed. The results show that bilateral filter perform better in filtering and edge information retaining. Then, the Canny operator is employed for rough edge localization, and the adaptive Zernike moment transition model is adopted for subpixel level localization. Finally, the coordinates and radius of the center of the circle are obtained by the least squares fitting circle method. Experiments prove that the optimized algorithm has a smaller average error in edge positioning with more accurate accuracy.

Abstract:The tower crane hook video has poor video stabilization due to complex working conditions at the construction site. To solve the problem, this study proposes a hook video stabilization algorithm integrating the improved ORB feature matching and fixed-lag Kalman filter. In the image motion estimation, the classical ORB algorithm is improved by adopting image chunking and adaptive thresholding for feature point extraction and introducing the image quadratic tree algorithm to improve the uniformity of image feature point distribution; on this basis, the background compensation combined with inter-frame difference method is used to quickly identify local motion targets and reject them, which improves the accuracy of global motion parameter estimation; in the motion filter and compensation stage, a fixed-lag Kalman filter algorithm is used to remove the random jitter components, so as to obtain the motion compensation parameters for video stabilization, thus achieving video monitoring stabilization of the tower crane hook visualization system. The experimental results show that compared with that of the classical ORB and Kalman filter algorithm for image stabilization, the inter-frame transform fidelity (ITF) of the proposed algorithm is improved by about 9.12%, and the average value of structural similarity ($ \overline {{\text{SSIM}}} $) is improved by about 2.75%, resulting in a better image stabilization effect, and inter-frame processing speed (FPS) reaches 29.65 f/s, meeting the real-time monitoring requirements of the tower crane.

Abstract:In order to effectively maintain the significant geometric features of scattered point clouds and improve the accuracy and efficiency of point cloud simplification, a point cloud simplification method based on point importance judgment is proposed. Firstly, the importance of points in point clouds is calculated, and feature points are extracted according to the importance. Then, the octree algorithm is used to simplify the non-feature points, so as to retain the main details of the point cloud and realize the simplification of the point cloud. Finally, the point cloud simplification method is verified by simplifying the data model of public point clouds and cultural relics point clouds. The results show that the point cloud simplification method based on point importance judgment can effectively simplify the point cloud while maintaining the detailed geometric characteristics of the point cloud. It is a fast and high-precision point cloud simplification method.

Abstract:The temperature of Lanzhou City is a non-stationary sequence with typical characteristics of loud noise and instability. As the temperature changes greatly, the sequence gets unstable. In order to improve the prediction accuracy, strengthen the generalization ability, and reduce the sensitivity of parameter selection of support vector machine (SVM) in temperature prediction, in this study, the improved particle swarm optimization (IPSO) algorithm is proposed to optimize the temperature prediction model of SVM. Firstly, the adaptive inertia weight is introduced into the particle swarm optimization (PSO) algorithm to improve the global optimization ability and local development ability of the PSO algorithm. Secondly, the improved IPSO is used to optimize the penalty factor and kernel function parameter of the SVM, and the optimized model (IPSO-SVM) is applied to the temperature prediction. The actual data of Lanzhou ground observation station are taken as sample data, and Matlab experimental tools are used for training and prediction. The experimental results show that the IPSO-SVM model in this study has a stronger generalization ability and better fitting degree than back propagation (BP), SVM, GRID-SVM, GWO-SVM, ABC-SVM, and ACO-SVM. It can predict the change in temperatures more accurately, which further verifies the feasibility of this model in temperature prediction.

Abstract:Violence can be easily occluded, and the recognition accuracy is low. At present, some algorithms add multi-view video input to solve the occlusion problem and fuse all view data with equal weight. However, video from different views differs in recognition due to shooting distance and occlusion itself. To solve this problem, this study proposes a violence recognition method based on view confidence and attention to improve the accuracy of violence recognition. The input of the temporal difference module (TDM) is expanded to a multi-view angle. The channel attention mechanism is applied to the segment dimension to enhance the ability of cross-segment feature extraction in TDM. The background suppression method is used to highlight the texture features of moving objects and calculate the image confidence of each view. The bilinear pooling method is introduced to fuse multi-view video features, and the weight of local features of each view is assigned according to the view confidence. In this study, validation is performed on both the public dataset CASIA-Action and the self-made dataset. Experiments show that the view confidence method proposed in this study is better than the bilinear pooling method before improvement, and the accuracy of violence recognition is better than that of the existing behavior recognition methods.

Abstract:During human-computer interaction, excessive mental workload is an important factor to produce operation errors. At present, EEG signals are often employed for the evaluation of mental workload based on their characteristics of high time resolution and good portability. In recent years, the rapid development of deep learning leads to its widespread application in brain electricity and better results are yielded than traditional machine learning. The n-back task can induce different degrees of psychological loads by setting different n values. In this study, the n-back paradigm based on vision and hearing is designed to avoid a single dimension. Additionally, a new convolutional neural network model is proposed. The data collected by 64-channel eego EEG equipment are preprocessed by eeglab for the training of the model. Compared with the performance of EEGNet, FBCNet, and ShallowConvNet in the test set, the classification accuracy of the proposed model is significantly improved, and thus this study has certain application potential in the evaluation of mental workload, especially in the classification of multi-dimensional n-back tasks.

Abstract:Bionic fish has broad prospect for engineering application. For the control of bionic fish, the first thing to solve is the tracking problem. However, the existing fish control methods based on CFD methods and traditional control algorithms feature high training data acquisition costs and unstable control. This study proposes an intelligent control method based on the PPO algorithm for bionic fish tracking. The surrogate model is employed instead of CFD to generate training data to improve the data generation efficiency. The efficient PPO algorithm is introduced to accelerate the learning speed of the strategy model and improve the utility of the training data. The speed parameter is introduced to solve the problem that the fish cannot track smoothly in the sharp turning area. Experiments show that the proposed method has faster convergence speed and more stable control ability in various paths, with guiding significance for the intelligent control of bionic robotic fish.

Abstract:In recent years, student cognitive diagnosis has been an important research topic in educational data mining, which is of great significance for accurate feedback in modern education. However, traditional cognitive diagnosis models have problems such as low prediction accuracy and low efficiency when dealing with large-scale data. Moreover, the existing research is mainly focused on traditional offline teaching and learning, and more research is needed in programming education. To solve the above problems, a programming-performance-based fuzzy cognitive diagnosis framework (P-FuzzyCDF) is proposed from the analysis of the characteristics of programming education. First, to deal with the case of partially correct programming questions, the model fuzzes the students’ mastery of the knowledge points. Second, fuzzy set theory is combined with educational assumptions to model student mastery of the questions. Finally, students’ scores on each problem are generated by considering plagiarism factors. Notably, the model takes advantage of the visualization and accuracy of programming education data to quantify the parameters for each model component. Experiments are conducted based on real data sets, and the results show that P-FuzzyCDF can achieve high accuracy, where the values of MAE, MSE, and RMSE assessment indexes are 0.07, 0.09, and 0.01, respectively. In addition, when comparing P-FuzzyCDF with existing classical methods such as DINA, IRT, and FuzzyCDF, the results of P-FuzzyCDF are significantly better than these methods in terms of MAE, MSE, and RMSE.

Abstract:Distributed average consensus and decentralized machine learning are widely employed decentralized computing methods. The convergence rates of the two methods are mainly determined by the spectral gap of the topology. The heterogeneity of the network environment among nodes includes the difference in node bandwidth and inter-node connection availability. The heterogeneous network environment poses a challenge to decentralized computation efficiency. This work studies the topology design of maximizing the spectral gap under a heterogeneous network environment. The gradient of the spectral gap for any edge of the topology is derived and an edge-addition and deletion algorithm is designed based on this gradient to construct the target topology. The generated topology has larger spectral gaps and similar data communication time of each node. The performance of this algorithm remains stable under different levels of heterogeneous network environments. The generated topology achieves convergence with a faster convergence rate and shorter time in distributed consensus. Based on this algorithm, this paper further verifies the recently discovered weak relationship between the spectral gap and convergence rate of decentralized machine learning.

Abstract:The Ethereum virtual machine (EVM) is a key component of the Ethereum blockchain, and its defects will cause deviations in the execution results of transactions, which will bring serious problems to the Ethereum ecosystem. The existing work on EVM defect detection only treats the virtual machine as an independent smart contract execution tool and does not fully test its workflow, resulting in blind spots in defect detection. To solve the above problems, a defect detection method for the whole process of EVM operation (ETHCOV) is proposed. ETHCOV first combines the weight strategy to guide smart contracts, contract interface parameter inputs, and transaction sequences to vary at different granularities. It then packages them with block state and world state as test cases and finally inputs the test cases into the EVM to trigger the run and compare the test run results, so as to detect the vulnerabilities in the EVM. Based on the above method, a prototype system is implemented and more than 20 000 real smart contracts are tested as input to the EVM for defect detection. Experimental results show that compared with the existing tool EVMFuzzer, ETHCOV improves the test efficiency by 339% and the code coverage by 125%, and the inconsistent output of three sets of test cases is detected. These results show that ETHCOV can effectively detect defects in the EVM.

Abstract:Based on previous work, this study proposes that the self-attention mechanism guided by syntactic dependency can integrate syntactic dependency knowledge to improve the performance of Chinese word segmentation so that the self-attention mechanism can only focus on those characters that have syntactic dependency influence on the current character’s word segmentation label and learn their influence degree on the current character. In addition, this study performs positional encoding on the self-attention mechanism guided by syntactic dependency trees. The experimental results show that the model has improved its performance compared with the baseline, and the recognition ability of the model for unregistered words has been strengthened.

Abstract:To solve the problems such as high data transmission delay and weak dynamic adaptability of underwater acoustic communication, this study proposes an intelligent underwater acoustic network routing protocol based on Q-learning ant colony optimization (QACO). The protocol includes routing behavior and intelligent decision. In the route discovery and maintenance phase, the construction of the network topology environment and information exchange among nodes as well as the network maintenance are carried out by intelligent NetAnts. In the Q-learning phase, the node energy and depth and network transmission delay learning characteristics are quantified as discount factors and learning rates to extend the network lifecycle and reduce system energy consumption and delay. Finally, simulations are carried out through the underwater acoustic network environment, and the experimental results show that QACO outperforms the Q-learning aided ant colony routing protocol (QLACO), Q-learning-based energy-efficient and lifetime-aware routing protocol (QELAR), and depth-based routing (DBR) algorithm in terms of energy consumption, delay, and network lifecycle.