Graph conjoint attention (CAT) network is one of the best graph convolutional networks (GCNs) frameworks,
which uses a weighting mechanism to identify important neighbor nodes. However, this weighting mechanism is
learned based on static information, which means it is susceptible to noisy nodes and edges, resulting in significant
limitations. In this paper, a method is proposed to obtain context dynamically based on random walk, which allows
the context-based weighting mechanism to better avoid noise interference. Furthermore, the proposed context-based
weighting mechanism is combined with the node content-based weighting mechanism of the graph attention (GAT)
network to form a model based on a mixed weighting mechanism. The model is named as the context-based and
content-based graph convolutional network (CCGCN). CCGCN can better discover important neighbors, eliminate
noise edges, and learn node embedding by message passing. Experiments show that CCGCN achieves state-of-the-
art performance on node classification tasks in multiple datasets.

Membership inference attacks on machine learning models have drawn significant attention. While current research primarily utilizes shadow modeling techniques, which require knowledge of the target model and training data, practical scenarios involve black-box access to the target model with no available information. Limited training data further complicate the implementation of these attacks. In this paper, we experimentally compare common data enhancement schemes and propose a data synthesis framework based on the variational autoencoder generative adversarial network (VAE-GAN) to extend the training data for shadow models. Meanwhile, this paper proposes a shadow model training algorithm based on adversarial training to improve the shadow model's ability to mimic the predicted behavior of the target model when the target model's information is unknown. By conducting attack experiments on different models under the black-box access setting, this paper verifies the effectiveness of the VAE-GAN-based data synthesis framework for improving the accuracy of membership inference attack. Furthermore, we verify that the shadow model, trained by using the adversarial training approach, effectively improves the degree of mimicking the predicted behavior of the target model. Compared with existing research methods, the method proposed in this paper achieves a 2% improvement in attack accuracy and delivers better attack performance.

To solve the privacy leakage problem of truck trajectories in intelligent logistics, this paper proposes a Quadtree-based Personalized Joint location Perturbation (QPJLP) algorithm using location generalization and local differential privacy techniques. Firstly, a flexible position encoding mechanism based on the spatial quadtree indexing is designed, and the length of the encoding can be adjusted freely according to data availability. Secondly, to meet the privacy needs of different locations of users, location categories are introduced to classify locations as sensitive and ordinary locations. Finally, the truck invokes the corresponding mechanism in the QPJLP algorithm to locally perturb the code according to the location category, allowing the protection of non-sensitive locations to be reduced without weakening the protection of sensitive locations, thereby improving data availability. Simulation experiments demonstrate that the proposed algorithm effectively meets the personalized trajectory privacy requirements while also exhibiting good performance in trajectory proportion estimation and Top-K classification.

An artificial rabbit optimization algorithm based on chaotic mapping and Levy flight improvement is proposed, which has the advantages of good initial population quality and fast convergence compared with the traditional artificial rabbit optimization algorithm, called CLARO. CLARO’s improvement method starts from three aspects: to optimize the quality of the initial population of the algorithm a chaotic mapping is brought in to initialize the population; to avoid the algorithm from falling into local optimum Levy flight is added in the exploration phase and the threshold of energy factor A is optimized to better balance exploration and exploitation. The efficiency of CLARO is tested on a set of 23 benchmark function sets by comparing it with ARO and different meta-heuristics algorithms. At last, the comparison experiments conclude that all three improvement strategies enhance the performance of ARO to some extent, with Levy flight providing the most significant improvement in ARO performance. The experimental results showed that CLARO has better results and faster convergence compared to other algorithms, while successfully addressing the drawbacks of ARO and being able to face more challenging problems.

目前，大多数构建的知识图谱无论其规模如何，大多有不完备性问题。这种不完备性会对基于知识图谱的应用产生负面影响。作为知识图谱补充的重要方法，链接预测近年来已成为热门研究课题。本文提出了一种基于半监督学习和模型汤思想的链接预测模型性能增强方案，通过对模型架构进行微小改变，有效提高了几种主流链接预测模型的性能。这一创新方案主要包括两个部分：（1）使用半监督学习策略预测图中的潜在事实三元组，（2）创造性地结合半监督学习和模型汤，进一步提高最终模型的性能，而不增加显著的计算开销。我们通过实验证实了该方案在各种链接预测模型上的有效性，特别是在具有密集关系的数据集上。对于测试的模型中整体性能最佳的模型CompGCN，在经过增强方案后，在FB15K-237数据集上的Hits@1指标提高了14.7%，在WN18RR数据集上提高了7.8%。同时，我们观察到增强方案中的半监督学习策略对于多类链接预测模型有显著改进，并且模型汤的引入带来的性能改进与具体的测试模型有关，某些模型的性能得到了改善，而其他模型的性能基本保持不变。

In this paper, the online power control and rate adaptation for a wireless communication system with energy harvesting are investigated, in which soft decision decoding is adopted by the receiver. To efficiently utilize the harvested energy and maximize the average actual information transmission rate, transmit power, modulation order and code rate are jointly optimized. The Lyapunov framework is utilized to transform the long-term optimization problem into an optimization problem per time slot. Since there is no theoretical formula for the error rate of soft decision decoding, the optimization problem cannot be analytically solved. A table to find the optimal modulation order and code rate under the different values of signal-to-noise ratio is built first, and then a numeric algorithm to find the solution to the optimization problem is given. The feasibility and performance of the proposed algorithm are demonstrated by simulation. The simulation results show that compared with the algorithms to maximize the theoretic channel capacity, the proposed algorithm can achieve a higher actual transmission rate.

A new procedure of distributed self-control coverage for monitoring the dynamic targets with mobile sensor network is proposed. A special model is given for maintaining the nodes bi-connectivity and optimizing the coverage of the moving targets. The model consists of two parts, the virtual force model is proposed for motion control and the whale optimization algorithm is improved to further optimize the node positions and to reach the steady state quickly. The virtual resultant force stretches the network toward the uncovered targets by its multi-target attractive force, and maintains the network connectivity by its attractive force while network stretching, and avoids node collisions by its repulsive force while nodes moving. The operating mechanism of multi-target attractive force and other forces is also profoundly anatomized. The adjustment criteria for the model in different application scenarios are also given. Finally, the comparisons are shown to be significant advantages over other similar kinds.

In the classification problem, deep kernel extreme learning machine (DKELM) has the characteristics of efficient
processing and superior performance, but its parameters optimization is difficult. To improve the classification
accuracy of DKELM, a DKELM algorithm optimized by the improved sparrow search algorithm (ISSA), named as
ISSA-DKELM, is proposed in this paper. Aiming at the parameter selection problem of DKELM, the DKELM
classifier is constructed by using the optimal parameters obtained by ISSA optimization. In order to make up for the
shortcomings of the basic sparrow search algorithm (SSA), the chaotic transformation is first applied to initialize the
sparrow position. Then, the position of the discoverer sparrow population is dynamically adjusted. A learning
operator in the teaching-learning-based algorithm is fused to improve the position update operation of the joiners.
Finally, the Gaussian mutation strategy is added in the later iteration of the algorithm to make the sparrow jump out
of local optimum. The experimental results show that the proposed DKELM classifier is feasible and effective, and
compared with other classification algorithms,the proposed DKELM algorithm aciheves better test accuracy.

The continuous phase modulation (CPM) technique is widely used in range telemetry due to its high spectral
efficiency and power efficiency. However, the demodulation performance of the traditional maximum likelihood
sequence detection (MLSD) algorithm significantly deteriorates in non-ideal synchronization or fading channels. To
address this issue, this work proposes a convolutional neural network (CNN) called the cascade parallel crossing
network (CPCNet) to enhance the robustness of CPM signals demodulation. The CPCNet model employs a multiple
parallel structure and feature fusion to extract richer features from CPM signals. This approach constructs feature
maps at different levels, resulting in a more comprehensive training of the model and improved demodulation
performance. Simulation results show that under Gaussian channel, the proposed CPCNet achieves the same bit
error rate (BER) performance as MLSD method when there is no timing error, but with 1/4 symbol period timing
error, the proposed method has 2 dB demodulation gain compared with CNN and convolutional long short-term
memory deep neural network (CLDNN). In addition, under Rayleigh channel, the BER of the proposed method is
reduced by 5% -87% compared to that of MLSD in the wide signal-to-noise ratio (SNR) region.

In this paper, a W-band broadband vialess microstrip (MS)-to-MS vertical transition in multilayer liquid crystal
polymer (LCP) substrate is presented, which consists of two MS lines in the top layer, a common ground plane and
slotline resonators in the second layer, and a close-loop transmission-line in the third layer. To increase the
passband of the vialess vertical transition, an H-shaped slotline resonator is introduced, which greatly improves the
impedance performance of the slotline resonator, and the full-wave simulated results indicate that insertion loss
(IL) is less than 2 dB and return loss (RL) is better than 10 dB at W-band. To verify this design, the broadband
vertical transition is fabricated and measured. The measured results indicate that a broadband vertical transition
with RL better than 10 dB and IL less than 5.67 dB can be obtained in the frequency range from 70.00 GHz to
104.09 GHz. Due to the fabrication error in the preparation process, the measured results are deteriorated
compared to the simulated results, and the investigation indicates that the deviation is caused by the thickness error
of the LCP substrate.

In order to improve the accuracy of used car price prediction, a machine learning prediction model based on the
retention rate is proposed in this paper. Firstly, a random forest algorithm is used to filter the variables in the data.
Seven main characteristic variables that affect used car prices, such as new car price, service time, mileage and so
on, are filtered out. Then, the linear regression classification method is introduced to classify the test data into high
and low retention rate data. After that, the extreme gradient boosting ( XGBoost) regression model is built for the
two datasets respectively. The prediction results show that the comprehensive evaluation index of the proposed
model is 0. 548, which is significantly improved compared to 0. 488 of the original XGBoost model. Finally,
compared with other representative machine learning algorithms, this model shows certain advantages in terms of
mean absolute percentage error (MAPE), 5% accuracy rate and comprehensive evaluation index. As a result, the
retention rate-based machine learning model established in this paper has significant advantages in terms of the
accuracy of used car price prediction.

Lumped element lowpass filter (LPF) for ultra-high frequency (UHF) radio frequency (RF) front-end system is
presented based on multilayer liquid crystal polymer (LCP). The lumped element LPF can achieve miniaturization
and one transmission zero on the stopband by the 8-shaped inductor. The lumped element LPF is fabricated on a 4-
layer LCP substrate with a compact size of 9 mm × 14 mm × 0.193 mm. The measured cut off frequency of the
lumped element LPF is 0.5 GHz with insertion loss (IL) less than 0.37 dB. Both measured and simulated results
suggest that it is a possible candidate for the application of UHF RF front-end system.

Learning archives management in traditional systems faces challenges such as inadequate security, weak tamper resistance, and limited sharing capabilities. To address these issues, this paper proposes LRChain, a method based on consortium blockchain, for lifelong learning archives data protection and sharing. LRChain employs a combination of on-chain and off-chain cooperative storage using a consortium chain and InterPlanetary File System (IPFS) to enhance data security and availability. It also enables fine-grained verification of learning archives through selective disclosure principles, ensuring privacy protection of sensitive data. Furthermore, an attribute-based encryption algorithm is utilized to establish authorized access control for learning archives, facilitating safe and trusted sharing. Experimental evaluations and security analyses demonstrate that this method exhibits decentralization, strong security, tamper resistance, and performs well, effectively meeting the requirements for secure sharing of learning archive data.

In frequency division duplex ( FDD) massive multiple-input multiple-output ( MIMO) systems, a bidirectional positional attention network ( BPANet) was proposed to address the high computational complexity and low accuracy of existing deep learning-based channel state information ( CSI) feedback methods. Specifically, a bidirectional position attention module ( BPAM) was designed in the BPANet to improve the network performance. The BPAM captures the distribution characteristics of the CSI matrix by integrating channel and spatial dimension information, thereby enhancing the feature representation of the CSI matrix. Furthermore, channel attention is decomposed into two one-dimensional (1D) feature encoding processes effectively reducing computational costs. Simulation results demonstrate that, compared with the existing representative method complex input lightweight neural network ( CLNet), BPANet reduces computational complexity by an average of 19. 4% and improves accuracy by an average of 7. 1% . Additionally, it performs better in terms of running time delay and cosine similarity.

For the task of content retrieval, analysis and generation of film and television scene images in the field of
intelligent editing, fine-grained emotion recognition and prediction of images is of great significance. In this paper,
the fusion of traditional perceptual features, art features and multi-channel deep learning features are used to reflect
the emotion expression of different levels of the image. In addition, the integrated learning model with stacking
architecture based on linear regression coefficient and sentiment correlations, which is called the LS-stacking
model, is proposed according to the factor association between multi-dimensional emotions. The experimental
results prove that the mixed feature and LS-stacking model can predict well on the 16 emotion categories of the self-
built image dataset. This study improves the fine-grained recognition ability of image emotion by computers, which
helps to increase the intelligence and automation degree of visual retrieval and post-production system.

In convolutional neural networks ( CNNs), the floating-point computation in the traditional convolutional layer is enormous, and the execution speed of the network is limited by intensive computing, which makes it challenging to meet the real-time response requirements of complex applications. This work is based on the principle that the time domain convolution result equals the frequency domain point multiplication result to reduce the amount of floating- point calculations for convolution. The input feature map and the convolution kernel are converted to the frequency domain by the fast Fourier transform( FFT), and the corresponding point multiplication is performed. Then the frequency domain result is converted back to the time domain, and the output result of the convolution is obtained. In the shared CNN, the input feature map is much larger than the convolution kernel, resulting in many invalid operations. The overlap addition method is proposed to reduce invalid calculations and speed up network execution better. This work designs a hardware accelerator for frequency domain convolution and verifies its efficiency on the Xilinx Zynq UltraScale + MPSoC ZCU102 board. Comparing the calculation time of visual geometry group 16 ( VGG16 ) under the ImageNet dataset faster than the traditional time domain convolution, the hardware acceleration of frequency domain convolution is 8. 5 times.

Due to the diversity of graph computing applications, the power-law distribution of graph data, and the high compute-to-memory ratio, traditional architectures face significant challenges regarding poor flexibility, imbalanced workload distribution, and inefficient memory access when executing graph computing tasks. Graph computing accelerator, GraphApp, based on a reconfigurable processing element ( PE) array was proposed to address the challenges above. GraphApp utilizes 16 reconfigurable PEs for parallel computation and employs tiled data. By reasonably dividing the data into tiles, load balancing is achieved and the overall efficiency of parallel computation is enhanced. Additionally, it preprocesses graph data using the compressed sparse columns independently ( CSCI) data compression format to alleviate the issue of low memory access efficiency caused by the high memory access-to-computation ratio. Lastly, GraphApp is evaluated using triangle counting ( TC) and depth-first search ( DFS) algorithms. Performance analysis is conducted by measuring the execution time of these algorithms in GraphApp against existing typical graph frameworks, Ligra, and GraphBIG, using six datasets from the Stanford Network Analysis Project ( SNAP) database. The results show that GraphApp achieves a maximum performance improvement of 30.86 % compared to Ligra and 20.43 % compared to GraphBIG when processing the same datasets.

Mobile manipulators are used in a variety of fields because of their flexibility and maneuverability. The path
planning capability of the mobile manipulator is one of the important indicators to evaluate the performance of the
manipulator, but it is greatly challenged in the face of maps with narrow channel. To address the problem, an
improved hierarchical motion planner (IHMP) is proposed, which consists of a two-dimensional (2D) path planner
for the mobile base, and a three-dimensional (3D) trajectory planner for the on-board manipulator. Firstly, a
hybrid sampling strategy is proposed, which can reduce invalid nodes of the generated probabilistic roadmap.
Bridge test is used to locate the narrow channel areas, and a Gaussian sampler is deployed in these areas and the
boundaries. Meanwhile, a random sampler is deployed in the rest areas. Trajectory planner for on-board
manipulator is to generate a collision-free and safe trajectory in the narrow channel with collaboration of the 2D path
planner. The experimental results show that IHMP is effective for mobile manipulator motion planning in complex
static environments, especially in narrow channel.

Addressing the issue of low pulse identification rates for low probability of intercept ( LPI) radar signals under low signal-to-noise ratio ( SNR) conditions, this paper aims to investigate a new method in the field of deep learning to recognize modulation types of LPI radar signals efficiently. A novel algorithm combining dual efficient network ( DEN) and non-local means ( NLM) denoising was proposed for the identification and selection of LPI radar signals. Time-domain signals for 12 radar modulation types were simulated, adding Gaussian white noise at various SNRs to replicate complex electronic countermeasure scenarios. On this basis, the noisy radar signals undergo Choi-Williams distribution ( CWD ) time-frequency transformation, converting the signals into two- dimensional (2D) time-frequency images ( TFIs). The TFIs are then denoised using the NLM algorithm. Finally, the denoised data is fed into the designed DEN for training and testing, with the selection results output through a softmax classifier. Simulation results demonstrate that at an SNR of - 8 dB, the algorithm can achieve a recognition accuracy of 97.22% for LPI radar signals, exhibiting excellent performance under low SNR conditions. Comparative demonstrations prove that the DEN has good robustness and generalization performance under conditions of small sample sizes. This research provides a novel and effective solution for further improving the accuracy of identification and selection of LPI radar signals.

Great challenges and demands are presented by increasing edge computing services for current power Internet of things ( Power IoT) to deal with the serious diversity and complexity of these services. To improve the matching degree between edge computing and complex services, the service identification function is necessary for Power IoT. In this paper, a naive long short-term memory ( Naive-LSTM ) based service identification scheme of edge computing devices in the Power IoT was proposed, where the Naive-LSTM model makes full use of the most simplified structure and conducts discretization of the long short-term memory ( LSTM) model. Moreover, the Naive-LSTM based service identification scheme can generate the probability output result to determine the task schedule policy of Power IoT. After well learning operation, these Naive-LSTM classification engine modules in edge computing devices of Power IoT can perform service identification, by obtaining key characteristics from various service traffics. Testing results show that the Naive-LSTM based services identification scheme is feasible and efficient in improving the edge computing ability of the Power IoT.