This PDF file contains the front matter associated with SPIE Proceedings Volume 12718, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

In order to solve the problem of reasonable coupon issuance under the intelligent marketing scenario, this paper proposes a coupon recommendation method based on machine learning and natural heuristic algorithm. First, the machine learning algorithm is used to predict the probability of consumers using coupons. Then, according to the marketing cost constraints and the prediction results in the first stage, the optimization objective function is established, and the natural heuristic algorithm is used to simulate annealing (SA) to solve the optimal value, so as to achieve the purpose of issuing coupons to the most suitable consumers. The recommendation system designed according to this method can build data- driven customer attributes, effectively improve the order conversion rate, promote revenue growth, and enhance brand influence.

With the continuous development of educational informatization, students have produced a large amount of data in the process of learning and life. The educational management system with simple search and query function can not find the potential value behind large-scale data and can not meet the practical needs of teaching management. In view of the above problems, in order to mine useful information from student behavior data, this paper collected and arranged student behavior data and achievement data, constructed an index system to describe student behavior, clustered student behavior data with K-means clustering algorithm, and generated student behavior characteristic description data. Apriori algorithm was used for association mining of feature data sets. Based on the mining results, the association rules between students' behavior and grades were analyzed, which provides a reference for students' behavior guidance and teaching management.

Distributed crowdsourcing platforms technology has been widely used in existing applications based on network including Amazon, eBay and Alibaba. Apparently, current distributed method can response almost requests immediately when the system received and prevent almost attacks from malicious users and make the whole platforms with high robustness. However, the privacy information attack can hardly be prevented by distributed servers and may cause serious damage for individuals due to all servers contain the privacy information from the platform users by current system structure. In this paper, we propose a novel privacy information protocol for distributed crowdsourcing systems and simulated our method in the real crowdsourcing system to evaluate the performance of our mechanism. From our extensive experimental result, we can conclude that our proposed method can prevent the information leakage from malicious attackers and guarantee the security of the whole crowdsourcing system with reasonable communication and computation costs.

In this paper, the Inception of lightweight encrypted traffic classification with depthwise separable convolution is proposed. By fusing the depthwise separable convolution on the basis of Inception model, the model reduces the parameters and computation of the model to a large extent without reducing the accuracy. The main motivation of this paper is to reduce the consumption of the network as much as possible, and to develop a model with small computation and high precision. The effectiveness of the proposed method has been verified in the internationally open ISCX VPN-non VPN dataset test experiment. Firstly, the original data were cleaned and converted to transform the encrypted traffic samples into gray maps. Secondly, 1D-CNN and Inception model integrated with depthwise classifiable convolution were used for classification. The experimental results show that the classification accuracy rate, precision rate, recall rate and F1-score are 96.27%, 95.49%, 97.39% and 96.43% respectively for encrypted traffic classified by 12 categories.

Due to the high economy and drivability performance objectives, the dual-drive hybrid system which has two motors is developed. In the diverse conditions, the drive modes which are on account of the engine status, the requested driver torque and power, C0 status are presented. The unique challenge of the dual-drive hybrid system is drive modes transition. Specifically, the path transition from the electric mode control to parallel mode control in the acceleration and from parallel mode to electric mode in the brake acceleration. The drive mode control process is divided into five control phases. Finally, tests are successfully performed in the real vehicles to prove that the drivability performance are greatly improved and the innovative control strategy is also demonstrated.

In order to further improve the network information security and solve the shortcomings of the previous encryption algorithms, in this research, based on the mainstream encryption algorithms RSA and AES, and integrating their advantages, the encryption algorithm of RSA and AES is developed and designed. On the basis of discussing the configuration of hardware and software supporting the algorithm, the performance test of the algorithm developed this time is carried out. The test results show that the comprehensive performance of the algorithm is outstanding, and it is expected to be gradually popularized and applied in the future information encryption work.

QR code is widely used in many scenarios including online transactions, warehousing, manufacturing and authentication. However, in the era of continuous development of quantum technology, the traditional QR code encryption method has low security in front of quantum technology, and information is easy to be stolen and forged. In response to the above problems, we consider combining post-quantum cryptography with QR code. In this work, we propose a quantum-safe, anti-counterfeiting and tamper-proof QR code based on lattice-based cryptography. The proposed QR code technique features two advantages: (a) it can be easily deployed on mobile phone apps and the verification is computationally efficient; (b) it is safe under quantum attacks and impossible to be forged and imitated in plausible quantum future.

In this paper, CAEG (Crash-based Automatic Exploit Generation) is proposed to automatically generate binary program exploits based on known crash inputs. In CAEG we analyze the input causing a program crash and address two main challenges: how to reproduce the crash state and how to automatically generate control flow hijacking exploits. Firstly, a path-oriented algorithm is proposed to find the crash path using symbolic execution techniques by treating the crash input as a symbolic value. Secondly, we summarize the principle of multiple control flow hijacking vulnerability exploits. In addition, consider bypassing the vulnerability mitigation mechanism by using springboard instructions to bypass Address Space Layout Randomization (ASLR) and return-to-libc to bypass Non-executable bit (NX). We tested 11 open source software with vulnerabilities (8 from the test sets of AEG and MAYHEM, and 3 from the CVE and EDB vulnerability repositories). The experimental results show that our scheme is more efficient than AEG and MAYHEM.

In this paper, we propose an offloading method for edge computing that takes into account the energy consumption of task requests, constructs and analyzes a co-tracked multi-star structure, calculates the individual task delay and builds a multi-layer edge projection pursuit (MLEPP) model to calculate the optimal offloading location. The experimental results show that this method has lower average offloading delay and higher offloading efficiency.

In order to meet the requirement of remote dynamic authorization access to experimental equipment in the Mitsubishi Electric Automation Laboratory of Tianjin University, an online remote experimental teaching system has been designed and developed. The system mainly consists of two parts, one is the reservation management subsystem of experimental equipment based on LAMP architecture, and the other is the programmable VPN subsystem based on OCServ. The main functions of the experimental equipment reservation management subsystem include experimental resource information, available resource display, reservation management, reservation approval, history query, and so on. The programmable VPN subsystem can dynamically add and delete VPN users in batch according to the reservation of experimental equipment, and realize remote dynamic authorization access to the equipment. With the help of this system, universities in Tianjin can share the laboratory equipment and improve the utilization rate of the laboratory equipment. The system solves the problem of remote security communication control over the Internet by VPN technology, and is suitable for all experimental equipment using IP protocol to communicate.

For the purpose of improving the phenomenon of short period of the sequence result in using single chaotic map, this paper proposes a pseudorandom sequence generator scheme based on Compertz map and piecewise map. The sequence generated by our scheme passed the National Institute of Standards and Technology (NIST) randomness tests, compared with other schemes, the sequence generated by our scheme shows good randomness and statistical characteristics. In the aspect of application, the sequence generated by our scheme can be use as pseudorandom numbers and secret key. In terms of confidentiality, the seed key consist of the combination of control parameters of piecewise map, initial value of chaotic system, number of initial iteration rounds of chaotic map and left shift of key sequence. In the aspect of security, the secret key generating by our scheme will cost a lot for the attacker to crack.

The main task of a ciphertext database is to store ciphertext data and perform simple operations on ciphertext data while ensuring data security. In this paper, the problem that mutable order-preserving encryption(mOPE) scheme in the ciphertext database system requires frequent interaction between client and server to locate the ciphertext data location during ciphertext insertion and ciphertext retrieval is addressed, which leads to performance loss. B-mOPE scheme is proposed as an optimized mOPE scheme for ciphertext databases by designing a new data structure, a boundary search tree. Experimental results show that B-mOPE scheme has a significant improvement in execution performance compared with original mOPE scheme and satisfies the security criteria of the order-preserving encryption (OPE) scheme.

With the continuous advancement of network attack technology, the complexity of defending networks has become more and more challenging, resulting in increasingly prominent network security issues. At the moment, research on network defence security focuses on strategy selection and technology implementation, yet there is a lack of a unified definition and a common model framework for the network defence process. This paper examines the overall process of network defense decision-making and addresses the issues of indistinct outlines of the network defense system and a lack of clear explanations of the command and decision-making process, beginning with the architecture. Furthermore, it analyzes the operational view of network defense based on the established architecture framework DODAF(Department of Defense Architecture Framework). The core operational view models, such as the Operational Concept Diagram, Organizational Relationship Diagram, Operational Activity Diagram, and Operational Node Connection Diagram, are initially constructed to describe and analyze the network defense operational system structure and operational process from a macro perspective, and to clarify the interrelationship between defense operational objectives, major activities, combat roles, and major information flow elements.

Network intrusion detection system (NIDS) plays an important role in network security. With the continuous development of technology, machine learning and deep learning are gradually becoming the main methods of NIDS. However, a large amount of network traffic data has the problem of manual labeling, which results in a limited train datasets, and reduces the performance of NIDS. Semi-supervised learning is a new approach that combines supervised and unsupervised learning to analyze large unlabeled datasets with a small number of labels. In this paper, we propose a semi-supervised deep learning method, which uses improved tri-training algorithm, and combines with deep learning model. We verified the performance of the proposed method on CICIDS2017 datasets. The experimental results show that the proposed method can improve performance of NIDS and outperform other semi-supervised learning methods.

Remote piano teaching is an important part of piano teaching curriculum construction music and art colleges. At the same time, as a new model and new normal, distance education conforms to the goal of cultivating new applied talents. IPSecV6 provides a good technical guarantee for the development of distance piano teaching with its characteristics and advantages of bandwidth guarantee, efficient transmission, safety and reliability. In the application design of IPSecV6 for remote piano experiment simulation, network security is the first technical element of campus network technology research from beginning to end. In view of this, the article analyzes and studies the remote piano teaching of IPSecV6 security simulation. The simulation results show that the scheme can provide network security guarantee of access control, data encryption and security verification for remote piano teaching. At the same time, this model also provides new ideas and reference for modern distance education teaching.

Efficient and flexible information flow mode is an important guarantee for the victory of information war. The current research on battlefield information flow mode is difficult to adapt to the network centric operational concept. The performance analysis of information flow mode is mostly focused on the topological structure characteristics of complex networks, and there is no method to analyze the overall performance of information flow. Aiming at the construction of battlefield information flow mode under information conditions, a general model of battlefield information flow based on complex networks is established, an information flow network of basic campaign scale is generated, and key indicators describing the overall performance of the information flow network are established. Through simulation experiments, analysis and comparison are made, and the overall performance of the information flow mode constructed in this paper has certain performance advantages.

The use of fileless technologies in malware continues to grow and fileless malware becomes more dangerous and difficult to detect. To address this challenge, we propose a novel visual method for classifying fileless malware based on few-shot learning. First, we built a fileless malware dataset, which is executed through a local virtual environment to collect malware memory dumps. Secondly, memory dumps are clipped and visualized. We developed a new memory dumps trimming method and a novel binary file visualization technique, which can remove redundant data from memory dumps, significantly compress the file size, and then represent the trimmed memory dumps as RGB images. Finally, we propose a few-shot learning framework, namely MMEL (MAML + Mean_subtraction + Euclidean_normalization + Label_Smothing), to improve the performance of the classification method. Experimental results show that our visualization technique and framework outperform other state-of-the-art few-shot learning methods.

In recent years, the few-shot text classification task has received increasing attention, researchers have proposed various methods to enable the model to classify text in a few labeled text instances. The traditional metric-based learning models map text features into a new vector space, allowing similar points closer and dissimilar points farther away, but these models focus too much on the semantic information of the text itself, ignoring the potiental relationship information between texts. However, the graph neural network is good at discovering such information, so a prototype network based on the distributed graph neural network is proposed to solve this problem. First, the distributed graph neural network will use both distribution-level information and instance-level information, and let them convert to each other when updating to make them fully fused, and then use the prototype network for classification. The experimental results demonstrate that the proposed model has improved classification accuracy compared with other few-shot text classification models.

The relatively large variation in samples from different categories in the object detection dataset leads to the problem of sample imbalance, which in turn affects the accuracy of the object detector. In this paper, a dynamically weighted approach to label assignment is proposed to enable model to suppress abnormal samples and mine hard samples during training, thus solving the problem of sample imbalance. Different dynamic weighting strategies are applied to positive and negative samples to calculate the classification loss, and a dynamically weighted EIOU method is used to calculate the localization loss. In addition, modifying the single convolutional prediction head to a dual prediction head structure enables the model to focus more on performing classification and regression tasks. This structure uses a fully-connected head to predict the classification results and a convolutional head to predict the localisation results. Experiments conducted on PASCAL VOC2007+12 show that the proposed method can help YOLOv5 to achieve better accuracy and reach 72.3 mAP in 100 epochs, an improvement of 2.4 mAP over the original model. Furthermore, we have done several experiments on MSCOCO to demonstrate the effectiveness of the proposed dynamic weighting method, which has implications for building faster and stronger object detectors.

This paper is based on end-to-end speech recognition based on convolutional neural network technology. The problem that convolutional neural networks are difficult to balance accuracy and model size is analysed and studied. A new acoustic encoder is proposed to optimize the extraction of speech features. The effectiveness of the proposed method is verified by an end-to-end speech recognition model. The new acoustic encoder focuses the Global context information with the local information obtained by convolution, Global context information is added. The convolution depth is increased while the convolution kernel size is reduced. It also reduces the amount of parameter calculation. At the same time, RNN-Transducer is used as the model architecture to jointly optimize acoustic features and text features, in order to obtain more effective context information. On the LibriSpeech dataset, the improved acoustic encoder achieves 9.83% word error rate and 4.61% sentence error rate with only 7M model. Compared with the baseline model of convolutional neural network, the word error rate is relatively reduced by 0.52%, and the sentence error rate is relatively reduced by 1.22%.

In this study, we investigate the effectiveness of machine learning models for the detection and mitigation of Advanced Persistent Threats (APTs) in cloud environments, which pose a significant risk to cybersecurity. Using a publicly available APT malware dataset, we evaluate the performance of Random Forest and Support Vector Machines (SVMs) models. Our results demonstrate that both models achieve high accuracy scores, with the Random Forest model achieving a low mean squared error. We present the results of ROC analysis and cross validation scores for the Random Forest model, which further demonstrate its potential for APT detection and mitigation. Our study highlights the significant potential of machine learning-based approaches for improving cybersecurity in cloud environments. However, further research is necessary to evaluate the performance of both models on larger datasets and in different scenarios. To enhance the accuracy and effectiveness of APT detection and mitigation, future work will focus on investigating other machine learning algorithms and techniques, such as deep learning and natural language processing. Overall, our findings provide a promising starting point for further research in this area, emphasizing the potential for machine learning-based approaches to enhance cybersecurity in the cloud. By leveraging these advanced techniques, we can mitigate the risks associated with APT attacks and better protect sensitive data and information.

Until now, the video de-blurring is still a challenging task. In many video de-blurring tasks, high-quality face images in videos is particularly important. However, due to the mutual motion between the camera and the subject during video recording, motion blur often occurs, especially in some frames the blurring of face is particularly serious. But before or after this frame, it may be possible to obtain different local information of the blurred face. For this reason, this paper proposes a new video face de-blurring method based on GAN prior. First, we learn the GAN prior network that generates high-quality face images, and embed it into the whole network as a prior module. Then, residual dense blocks are used in the RNN cell to extract the spatial features of the face in the current frame. In addition, a fusion module is used to fuse the hierarchical features of past and future frames to de-blur the current frame and generate a low quality low fuzzy face. Finally, this face is sent to a trained prior generator to generate a high-quality face. After a series of experiments, the results show that the de-blurring effect of the method we proposed is better than the existing algorithm.

The current axis temperature discrimination system is gradually emerging a method of axis temperature status discrimination combined with neural networks. The principle is to train a neural network model with the existing axis temperature data and use the output of this model to determine the current axis temperature status. However, because the data used for the training of the axis temperature state discriminative model are less for the strong thermal class, the scale of the model used to discriminate the axis temperature state class cannot be too large, so the model discriminative results are often not particularly satisfactory. In this study, the temperature rise, column temperature rise difference, and vehicle temperature rise difference are used as feature data. Using the optimized generation network generates simulated data with probability distribution similar to that of the real strong thermal class feature data. And the T-SNE visualization and the discrimination of the actual axis temperature characteristics are used to prove the validity of the generated strong thermal class features data. The simulation results show that the generated axis temperature feature data of the strong thermal class is valid and can be used for the subsequent training of the axis temperature discriminative network.

The structure of wind turbine is complicated and the operating environment is harsh, which leads to the high probability of failure of wind turbine, and the maintenance is more difficult. Once an accident occurs, it will force the wind turbine to shut down, causing huge economic losses. In addition, it will affect the safe and reliable operation of the whole power system. Therefore, fault diagnosis of wind power generation system is very important. The traditional fault diagnosis algorithm can only identify the known types of faults in wind turbines. If a new fault category appears in the wind turbine generator, the traditional fault diagnosis algorithm can only identify it as a known fault category. To solve this problem, a new classification fault diagnosis method based on semi-supervised deep learning is proposed. The benchmark model of offshore wind turbine is built in FAST software, and 10 different types of faults such as sensors and systems are simulated through model simulation. Different responses of fault signals are analyzed, and 15 parameter signals are selected as input. Firstly, fault features are extracted through multi-scale convolutional self-coding network. Secondly, the initial model is established by using the compressed feature and the error feature map as the input of the classifier and the detector respectively. Finally, the detector judges and puts the new category of fault instances into the buffer. When the buffer reaches the maximum value and starts to overflow, the algorithm starts to update, so as to realize the diagnosis task of new types of faults. Compared with the experimental results of isolated forest, local anomaly factor and single-class support vector machine, the accuracy of our proposed method and other indicators are significantly better, with an accuracy of 99.2%, which can effectively solve the problem of new class fault identification of wind turbines. It can effectively solve the problem of fault identification of the new type of fan, and is conducive to avoiding the occurrence of shutdown accidents, thus maintaining the safe and reliable operation of the power system.

In recent years, the continuous increase in the number of renewable energy sources and users has led to an increase in the load on the electricity market. In this environment, the personal information data and management flexibility cannot be guaranteed in a market that integrates producers and consumers. It seems that the new energy market will face many challenges, such as how to price the output power of power plants according to different energy sources, which price that consumers can sell to other market users through self-generated electricity, how to communicate with consumers and producers in order to maintain the balance of the market and encourage consumers to reduce electricity consumption, how to calculate the market load in order to reduce carbon emissions and protect the environment. In the face of many of the above challenges, it is beneficial to introduce blockchain technology into the energy market because of the unique characteristics, which effectively reduces the cost of maintaining the market and enhances the flexibility of the market. But different market users are sensitive to market information with the advent of information transparency. Therefore, how to restrain the user behavior of the market and maintain the stability of the market will become a major problem. In this paper, we propose a market-based access control strategy to construct a blockchain framework and design an access control-based smart contract to constrain the behavior of each user. In order to further strengthen the relationship between producers and consumers, the concept of prosumer is used to enable consumers to participate in trading activities as energy producers. The experimental results are proved that the proposed solution works successfully in the blockchain, with timely responses between prosumers and consumers, good execution time and the admission of access control provides better market efficiency.

Multivariate time series anomaly detection aims to identify time periods in time series data that are abnormal or deviate from historical data patterns. The increasing dimension of multivariate time series data and the complex temporal dependence relationships make anomaly detection more difficult. Therefore, dealing with the temporal and spatial dependence of the time series data becomes the key to anomaly detection. In this paper, we propose an anomaly detection model GNN-GRUAD based on Graph Neural Network and Gate Recurrent Unit, which can obtain the graph structure of inter-feature dependencies without relying on a priori knowledge, and use it together with the temporal data as the input of the graph neural network to capture the data's spatial dependence relationships. The model also incorporates GRU to model the temporal dependencies of the data, to finally predict the data and achieve anomaly detection based on the prediction error. This paper compares with several other baseline models on the SWaT and WADI datasets, and the model in this paper achieves good anomaly detection performance with an improvement of 0.37% and 7.02% compared to the best baseline on the two publicly available datasets, respectively.

With the development of the Internet of Things (IoT) and the intersection of electrification, networking and intelligence, the risks of vehicle operational safety, data security and network security are intertwined and overlapping, and the security situation is more complex and severe. Thumbur propose a certificateless aggregated signature for vehicular ad hoc networks (VANETs) to solve the authentication and efficiency problems in vehicular networks, and prove the unforgeability of the scheme based on the elliptic curve discrete logarithm problem. In this paper, we propose a new certificateless aggregated signature scheme based on this scheme and prove its security under the random prediction model. Finally, we compare and analyse the computational and communication efficiency of the scheme and show that the computational and communication costs of the scheme are lower.

We propose a format-preserving encryption scheme that takes the field as the unit of encryption for sensitive data in a database. First of all, we need to obtain the data type and length of plain text, and then convert plain text to an array of integers. Moreover, the proposed scheme in this paper is used to encrypt and decrypt the array of integers. Finally, the data after encryption or decryption is converted to cipher text with the same format as plain text, and then is stored in the database. The proposed scheme first uses ZUC stream cipher to whiten, makes the plain text more confusing, and then uses variable block cipher based on the Lai-Massey structure to encrypt, which increases security while preserving the data format. We present the experimental results of the encryption of sensitive data in a database using AES, FFX, and the proposed format-preserving scheme respectively. The results show that the proposed scheme in the MySql database can be performed correctly and efficiently. And the proposed scheme roughly has the same performance as the AES algorithm. As well as it is approximately 5 times faster than FFX while encrypting 1000 records, the more records, the more efficient it is.

In the era of big data, data has gradually become an important asset. In order to fully protect the value of data and ensure the effective conduct of data transactions, this paper combines bulletproof and blockchain to ensure that data transactions can be carried out in a safe and reliable environment, while also hiding the value of data itself. The commitment of bulletproof ensures that all users on the blockchain can verify the validity of the transaction but not the details of the transaction. In this paper, the bulletproof is simulated, and the simulation results verify the feasibility and practicability of the scheme.

The original energy trace dataset have low signal-to-noise ratio which extremely seriously affects the efficiency of side channel attack. For features in energy analysis attacks, this paper proposes a noise reduction method solving the problem that the traces mainly focus on desynchronization and gaussian noise. By combining the low-pass filtering and the DAE model used in this paper, the energy trace is reduced to achieve a high signal-to-noise ratio and expose the POI position of the leakage point clearly. As the image indicates, this work increases the signal to noise ratio of the noise energy trace by 75.9%.

In the information era, rumor detection has gradually become a technology that people urgently need. Rumor detection with a single feature dimension is often not suitable for more complex scenarios. This paper proposes a model based on time series-propagation structure fusion, which can achieve better performance by extracting the characteristics of rumors in the time dimension and the potential features in the propagation process. Experiments are currently carried out on two public Twitter datasets, and the results show that the fusion model in this paper has achieved better performance than the previous method.

Due to the sporadic alienation of the underwater sensor, the node deployment is very scattered, and the physical safety of the node is difficult to get guaranteed. Traditional solutions to ensure security encryption algorithms will increase energy consumption. The reliability of traditional encryption solutions based on difficult mathematical problems with difficulty mathematics issues is getting lower and lower. In order to ensure the secure transmission of communication data, a physical layer security solution is proposed. This scheme through the legal communication between the two parties to develop channel detection sequences, extract channel status information, obtain channel feature value, pre-process the feature value, and use the improved CQG algorithm to generate key. Finally, key negotiation and privacy enhancement are performed on the key. The simulation results show that compared with the traditional CQG algorithm, the improved CQG algorithm can make the key generation rate significantly improved when the key inconsistent rate is slightly reduced.

Through information system, the virtual power plant realizes the aggregation and collaborative optimization of distributed power, energy storage, controllable load, electric vehicle and other distributed resources. As a special power plant, it can participate in the coordinated management system of power market and grid operation. China's power grid is in the transition period of new and old power systems, and power transmission will change from "generation, transmission and distribution" to "source network load and storage". After the establishment of the dual carbon target of the Chinese government, the addition of distributed energy and energy storage has made the original power system more complex. Through the analysis of actual production problems, this paper innovatively proposes the research of key technologies for the operation of virtual power plants to realize the collaborative management, cluster control and optimal operation of multiple distributed energy sources, thus serving the power grid's absorption capacity of green power, and reducing the impact of wind power, photovoltaic and other new energy source output fluctuations on the power grid.

As the number of users using Android continues to grow, the Android app market is becoming the dominant platform in the smartphone market and the correct use of Android app permissions is one of the key factors in the success of Android apps. When installing or running their apps, users need to agree to these permission requests. However, in practice, despite the Android platform introducing official documentation on how to use these permissions properly, there are still many cases of permission abuse. The average user does not have any tools to detect excessive permissions in apps, so most users grant any permissions declared by the app, which exacerbates the risk of private information leakage. Although some previous studies have attempted to address the issue of granting excessive privileges, their methods are now inapplicable due to advances in app protection technology and updates to the Android system. Most existing permission testing techniques are generally based on application development documentation or user decision feedback data, but application development documentation is generally missing, incomplete in description and unreliable in user behaviour data mainly due to its large number of irrelevant characters and non-standardised language text. Many misuses have been reported, and these occur even for the most popular permissions This paper presents a static analysis of permissions based on the classification of apps in the Android App Store.

In order to solve the problem of low classification accuracy of the cloud platform anomaly detection model, this paper proposes a cost-sensitive random GBDT anomaly detection method, considering the different classification costs of different types of data. In this paper, a cost- sensitive stochastic GBDT-based anomaly detection method is proposed based on considering the different classification costs of different categories of data. The method first introduces the cost matrix into the design of the loss function to establish a cost-sensitive loss function to reduce the classification cost of different categories of data; and in order to reduce the computational effort and improve the computational speed, the idea of stochastic gradient boosting is introduced into the GBDT method. And the effectiveness of the method proposed in this paper is proved by simulation experiments.

Microblog short text usually contains rich emotional information. It is a hot research topic in network data mining to grasp the dynamics of network public opinion through microblog emotion analysis. In order to improve the effect of Chinese microblog sentiment analysis, this paper first uses word embedding technology to quantify microblog short text from high dimension to low dimension vector space; Then, the global features of microblog data are extracted through BiGRU, and the Attention mechanism is introduced to obtain important features to build a Chinese microblog emotion analysis model. The feasibility and superiority of the model were verified with the public data set released by SMP2020. The accuracy, recall and F1 values of the model reached 78.65%, 78.57% and 78.41% respectively. The experimental results show that the feature vector of BiGRU combined with attention mechanism contains more rich emotion information of short text of microblog, which can effectively improve the performance of sentiment analysis of Chinese microblog. The experimental results show that the feature vector of Bi-GRU combined with the attention mechanism contains richer semantic information of the text, which can effectively improve the performance of emotion recognition of online public opinions.

This paper proposes a text classification model based on a bi-directional LSTM network with attention mechanism and subsampling in the word vector stage. Firstly, we use the Skip-gram model in Word2Vec for feature extraction, and then combine the two-way LSTM network with subsampling to extract and classify the key semantic information in the text, and integrate the Attention mechanism to optimize the generation weights, enhance the feature transfer between layers, and focus attention on the high weight words. Experimental results on different categories of open datasets show that the proposed model improves the recognition and extraction of high-featured content compared to a single network, and can effectively improve classification accuracy.

In Dunhuang mural image restoration, image enhancement techniques have effectively helped in image restoration. Based on pulse-coupled neural network (PCNN) has been widely used in image processing, in order to solve the low-lighting problem of Dunhuang mural images, on the basis of FC-MSPCNN model, the parameters such as synaptic weight matrix 𝑊_ijk1, link strength 𝛽, attenuation factor 𝛼 and attenuation adjustment parameter K are redefined in combination with adaptive parameter setting method, and the Performed-FCMSPCNN (PFC-MSPCNN) model. Finally, the linear transform, gamma transform, and histogram algorithms are used for image enhancement and compared with the PFC-MSPCNN model, respectively. It is verified that the PFC-MSPCNN model in this paper has a good enhancement effect on low-light images.

Most popular image enhancement algorithms are generally based on a series of specialized images collected by image photography devices. Hereinto, Pulse-Coupled Neural Network (PCNN), plays important roles in image enhancement aspect, with lower computational complexity and higher image enhancement accuracy. On the research, we propose an image enhancement method based on enhanced fire-controlled MSPCNN(EFC-MSPCNN) model, which gives the setting methods of designed adaptive parameters. Related experimental results demonstrate that our proposed method has good image enhancement performances.

In this paper, a secure and efficient access authentication scheme is designed between edge devices in the smart grid, to achieve secure communication between devices in the smart grid. Considering the characteristics of power grid equipment, identity based cryptography is introduced to embed the equipment identity, timestamp, etc., into the process of identity recognition and private key update. Combined with the characteristics of private key update in the cryptographic system, bilinear pairing is used to negotiate the communication key in each time interval, and only hash, XOR operation and concatenation operation are used to achieve lightweight authentication scheme. The scheme can not only meet the security requirements, but also minimize the burden on edge device resources. Experimental results show that the communication cost and computation cost of our scheme are less than the existing schemes.

Nowadays, DSP processors have been widely used in wireless network systems, since it can meet high-speed real-time information processing requirements. Meanwhile, the undocumented instructions in processors, with potentially catastrophic consequences, have become one of the main threats to system security. Therefore, the mining of undocumented instructions in processors is quite crucial for reinforcing the security of wireless network. To achieve this goal, we propose an undocumented instruction search method specifically for DSP processors, which can efficiently and accurately obtain undocumented instructions. First, to compress the instruction search space, we analyze the characteristics of DSP instruction format. Second, we build a precise instruction disassembly framework to identify all the undefined instructions. Finally, we build an automatic test tool for the executability of undefined instructions. We perform extensive experiments to evaluate the effectiveness of our method. The results reveal that our approach is valid for mining undocumented instructions in DSP processors. Specifically, we have successfully mined 335 instructions on 14 TI DSP processors from 6 different series.

The security of deep neural networks has triggered extensive research on the adversarial example. The gradient or optimization-based adversarial example generation algorithm has poor practicality and cannot combine high success rate and high efficiency; using GAN to generate adversarial examples has the problems of gradient disappearance and unstable generation. In this paper, we propose a novel adversarial example generation algorithm based on WGAN-Unet, which uses the structure of WGAN and Unet to form a generative adversarial network to improve the stability of network training, and uses the cosine loss function to measure the category loss and improve the success rate of adversarial attacks. The adversarial example generation using WGAN-Unet is compared with other algorithms in terms of human eye perception, time consumption, attack success rate, and image quality, proving our scheme’s superiority.

In real life, whether it is in the fields of weather forecast at airports, traffic control, weather services at ports, or industries such as agriculture, aquaculture, and transportation, all need to know the real situation of the weather. If it cannot be accurately identified and judged in time, will lead to threats to human life and property safety. In order to reduce the impact of weather on people's production and life in various industries and improve the accuracy of weather recognition, this paper establishes an improved convolutional neural network model based on the AdaBoost algorithm to identify three types of weather: rainy, snowy, and foggy. First, the convolutional neural network is trained and iterated many times, and then each convolutional neural network is used as a weak classifier, and multiple weak classifiers are weighted and combined based on the AdaBoost algorithm to form the final strong classifier model. Experiments show that the strong classifier optimized by the AdaBoost algorithm has a stronger ability to identify misclassified data and can realize dynamic reinforcement learning of data.

As the core component of the subway running gear, the safety and reliability of the bearing has been widely concerned. Under the influence of high speed, heavy load, long routing, sand erosion, rain corrosion and other operating conditions, the service conditions of bearings become very harsh, and damage inevitably occurs. At present, most of the nondestructive testing technologies used in metro bearings are black magnetic particle testing technology, which has the advantages of high sensitivity, high reliability and easy identification of defects. It is one of the most effective technologies widely recognized for the effective inspection of the surface or near-surface of ferromagnetic materials. To detect the surface defects of rolling bearings of metro vehicles, the images of black magnetic particle flaw detection defects of metro vehicles’ rolling bearings are collected on the spot. By using image processing technology, all binary images of surface defects are extracted, and a data set including the original image and binary images of surface defects is established. It is pointed out that there are great difficulties and limitations in using traditional machine vision technology to identify subway bearing flaw detection defects. Thus, convolutional neural network is chosen to identify bearing defects to realize deep feature fusion. Through the comparison with several models, it is concluded that there is a certain improvement in the effect and accuracy of surface defect identification in this paper, which can not only reduce the labor intensity and training costs of personnel, but also ensure the efficiency and accuracy of bearing defect detection identification. It is the trend and hot spot of future development.

The session-based recommendation (SBR) is a task to predict user next action based on anonymous behavior sequences. Most existing studies utilize global context in conjunction with current session information to improve recommendation results. However, the local and global perspectives have an inherent conflict, and it has been a great challenge to integrate the two types of information. In this paper, we propose a novel recommendation method, namely, dual-grained global graph neural network (DG-GNN), to capture collaborative information from all sessions in a more subtle manner to model current user preferences. Specifically, we convert the local and global content into two graphs and process them with the corresponding encoders. The local section exploits the graph attention network to learn a set of item-level embeddings. The global section collects the pairwise item transitions relevant to the current session for constructing graphs and exploits a lightweight encoder to learn a single session-level embedding, which reduces the influence of irrelevant information from global content on the local section. Moreover, a new position embedding mechanism is proposed to generate different position embeddings depending on the length of sessions for enhancing the ability to capture position information. Experiments on real-world datasets demonstrate that DG-GNN outperforms other state-of-the-art methods.

The rapid development of new energy vehicles in recent years has accelerated the process of smart and networked cars in the automobile industry. Correspondingly, cars are also facing more and more network security threats. However, most of the current car internal communication protocols are still CAN or CAN FD, which were designed without considering communication security issues. Therefore, the ECU inside the car is unable to identify the reliability of the source of the information. In order to help smart vehicles equipped with CAN or CAN FD protocol to better deal with security threats, this study proposes a device fingerprint recognition method based on the root of spectral amplitude share (RSAS) of signal from a physical security perspective and builds a high-fault tolerance signal identity recognition system based on it. The experimental results on CAN FD devices show that our system can achieve near 100% accuracy in identifying signals with just 6.8% of the data in the DATA field of a single frame, which is superior to the result of similar studies. Overall, our method can achieve simple and efficient signal identity authentication without changing the existing communication protocol and with very low computational costs. This lays a foundation for the further development of physical security technology in the field of automobiles and has important practical significance for the construction of a secure device communication system in automobiles.

Aiming at the increasingly serious problems of phishing websites and the poor characteristics of identifying phishing URL, a phishing URL recognition model based onordered neurons long short-term memory neural network( ON-LSTM) and attention mechanism feature fusion is proposed. Firstly, the URL is divided based on the word level and character level of sensitive words and special characters, and then the URL vector after word segmentation is input into the onlstm model to realize feature extraction. In order to enrich the feature information, the feature fusion layer adopts the feature splicing method to fuse the features of sensitive word level and character level, and introduces the attention mechanism to further enhance the feature ability and depth of the module to extract the URL. Finally, the softmax function of the fully connected neural network is used to realize the classification and recognition of URL. The experimental results show that the phishing URL recognition model based on the fusion of ON-LSTM and attention mechanism features has better ability to identify phishing URLs, with an accuracy of 98.95% and an F1 value of 98.92%. Compared with the comparison model, this method has better recognition performance.

Facing the pressure of the extremely inflated network information and data overload surplus, it is extremely important to locate “valuable information” efficiently and accurately. Text summarization technology in the field of Natural Language Processing (NLP) is an effective means to analyze and process network information. In this paper, we propose an abstractive text summarization model based on bidirectional encoder representations from transformers (BERT) vectorization and bidirectional decoding. The BERT is adopted to obtain a more global vector representation, which helps the subsequent encoder and decoder to fuse the full-text information to generate a summary with high generality. The decoding phase adopts a bidirectional decoding structure and combines the attention mechanism to maintain the bilateral decoding result to generate summaries. The bidirectional decoding structure can be fine-tuned according to the bidirectional results, which can overcome the tilt problem of the unidirectional structure, and the generated summaries are more consistent. The experimental results on the NLPCC2017 text summarization dataset show that the summaries generated by our model have the higher coherence at the word and sentence level, and the stronger generalization of the full text.

Most of the existing malicious webpage detection technologies are single-mode, without considering the internal relationship between the content of the webpage and the image. In view of the above problems, this paper proposes a malicious webpage detection technology based on Transformer-MDA-ESIM multi-modal fusion. Use the BERT model and the VGG16 model to extract the feature vectors of the content of the web page and the image. The two extracted feature vectors are input to the Transformer layer for encoding, which solves the problem of difficulty in capturing long-distance dependencies. The text semantics is enhanced by mining the connection between the content of the web page and the image through the dual attention mechanism of the semantic enhancement layer. The ESIM interactive attention layer is used to obtain the similar features of text content and images, realize the fusion of text features and image features and improve the model's ability to understand web pages. The experimental results show that this method has better performance in malicious webpage identification compared with single modality malicious webpage detection technology.

The services required by wireless network users are generally based on the location requirements. How to accurately locate the users in the wireless network and select the network which can provide the best service for access is the key to improve the user experience. At present, accurate positioning can be achieved in outdoor environment, but the accuracy of indoor positioning method needs to be improved. In addition, how to achieve effective switching in different networks according to users'needs during the mobile process is also a major difficulty. This paper collects user positioning information in wireless networks by studying high-precision indoor positioning algorithms and effective mobility management strategies. Thus, a fingerprint positioning model based on RSSI is built on the premise of guaranteeing good service quality for end users. RSSI of each access point in the actual location area is collected to form an offline fingerprint database. At the same time, an improved extreme learning machine algorithm location model based on particle swarm optimization is proposed to match and locate RSSI in user and fingerprint libraries. Finally, the simulation results show that the positioning scheme proposed in this paper can effectively reduce the positioning error compared with the traditional positioning algorithms.

The underwater environment of the sea bottom is complex, and it is easy to be affected by light, underwater impurities and shooting angle. Because of the dense and mutual occlusion of underwater targets, the collected images are not clear, fuzzy and easy to be resolved, which affects the accuracy of target detection and recognition. Aiming at the above problems, this paper proposes an improved YOLOV5 method for underwater seafood target detection. Firstly, the C3 module of high-level features is fused with swin transformer to create the C3STR module, and the improved C3STR module can enlarge the receptive field of the network, obtain the global information better and increase the ability of feature extraction, the DC3 module is constructed by adding convolution to the low-level feature C3 module, which improves the model detection ability with a little more computation. Finally, the network feature fusion module is improved, which integrates the deep-level information with the shallow-level information, the combination of semantic information and position information can improve the precision of model recognition and enhance the ability of target detection. The experimental results show that the detection accuracy of the improved YOLOv5 algorithm is increased to 88.3%, which is 2% higher than that of the original YOLOv5 algorithm.

Aiming at the problem of poor detection results due to the large difference in scale of detection targets and numerous small targets in remote sensing images, a remote sensing target detection algorithm DCB-YOLO based on atrous convolution and multi-scale feature fusion is proposed. Based on the YOLOv5 algorithm, the algorithm designs a bidirectional feature pyramid structure (MDCB-BiFPN) with multi-branch hole convolution. The MDCB-BiFPN structure can better extract and fuse features of different scales, thereby improving the model's ability to detect multi-scale remote sensing targets. The experimental results on the remote sensing public dataset DIOR show that the DCB-YOLO algorithm can better accurately identify remote sensing targets at different scales.

As an important research topic, text classification is often deeply studied by scholars. Convolutional neural network is a commonly used deep learning algorithm for text classification. However, CNN will lose some information features in the pooling process due to the limitation of fixed-size convolution kernel when extracting features from complex semantic information. In order to avoid CNN's problems in text classification tasks, Capsule network is used to extract text features. CapsNet model has the problem of redundant information transmission. In order to reduce redundant information, a capsule screening network is proposed in this paper to remove the weak connection between the screened child capsule and the parent capsule, optimize the dynamic routing algorithm, and reduce the redundant information transmission of the capsule layer. In addition, in the classification of text, the text may exist, such as less useful information, serious colloquial language, unable to reflect the importance of words and so on. In order to improve this problem, this paper introduces the attention model. Combined with improved CapsNet to handle text sorting tasks. Three experimental results show that CapsNet with improved mechanism of attention is the best in text classification task.

The task of dense video captioning is to generate descriptions for multiple events in an untrimmed video. Previous works generate descriptions for each event, and then concatenate the individual sentence descriptions into a final paragraph, often leading to inconsistency and redundancy problems. To overcome these limitations, we propose a context-aware model that leverages context event information from both past and future events to provide conditional dependence on the current event description. Moreover, context event information is encoded by a strong pre-trained context encoder and is incorporated into the captioning module through a gate-attention mechanism. Experimental results on both YouCookII and ActivityNet datasets demonstrate that our model outperforms the existing context-aware and pre-trained model by a margin.

There are many excellent community detection algorithms in the field of overlapping community detection, which can respond well to the needs of various scenarios, traditional community partition method due to excessive ignored the high computational cost brought by the use of global information, in the face of increasingly large scale appear struggling, this article hopes to local information using traditional community partition algorithm, combined with seed development principle, and experiment with human networks and real network data to get better community results.

For time-sensitive and large-scale text information such as news, traditional abstract summarization methods may lead to deviation from the subject, distortion and incompleteness of information. To this end, we propose an improved scheme based on automatic text summarization technology, which combines the characteristics of BERTSUM(BERT-based Text Summarization) and BART (Bidirectional and Auto-Regressive Transformers) models. Specifically, we input the same text information into the BERTSUM and BART models respectively, and obtain the key sentence granules of relevant text information through the BERTSUM model and the summary text processed by the BART model, and combine the processing results of the two models into a mixed data Set, and generate a summary through the BART model again to get the final summary result. The effectiveness of the BERTSUM-BART method was verified through experiments based on news long text data. The average recall rate of ROUGE-1, ROUGE-2 and ROUGE-L reached 39.83, 18.05 and 36.46.

Automatic fake news detection is crucial for society. Existing methods mainly focus on the post's content or taking advantage of external sources to make a decision. Recently a new approach called NEP has been proposed, it constructed out news environment for each news to capture popularity and novelty as other evidence. However, NEP neglects the changeable environment and it is weak when news breaks out or is published several times. Commonsense knowledge related to a post has sufficient reliability compared with the news environment or external news evidence, and it can bring a stable benefit for distinguishing fake news. Based on this, we search out commonsense knowledge for each post and propose the News Environment-Knowledge Perception (NEKP) based on NEP. For each post, we search out the related knowledge items on Wikimedia. Then we fuse this knowledge through an existing fake news detector–DeClarE. Finally, we fuse the news environment, knowledge, and the news itself to make a detection. Experiments on NEP datasets show that commonsense knowledge is another helpful piece of evidence.

Falls have become an essential factor affecting life safety in recent years. Given the high energy consumption associated with running models over extended periods in practical engineering, this research presents a network architecture, Spiking-Rep-YOLOv3, which is more suitable for transformation into Spiking neural networks (SNNs). We redesign the RepVGG-based backbone network by improving the YOLOv3-tiny network. Second, we speed up inference by employing neuron normalization to reduce information loss during inference after neural network transformation. We experimented with Spiking-Rep-YOLOv3 on the E-FPDS public dataset and showed that the network achieved similar accuracy as YOLOv3-tiny and only 40.7% of the energy consumption of the latter.

With the rapid development of graph neural network technology, its application in the field of natural language processing is more and more extensive, text classification is one of the important applications, everyday life will produce a large number of non-Euclidean text data, while the traditional classification methods in the graphic structure of text data has been a great challenge. Graph convolutional neural network(GCN) is considered to be able to model the structural attributes and node feature information of graphs well, and is gradually becoming a good choice for text classification of graph data. This paper proposes a text classification model based on graph convolution network and neural network local enhancement. On the basis of using GCN to extract features, Bi-LSTM method is used to balance the experimental results, enrich the feature information by capturing local information, integrate the attention mechanism, and fuse the evaluation values to improve the accuracy of classification. It is verified that this method has achieved better results than the existing classification methods in many classical data sets such as 20NG and OHSUMED.

As a traditional Chinese flower, chrysanthemum has the advantages of long flowering period, many varieties, wide application, and bright colors. However, different varieties of chrysanthemums require different growth environments and have different ornamental and medicinal values. Therefore, the fine-grained and accurate identification of chrysanthemum varieties has important guiding significance for the breeding and cultivation of chrysanthemums. Finegrained image recognition is a more detailed subclassification of coarse-grained large categories. Due to the subtle interclass differences between sub-categories, compared with general image recognition and classification, it is necessary to use a deeper and more complex network structure to extract more complex and higher-order features of images to improve the the accuracy of fine-grained classification of chrysanthemum images. However, the complexity of the network structure will lead to a large number of parameters of the convolutional neural network model, long training time and weak generalization ability. In order to solve the above problems, this study proposes an improved multi-scale, multi-parallel convolutional neural network, VGG-Inception, based on the traditional VGG-16 convolutional neural network. This model increases the parallel network structure and widens the network while ensuring the depth of the network. VGG-Inception uses the global pooling layer to replace the full connection. While ensuring the accuracy of model classification, the number of model parameters is reduced to 9% of the original VGG-16. At the same time, the auxiliary classifier is used to conduct the gradient forward, avoid the gradient disappearance, and improve the generalization ability of the model.

Aiming at the problems of complex feature extraction process and low recognition accuracy of traditional methods in infrared fire image recognition, an infrared fire image recognition algorithm based on ResNet50 and transfer learning was proposed in this paper. Firstly, the infrared fire image was normalized, and the sample was expanded by data enhancement operation. Then, all parameters of ResNet50 convolution layer and pooling layer were frozen by transfer learning method. The top layer network was designed as convolution layer, discard layer, batch standardization layer and full connection layer, and the infrared fire image recognition model was constructed. Several groups of comparative experiments were carried out with CNN, VGG16, ResNet101 and ResNet152. The experimental results show that the algorithm was proposed in this paper had a recognition accuracy of 99.63% on the infrared fire image test set, which was much higher than other algorithms. The recognition rate on the visible light fire data set was almost close to 99.82%. The effectiveness of the proposed algorithm in infrared fire image recognition had been verified.

Object tracking is widely used in daily life. The existing visible camera-based tracking methods are difficult to track objects accurately in the face of degraded scenes such as fast movement, high contrast, and low illumination, which will cause the loss of tracking performance. An event camera is a biologically driven sensor with higher dynamic range, smaller time delay, and higher light sensitivity than traditional visible cameras. We propose a siamese network object tracking method that fuses visible and event cameras to realize more reliable tracking. We design a feature fusion method combining visible and event camera features based on attention mechanism. It realizes the cooperation of both sensors and improves the accuracy of tracking. Experiments on the VisEvent dataset reveal that our network outperforms single-modality trackers by 5.8% and outperforms other fusion-based methods by 2.1%.

Smoking is prohibited in many places. In order to detect smoking behaviour and make accurate judgments in a timely manner, machine vision technology is used to carry out research on the detection of small targets of cigarette sticks to improve the problem of difficult and inaccurate recognition of small targets in current detection algorithms and to improve the accuracy of cigarette sticks being recognized. In this study, we propose an improved method to detect smoking behaviour of Yolox based on the attention mechanism, in view of the fact that cigarette sticks occupy a very small part of the image and the presence of hand occlusion. The Attention mechanism module is added to the feature extraction network to focus on local information. Meanwhile, the use of the deep network is increased by adding a scale to focus attention within the target region; and the loss function is optimally replaced and the GIoU loss function is chosen to solve the drawback problem of IoU; experimental results show that the improved Yolox-c algorithm mAP reaches 95.96%, compared with Yolox and Yolov5 mAP, which are improved by 5.87% and 8.85%, respectively.

Combining the posture recognition algorithm with Taiji to realize the movement recognition of it. Acquire the video key frame, detect the target of the standard posture of Taiji through the deep learning method, use the posture recognition algorithm OpenPose to extract the coordinates of the key points of the human body, use PAF to calculate the connection of the key points, through the double judgment of angle and distance, select the action with the smallest error from the standard action, which is more than the previous single judgment ,The accuracy rate of each movement has increased by about 4% on average, achieving the expected effect.

At present, there are distribution differences in the public data set of pedestrian attribute recognition, and the sampling scenarios are diversified, which leads to the reduction of the prediction accuracy of the pedestrian attribute recognition algorithm (PAR) in the cross-data set and the limitation of its application in the actual scene. For the appeal problem, this paper proposes a multi-source and multi-label pedestrian attribute recognition method based on domain generalization (DG). First, label multiple data sets as source domains through label alignment to determine the common attribute labels in each source domain. Secondly, add the countermeasure data enhancement module (ADM) to generate the data distribution of the target domain based on the difference of data distribution in the source domain. This model can predict the data in the unknown target domain, thus optimizing the performance of pedestrian attribute recognition (PAR) in cross-domain and actual scenarios. The model was tested on the public data sets PETA, RAP, PA-100K. Compared with HPNet, LGNet, PGDM, StrongBaseline and other models, the average recognition accuracy (mA) index and F1 value increased by 1.16% and 1.51% respectively, which proved the effectiveness of the model in pedestrian attribute recognition.

Few-shot image recognition is a challenging task that aims to train a model with limited labeled images for image classification. To overcome the issue of insufficient image information in few-shot learning (FSL), this paper proposes to incorporate textual semantic information for guiding knowledge transfer to achieve more robust models. However, existing multi-modal guidance methods often necessitate aligning feature spaces from different modalities, which not only adds extra model parameters and training costs but also increases overfitting risks in FSL. To address these issues, this paper proposes a multi-modal method based on ranking instead of training, called Ranking-Based Multi-Modal Knowledge Transfer (RMMKT). Specifically, our RMMKT method (1) measures inter-class correlations across different modalities and performs multi-modal fusion based on ranking to avoid the need for training multi-modal feature adapters, (2) utilizes the multi-modal metrics to transfer base classes information to novel classes, which improves the stability of model, and (3) introduces the maximum a posteriori estimation for prototype generation and image classification without training. The proposed method was evaluated on commonly used FSL datasets and achieved state-of-the-art performance. RMMKT provides a training-free transductive few-shot learning strategy that achieves multimodal fusion and knowledge transfer without introducing extra model parameters for novel tasks, thus enhancing model stability and accuracy.

Compressed sensing (CS) algorithms for images based on deep learning offer huge improvements over conventional algorithms to reduce time complexity and improve reconstruction quality. but existing deep compressed sensing (DCS) algorithms still have problems, among which our concern is the loss of detail features in image reconstruction. Based on this, We propose a parallel image foreground- background reconstruction network (PFBNet), and PFBNet combines the image segmentation (IS) domain. The initial reconstruction uses parallel networks to learn the features of the segmented foreground and background images, so each network focuses on learning the main features of the image to recover pixel details. Then, the deep network superimposes the foreground and background images into a complete image while performing error elimination to produce the final image. Experimentally, the PFBNet algorithm demonstrates better recovery performance and preserves the pixel and texture information of the image.

Aiming at the problem that the monolingual corpus (Chinese or English) is affected by the sentence structure in the sentiment analysis task, resulting in a large deviation of the extracted context information, and the traditional model does not consider the dependence between aspect words and context information words, which makes the sentiment classification inaccurate, a bilingual multi-feature sentiment analysis model with the degree of fusion relationship is proposed. Firstly, the attention mechanism is used to calculate the attention relationship degree of the aspect word in the sentence, and the Hadama product operation is performed with the context information to obtain the dependent information of the aspect word. Then, the Bidirectional Gated Recurrent Unit (BiGRU) is used to learn the global serialization information of the sentence to obtain the overall information characteristics of the sentence. Next, through the attention mechanism, the overall information characteristics and the aspect word-dependent features are integrated to obtain single-language information. Finally, through the attention mechanism, the overall information characteristics and the aspect word-dependent features are integrated to obtain single-language information. Experimental results on the public datasets Sem Eval2014 Task4 and AI Challenger 2018 show that the model can obtain sentiment information based on aspect words more comprehensively, and has higher accuracy of text sentiment polarity classification than the comparison model.

Federated Learning (FedL) emerged as a privacy-aware alternative, creating an effective means for multiple data providers to enable collaboration on training models without accessing the original data. Vertical federated learning (VFedL), as a crucial classification within FedL, has always been primarily utilized to train a machine learning model with non-uniform data from different providers. Despite the VFedL's benefits in facilitating collaborative training models while safeguarding data privacy, it remains a daunting challenge to incentivize more valuable data providers to participate in the VFedL due to the absence of scientific data pricing and precise measurement of data contributions from participants in practical operations. In this paper, we construct a scientific data pricing method based on the participants' data contribution score to federated models, so that all data providers can be compensated fairly. Firstly, an accurate measurement method of the data contribution score of each federated participant to the global model is constructed based on shapely values for Monte Carlo optimization. Then, taking the data contribution score as the input variable, we formulate a data pricing game model based on Stackelberg with the hosts as the leader and the guest as the follower in VFedL. We further solve our model and analyze the guest's optimal data usage strategy based on data contribution score and the hosts' optimal data pricing strategy. Our method has been proven through numerical experiments to precisely assess the data contribution score of participants with the Federated Logistic Regression model. These study findings can also offer management direction for the FedL service providers.

The visual SLAM algorithm based on the assumption of static environment will be affected by the motion feature points in dynamic environment leading to the degradation of system robustness and accuracy. To address this problem, a visual SLAM algorithm based on instance segmentation to remove dynamic feature points is proposed. The Yolact is used to detect potential motion objects, and the semantic information obtained is used to accurately reject dynamic feature points. In addition, considering the under-segmentation problem of the instance segmentation network, this algorithm combines the semantic information in historical key frames and depth maps to segment dynamic objects that are not detected by Yolact using a region growing algorithm, which further improves the robustness of the system. To ensure the real-time performance of the system, the instance segmentation network segments only the key frames, and improves the key frame selection strategy. The experimental results of the system based on TUM dataset show that compared with ORB-SLAM3, the ab-solute trajectory error and relative trajectory error of this algorithm in dynamic environment are reduced by 95.3% and 96.5%, respectively, the real-time performance of this algorithm is better compared with DynaSLAM and DS-SLAM.

The big data of coal mine was characterized by large scale, many influencing factors and weak correlation. The existing big data mining based on quantitative data analysis usually adopts fixed framework processing, which is easy to fall into the dilemma of overfitting and cannot acquire elastic knowledge across layers. From the perspective of coal mine monitoring and management, a multi-granularity time-varying cloud model is proposed based on a multi-granularity semantic transformation model and a grey prediction model to achieve the conceptual prediction of coal mine gas concentration risk. The new model first converts the gas concentration data into semantic concepts using a semantic transformation model, then processes the sequence of semantic concept sets for randomness and oscillation, and builds a grey prediction model for them, which is used to project future semantic concepts. The new model is compared with the existing model, and the experimental results show that the new method can not only predict coal mine risk concepts at different granularities, but also has high prediction accuracy.

With the diversification and rapid development of the network, the generation of cyber threat intelligence (CTI) is more and more comprehensive, and the knowledge graph of which is also developed and updated constantly. Of course, it is inevitable to encounter the expansion of the knowledge base of threat intelligence, and Intelligence entities are closely connected. In the Knowledge Graph (KG) field, link prediction is widespread because the entities are already available. And you need to determine the relationship between the two entities somehow. However, the more complex problem is extending the entities created after establishing the knowledge base into the existing one. Dealing with these problems often relies on extracting entities and relationships from the natural language of threat intelligence, but this approach only predicts undeclared intelligence entities or relationships. Embedded models often use structural rules to predict link relationships but cannot predict missing intelligence entities. In this paper, the latest MiNer method is combined with the effective STransE model to fully invoke the neighbor information around the entity and analyze the proportion of each neighbor. At the same time, this paper uses neighbor traffic statistics to assign corresponding weights to indicate which neighbor is more important to the entity to be predicted. We show that our method can use the existing cyber threat intelligence KG database to effectively predict some unknown entities, according to the importance of neighbor information, and add them to the existing KG database. At the same time, our method’s performance is greatly improved when the specific matrix is converted to the diagonal matrix.

In this paper, an optimized long short-term memory (LSTM) network is proposed for the remaining useful life (RUL) prediction of the rolling bearings based on whale optimized algorithm (WOA). The multi-domain features are extracted to construct the feature dataset as the single domain features are difficult to characterize the performance degeneration of the rolling bearing. Considering the possible gradient explosion by training of the rolling bearing lifetime data and the difficulties in selecting the key network parameters, an optimized LSTM network, namely, WOA-LSTM network is proposed. Experiment results show that, compared with the LSTM network, the RUL prediction accuracy of the rolling bearing are improved by the proposed WOA-LSTM network.

Whether in land, sea or air transportation, atmospheric visibility is an important factor related to the safety of people's lives and property. How to continuously monitor low-visibility weather and achieve better results, so as to provide indicators for the traffic control department to make corresponding traffic control measures in a timely manner, is of great significance to the driving safety of the people. In recent years, more and more facts are proving that the problem of low recognition accuracy of traffic signs, traffic lights, roadblocks and lane lines in low-visibility environments limits the rapid development of technology for unmanned driving technology. Therefore, there is a growing need to discover methods that can detect visibility efficiently and inexpensively. This paper reviews and summarizes the traditional and emerging visibility detection methods in recent years, analyzes and briefly expounds the advantages and disadvantages of each method. This paper aims to give researchers easy access to new and old technologies in the field and to provide them with basic directions for innovative methodological techniques based on them.

The detection and recognition of traffic signals, which are important elements in the driving environment, are part of the intelligent driving domain, and the results of the detection and recognition affect the driving safety of intelligent vehicles. However, the classic traffic light signal detection algorithm model is too large, which has certain performance requirements on the devices deployed at mobile terminals. In this paper, we propose a structurally reparameterized network structure that can achieve the performance of classical mainstream algorithms and also significantly reduce the number of parameters and operations of the model. Firstly, the C3 structure of the yolov5s backbone network was replaced with MobileOne structure. Depthwise Convolution was used to reduce parameters and operation costs on the neck part of the model. The CIOU which calculated the location loss was replaced with the EIOU which focused on the length loss and width loss of the anchor boxes. Finally, SE attention module is introduced to further improve the characterization ability of the model. Experiments show that compared with the original Yolov5s algorithm, the parameter number and GFLOPs of the improved structure reparameterized Yolov5s decrease by 29.8%. the map0.5 performance index increased by 0.7% and reasoning speed increased by 19.2%, which effectively achieved a more real-time effect on traffic signal light detection and has certain application value. It can provide effective support for the driving safety of smart cars.

The context of emergency plan text is very tightly connected, while fully connected self-attention architecture Transformer has advantages in parallelism and remote context modeling, and Bi-directional Long-Short Term Memory (BiLSTM) can also capture long-distance context information. In this paper, we propose a new model Bi_ComL-CRF, the most important of which is the Com-LSTM unit, by introducing an additional Transformer encoded representation for each recurrent unit, then enhancing the captured context information through the new gating mechanism, so as to improve the effect of emergency plan NER task. We have carried out experiments on the emergency plan data, and the results show that our proposed model is better than other NER models, and can provide help for calculating the text similarity of emergency plans and generating emergency plans quickly.

In this study, the dynamic model and immunity strategy for the risk propagation in complex weighted supply chain networks are investigated. First, the BA scale free network model is improved, and a dynamic complex supply chain network model is constructed. Second, the risk propagation in the supply chain network is simulated based on the traditional SIR(Susceptible-Infected-Recovered) model, and various types of immunity strategies for the nodes in the network are examined and proposed. The results of the simulation experiment indicate that the proposed compound targeted immunity strategy leads to better immunity than the traditional target immunity. Furthermore, the immunity of the network can be improved by adjusting the network structural parameters, and these findings can act as guidelines for evaluating the immunity strategy and controlling the risk propagation in complex weighted supply chain networks.

Aiming at the problems of the basic RRT algorithm in path planning of mobile manipulator, such as undirectivity, repeated search and node redundancy, an improved RRT algorithm was proposed. Firstly, the heuristic search idea is introduced, and the target points are extended to sampling points with certain probability to enhance the target orientation of the algorithm. Secondly, node rejection mechanism is introduced to cover the search area to avoid repeated search to generate new nodes. Finally, the path is pruned and smoothed. The experimental results show that the improved RRT algorithm has stronger search ability, greatly reduced path sampling points, reduced search time, and higher success rate.

With the development of the Internet, online shopping has become the main choice for people to shop. Fresh goods are also gradually turning to online sales mode. Fresh e-commerce APP emerges at the historic moment. According to the user's demand for fresh e-commerce APP interview and research, use the kano model to classify the user demand attributes, and according to the user satisfaction coefficient select the fresh e-commerce APP function design practice, improve user satisfaction.kano model was used to analyze different demand priorities of fresh e-commerce APP to provide reference for functional design and optimization of fresh e-commerce APP.

Most previous studies have analysed the similarity of industrial structures based on the current situation or the past, and have drawn conclusions in stages. Based on the findings of past studies on the impact of industrial structure similarity, and considering the small changes in the industrial structure similarity index in recent years, this paper combines machine learning with support vector regression analysis to predict the impact of the year to be predicted using the industrial structure similarity index for the three years prior to the year to be predicted. The paper uses GDP as an indicator reflecting the impact of similarity of industrial structure. The experimental results show that the impact of similarity of industrial structure can be predicted to a certain extent and is very close to the true value, which is a guide to the healthy and sustainable development of cities.

Appearance-based methods with deep learning can predict the point of gaze by using a monocular camera, which requires a large sample to learn. However, existing appearance-based gaze estimation methods with deep learning mainly use face and eye images or only use a single face image, ignoring the correlation between face features and eye features In response to this issue, we propose a coordination model where face feature extraction is the gaze estimation network and eye feature extraction is the coordination network, which deeply fuses the eye-face feature relationships to perform the gaze estimation task. The model achieves good results on MPIIFaceGaze dataset and GazeCapture dataset.

For the wordle game, we mainly want to determine whether its daily words have an impact on the proportion of people participating in the difficult mode. We need to predict the distribution of the number of attempts in the future difficult mode, so we can use neural network to establish the MIMO(Multiple-Input-Multiple-Output) BP neural network model. This model encodes words as input. We set the transformation function of the model as tansig and purelin, the integration function as trainlm, and the learning function as learngdm. Then we set the number of training to 1000, and the mean square error of training is 0.005. After that, we can train the model and draw a conclusion.

Revenue in the live streaming industry depends largely on the extent to which information is disseminated. Opinion leaders in social networks tend to speed up information diffusion. This work construct a live streaming e-commerce model based on evolutionary game theory to explore the internal mechanism of information dissemination in social networks where opinion leaders exist. The process of information transmission is associated with the behaviors to the utility of relationship updates and interactions among users, therefore, the paper inport the user’s behavior and social relationship updating mechanisms in the model. Furthermore, the influence factors of environmental reputation and reward are introduced to simulate user behavior decision-making in live streaming e-commerce. Simulation results indicate that the positive impact of rewards on information dissemination is obvious when relationship adjustments are infrequent, and the effect is weakened with the increase of reputation. Our work contributes to understanding the dynamics of information dissemination, and provides suggestions for the development of live streaming businesses.

A large number of semi-structured and unstructured texts have been accumulated during the operation of power administration. Scientific mining of the value behind these texts plays a key role in the digital transformation of power grid companies. In this paper, we adopt Bi-LSTM-CRF model for power administrative entity recognition, firstly classify and label the power administrative texts, and then adopt the above mentioned model for power administrative entity recognition, and compare it with HMM, CRF and BiLSTM models experimentally. The final experimental results prove that the Bi-LSTM-CRF model entity recognition has the highest accuracy rate and is more effective in identifying entities in electric power administration.

This paper mainly investigates monthly crude oil price prediction leveraging some traditional time series models and machine learning models. We combined variables suggested from past literature with models suggested from other literature to figure out an approach to improve model prediction accuracy on monthly crude oil price. We first did preliminary variable selections to choose the variables that are most likely related to the crude oil price. Then we established different time series models including a machine learning model on the selected variables. Our approach of modeling and predicting monthly crude oil prices performs well, with slight improvement compared with literature in terms of prediction accuracy. The OLS model plays an important role in variable selection, which improves the model’s performance. LSTM model has the best overall performance, especially did a good job in forecasting oil price during pandemic without overfitting.

With the emergence of AI application ecology, an increasing number of AI applications are being developed and deployed on end devices. For some applications, due to various reasons (such as delay, bandwidth and privacy issues), inference must be performed on the edge nodes. To realize the efficient deployment of multiple networks on different chips, this paper uses cloud service technology and container resource management technology to achieve cloud deployment and uses a variety of model optimization technologies, such as model format conversion, graph optimization, chip optimization, model low-precision calculation optimization, model cutting and distillation. To achieve the effect of saving considerable memory and reducing energy consumption on the premise of satisfying the accuracy.

Users' comments on cars are the focus of automobile companies, which can be used to analyze people's concerns and satisfaction with cars. Such a task is of great commercial value. There are more studies on end-to-end models for topic classification, but the method of manually extracting obvious features and transferring them to the model does not take much time and can improve the accuracy and convergence speed of the model. This study is based on a labeled car comments dataset. First, distinctive feature words are selected as anchor words, and the inclusion of anchor words will anchor this comment to the corresponding topic, making the topic classification and sentiment analysis more accurate. In addition, it is feasible to use a single Bert model to connect the token [CLS] to two linear layers respectively for topic classification and sentiment analysis. It shows that the BERT model with CorEx can improve the classification accuracy and convergence speed significantly, and has a good value in car review classification.

The digital economy is booming, but cybercrimes and telecommunication frauds are emerging one after another. How to detect fraudulent behaviours and prevent the occurrence of crimes is a significant challenge. This paper mainly conducts data mining and analysis on the bank card telecommunication fraud data set, first of all, data mining and feature engineering for the given data set, including analyzing the data integrity, the overall statistical analysis of the data and standardizing the data using the Z-Score standardization method, Use the Pearson correlation coefficient to explore the feature correlation, use the SMOTE method to balance the data set, and finally divide the training set and the test set. Subsequently, four machine learning classification models, including the logistic regression classification model, KNN classification model, decision tree classification model and XGBoost classification model, were established to predict and classify fraudulent behaviours preliminarily. To further mine the data set of bank card telecommunication fraud, the optimal solutions of the models are obtained by grid tuning and cross-validation for the four established models. After experiments, the logistic regression classification model, KNN classification model, decision tree classification model and XGBoost classification The prediction accuracy rates of the model in the test set are 93.45%, 99.85%, 99.92%, and 99.94%, respectively. It is preliminarily believed that the XGBoost and decision tree classification models have excellent classification capabilities. Use the obtained four optimal models to calculate the three performance evaluation indicators of prediction accuracy, recall rate and F1 value in the test set, respectively, and further evaluate the four machine learning models. Through comparative analysis, the XGBoost classification model has the best performance. Due to its classification ability, strong generalization ability and robustness, it is selected as the final bank card telecommunication fraud prediction model. In addition, the P-R curve and ROC curve of the classification results are drawn using the performance evaluation indicators to be intuitive. Analysis of the model's performance further shows that XGBoost has better generalization ability.

The recommendation algorithm based on graph neural network has advantage than the traditional recommendation algorithm that it has the power to extract the association relationship between users and items, which makes it become a research hotspot. However, most of the item recommendation algorithm models based on graph neural network are too focused on the interaction information between users and items, ignore the user preferences and item attributes contained in the review data of users and items. In order to solve this problem, this thesis proposed a new model. This model used the graph neural network to extract high-order feature from the interaction information between items and users, and deep learning method to get the general feature from review data. Then it got the final feature of items and users by carrying out feature fusion, which used the attention mechanism. The thesis conducted comparative experiments and ablation experiments on four sets of data sets and the evaluation indexes include the normalized discounted cumulative gain and recall rate. Experiments results show that this model is more effective than the existing algorithms, the feature fusion has positive effect on improving the recommendation accuracy.

The structural flexibility of the aircraft makes it simple for it to distort during construction as a result of material and working environment variables. Therefore, a deformation forecast model must be proposed in order to quickly identify deformation issues. Deep neural networks are now frequently used in clever production as a result of the advancement of artificial intelligence technology. In this article, aircraft deformation prediction using deep neural networks is combined with a multimodal fusion-based aircraft deformation prediction network is proposed (MMDP-Net). In order to forecast aircraft deformation, we take into account both the aircraft construction data and the operating state data. The characteristics of the aircraft construction mode and the operating state mode are extracted by MMDP-Net, respectively. To combine characteristics, we suggest a multimodal fusion network that uses the average pooling and Bayesian decision-making algorithms. On the aircraft deformation dataset, MMDP-Net gets greater categorization accuracy than the standard point cloud classification network.

Multi-UAV cooperative trajectory planning model with a complex battlefield environment is difficult to establish, and the solution is complex. Aiming at these problems, a more realistic optimal combination model is established, and a particle swarm optimization with based on levy flight and differential evolution (LFPSO-DE) algorithm based on embedded Levy flight strategy is proposed. First, this work establishes a multi-UAV cooperative trajectory planning problem model in a complex environment, and establishes a comprehensive cost function to represent the complex environment; Secondly, aiming at the problem that the PSO algorithm is easy to fall into the local extremum, the Levy flight strategy is introduced to improve the particle swarm position update formula, and the inertia weight and learning factor are optimized, and then combined with the improved DE algorithm, the convergence speed and accuracy of the particle swarm are improved , while enabling the algorithm to have a better ability to explore; Finally, LFPSO-DE algorithm is compared with other algorithms in a variety of test functions, and then applied to multi-UAV trajectory planning experiments. The experimental results show that the LFPSO-DE algorithm has a good convergence speed and can effectively deal with the problem of falling into a local extremum. The planned trajectory is better in terms of voyage, speed, and time-consuming voyage, which has practical application value.

Scalability has grown more important as blockchain technology is adopted. This study introduces RSA-accumulator-based stateless blockchain performance optimization. Verifiable computing technology and distributed design compress node state data and prove smart contract validity, while a new smart contract execution logic and transaction verification scheme solve the problem of transaction verification and execution under stateless conditions, increasing system throughput. The technique reduces consensus node storage and improves throughput compared to Ethereum blockchains, according to experiments. This study examines the scheme's concept, implementation, and blockchain scalability potential.