This paper proposes the RGLK model, which integrates the Linformer, Gated Recurrent Unit (GRU), Reversible Instance Normalization (RevIN), a dynamic gating mechanism, and the Kolmogorov–Arnold Network (KAN) to address common challenges in intrusion detection, such as class imbalance and limited time series modeling capabilities. To alleviate the class imbalance, SMOTE is adopted as a data-level augmentation strategy. The RevIN module is introduced to address distribution shifts in the time series while preserving the original features, thereby enhancing model stability. In the feature extraction stage, the GRU is used to capture local temporal dependencies, the Linformer efficiently models long-range dependencies, and the GRU–Linformer dynamic gating mechanism adaptively integrates the two, improving the flexibility and robustness of feature representation. Finally, the KAN replaces the traditional classifier to better fit the complex decision boundaries in the high-dimensional feature space. Experiments demonstrate that the RGLK model achieves an accuracy of 99.68% on the CICIDS2017 dataset and 99.17% on the NSL-KDD dataset for multi-class intrusion detection tasks, significantly outperforming existing deep learning methods.

Optimizing resource allocation and routing to satisfy service needs is paramount in large-scale networks. Software-defined networking (SDN) is a new network paradigm that decouples forwarding and control, enabling dynamic management and configuration through programming, which provides the possibility for deploying intelligent control algorithms (such as deep reinforcement learning algorithms) to solve network routing optimization problems in the network. Although these intelligent-based network routing optimization schemes can capture network state characteristics, they are prone to falling into local optima, resulting in poor convergence performance. In order to address this issue, this paper proposes an African Vulture Routing Optimization (AVRO) algorithm for achieving SDN routing optimization. AVRO is based on the African Vulture Optimization Algorithm (AVOA), a population-based metaheuristic intelligent optimization algorithm with global optimization ability and fast convergence speed advantages. First, we improve the population initialization method of the AVOA algorithm according to the characteristics of the network routing problem to enhance the algorithm’s perception capability towards network topology. Subsequently, we add an optimization phase to strengthen the development of the AVOA algorithm and achieve stable convergence effects. Finally, we model the network environment, define the network optimization objective, and perform comparative experiments with the baseline algorithms. The experimental results demonstrate that the routing algorithm has better network awareness, with a performance improvement of 16.9% compared to deep reinforcement learning algorithms and 71.8% compared to traditional routing schemes.

2D Human Pose Estimation plays a crucial role in analyzing performance in fitness activities. Current single-stage methods suffer from the lack of interaction between classification and regression branches, large network parameter sizes, and poor detection accuracy due to strong short-distance dependencies between keypoints. To address these issues in fitness, a lightweight dynamic task alignment framework based on Yolov8-pose is proposed. In order to enhance classification and regression alignment in single-stage networks, a dynamic task alignment detection head is proposed by leveraging label assignment strategies and learning task interaction features. To mitigate the issue of information loss caused by the unidirectional propagation in Yolov8-pose, the backbone is replaced with RevCol to enhance feature retention. Additionally, efficient self-adjusting weighted downsampling module is designed to retain more useful information. Furthermore, the C2f module in downsampling is enhanced with Context-Guided Blocks, integrating local and global feature fusion. Experimental results on a self-created fitness action dataset show that, compared to Yolov8n-pose, our proposed algorithm reduces parameters by 60.8%, decreases computational cost by 33.7%, and improves average detection accuracy by 1.89%.

In mobile edge computing (MEC), randomly offloading tasks to edge servers (ES) can cause wireless devices (WD) to compete for limited bandwidth resources, leading to overall performance degradation. Reinforcement learning can provide suitable strategies for task offloading and resource allocation through exploration and trial-and-error, helping to avoid blind offloading. However, traditional reinforcement learning algorithms suffer from slow convergence and a tendency to get stuck in suboptimal local minima, significantly impacting the energy consumption and data timeliness of edge computing task unloading. To address these issues, we propose Parallel Exploration with Asynchronous Training-based Deep Reinforcement Learning (PEATDRL) algorithm for MEC network offloading decisions. Its objective is to maximize system performance while limiting energy consumption in an MEC environment characterized by time-varying wireless channels and random user task arrivals. Firstly, our model employs two independent DNNs for parallel exploration, each generating different offloading strategies. This parallel exploration enhances environmental adaptability, avoids the limitations of a single DNN, and addresses the issue of agents getting stuck in suboptimal local minima due to the explosion of decision combinations, thereby improving decision performance. Secondly, we set different learning rates for the two DNNs during the training phase and trained them at various intervals. This asynchronous training strategy increases the randomness of decision exploration, prevents the two DNNs from converging to the same suboptimal local solution, and improves convergence efficiency by enhancing sample utilization. Finally, we examine the impact of different parallel levels and training step differences on system performance metrics and explain the parameter choices. Experimental results show that the proposed method provides a viable solution to the performance issues caused by slow convergence and local minima, with PEATDRL improving task queue convergence speed by more than 20

Currently, existing research on deploying deep reinforcement learning on software-defined networks (SDN) to achieve route optimization does not consider the network’s spatial–temporal correlation globally and has yet to reach the ultimate in performance. Given the above issues, this study proposes a Proximal Policy Optimization algorithm based on the Attention mechanism and Spatio–Temporal correlation (ASTPPO) to optimize the SDN routing issue. First, we extract temporal and spatial correlation features in state information using Gated Recurrent Units (GRU) and Graph Attention Networks (GAT), providing implicit information containing more environments for reinforcement learning decisions. Second, we use the skip-connect method to connect implicit and directly related information into a multi-layer perceptron, improving the model's learning efficiency and perceptual ability. Finally, we demonstrate the effectiveness of ASTPPO through static and dynamic traffic experiments. Benefitting from Spatio–Temporal correlation learning with a global view, ASTPPO performs better load balancing and congestion control under different traffic intensity requirements and network topologies than other reinforcement learning baseline algorithms. The simulation results show that the ASTPPO algorithm improved by 9.02

Load-balancing optimization in software-defined networking (SDN) has been researched for a long time. Researchers have proposed many solutions to the load-balancing problem but have rarely considered the impact of transmission delay between controllers and switches under high-load network conditions. In this paper, we propose an adaptive load-balancing architecture based on link-state prediction (ALBLP) in SDN that can solve the influence of transmission delay between controllers and switches on network load balancing. ALBLP constructs the prediction model of the network link status, adopts the long-term and short-term memory neural network (LSTM) algorithm to predict the network link-state value, and then uses the predicted value as the Dijkstra weight to calculate the optimal path between network hosts. The proposed architecture can adaptively optimize network load balancing and avoid the empty window period, in which the switch flow table does not exist by actively issuing the flow table. Under the network architecture, we collect the data set of the network link-state by simulating the GÉANT network, and we verify the effectiveness of the proposed algorithm. The experiment results show that the ALBLP proposed in this paper can optimize load balancing in SDN and solve the problem of transmission delay between controllers and switches. It has a maximum load-balancing improvement of 23.7% and 11.7% in comparison with the traditional Open Shortest Path First (OSPF) algorithm and the reinforcement learning method based on Q-Learning.

In text classification methods, the key to correct text classification is to extract semantic information accurately. However, important semantic information cannot be recognized, and existing models focus too much on local information and ignore global information often make the semantic bias, which in turn affects the classification results. To address these problems, this paper designs a text classification method based on BERT-Att-TextCNN model, which significantly improves the classification accuracy. First, this paper uses the feature of BERT using bidirectional architecture to obtain the accurate semantic representation of each word through multiple rounds of training iterative feedback, and then obtains the high-quality sentence vector. Secondly, this paper takes advantage of the attention mechanism in capturing important information to compute weights on the word vectors to enhance the semantic representation, and feeds the results into the TextCNN network to extract local key features. Then, the model in this paper fuses the hidden output of BERT with the pooled output of TextCNN to converge global semantic features and local semantic features to strengthen the judgment ability and robustness of the model. Finally, the classification results are obtained through the fully connected layer. Experiments show that this paper's method is tested on THUCNews, SST-2, and content security datasets, and the accuracy reaches 94.8%, 94.6%, and 99.4%, respectively, which is better than the classical classification model. The model in this paper has good generalization ability while improving the accuracy.

由于已有算法未能进行数据降噪,导致查全率和查准率偏低,运行时间较长.提出一种分布式供热管网运行数据异常在线辨识算法,通过基于贝叶斯的小波降噪方法对供热管网运行数据进行降噪处理.利用数据融合异常辨识特征进行分布式供热管网数据特征识别,构建特征提取模型.使用空间欠采样技术对数据特征进行重组,提取数据在线辨识的统计特征量,进而通过数据异常分布状态完成数据异常在线辨识,同时引入布谷鸟搜索算法进行在线辨识的自适应寻优,获取最终的辨识结果.仿真结果表明,所提算法能够有效提升查全率和查准率,同时有效降低运行时间.

Due to the rapid development of network communication technology and the significant increase in network terminal equipment, the application of new network architecture software-defined networking (SDN) combined with reinforcement learning in network traffic scheduling has become an important focus of research. Because of network traffic transmission variability and complexity, the traditional reinforcement-learning algorithms in SDN face problems such as slow convergence rates and unbalanced loads. The problems seriously affect network performance, resulting in network link congestion and the low efficiency of inter-stream bandwidth allocation. This paper proposes an automatic load-balancing architecture based on reinforcement learning (ALBRL) in SDN. In this architecture, we design a load-balancing optimization model in high-load traffic scenarios and adapt the improved Deep Deterministic Policy Gradient (DDPG) algorithm to find a near-optimal path between network hosts. The proposed ALBRL uses the sampling method of updating the experience pool with the SumTree structure to improve the random extraction strategy of the empirical-playback mechanism in DDPG. It extracts a more meaningful experience for network updating with greater probability, which can effectively improve the convergence rate. The experiment results show that the proposed ALBRL has a faster training speed than existing reinforcement-learning algorithms and significantly improves network throughput.

Graph neural network extends deep neural network to graph structured data, and realizes many tasks related to graphs. However, most of the existing models focus on design graph convolution operations, and the important role of graph pools in hierarchical learning representation is often ignored. In addition, the existing Top-K graph pooling method will cause the feature information of unselected nodes lost, which inevitably leads to the loss of a large amount of graph topology feature information. In order to solve the above problems, improved “intra-layer” & “inter-layer” joint feature extraction structure is proposed, which introduces a structural feature learning mechanism, combines the graph adaptive pooling operator with the existing GNN architecture. To mitigate the problem of insufficient feature extraction by LSTM, BiLSTM is used to extract the inter-layer feature. Experiment results show that our method can achieves better performance on benchmark data sets by using a reasonable number of parameters.

Previous research has shown that smaller cache capacity of a single small base station (SBS) in the 5G ultra-dense network (UDN) results in a lower hit rate, a scheme of partitioned cooperation edge caching is proposed. First, according to the location information and the download history, the cosine similarity and Gaussian similarity between the SBSs are calculated, then the comprehensive similarity matrix is obtained, which is the weight matrix of the Fast Newman Algorithm (FN). Finally, a content placement scheme basing on the local popularity in the cooperative cluster is proposed. The experimental results show that by dividing the cooperation cluster, the network performance can be further enhanced with a higher overall hit rate and the average content transmission delay can be reduced.

针对现有的最大容错块挖掘算法受容错度影响较大,在大型事务数据库和稀疏数据库性能不佳的情况,提出了基于并行PSO的最大容错块挖掘算法.通过删除小于最小支持度的项提高数据库的密度,利用高斯扰动优化的PSO算法结合Spark框架对目标数据库进行最大容错块挖掘.实验结果表明,多组不同规模数据集实验结果均优于利用整数线性规划设计的最大容错块挖掘算法,对于同一数据库算法效率不会受容错度的影响.

为解决传统图神经网络在图分类任务上存在训练过程中噪声信息过多以及不能完整地挖掘图的层次表征信息等问题,提出一种端到端的基于重要性池化的层级图表示学习方法.该方法以层内-层间联合特征提取结构为基础,主要包括层内特征提取模块和层间特征提取模块2个部分.利用池化方法将图粗粒化为高级子图结构,以缩减特征图的尺寸;利用循环单元,以抑制层间噪声的传播,并自适应地聚集层级表征.实验结果表明,在合理的时间复杂度下,该方法能使损失函数收敛于更小的值,且模型精度有明显的提升.

为了提高个性化PageRank算法在大型网络图中的计算效率,提出基于可达查询的PPR算法.该算法采用一定的分割方法将大图上的计算转移到子图上,通过可达查询算法快速删除所有与计算无关的节点与边,得到源节点的可达子图,在可达子图上计算节点的近似PPR估计值.实验结果证明,该算法可显著提高PPR的计算效率.

为了合理评估Android系统软件的风险,提出一种基于AHP贝叶斯网络的Android软件风险评估方法.利用层次分析法构建Android软件风险评估体系,将其映射为贝叶斯网络,并对节点实例化.利用三角模糊函数计算节点各风险状态的隶属度,将其添加到节点条件概率表.针对app种类和权限的不确定性,在网络中为其设置证据节点,在实际应用场景中通过改变证据节点的状态对app风险实时推理.Netica软件仿真结果表明,该方法能评估不同软件分类下的权限风险和app风险,具有实际可操作性和易扩展性.

针对广告点击率预估模型难以有效捕捉高阶特征组合的非线性关系且模型复杂度较高、实际应用困难的问题,从模型预估准确度和模型实用性出发,提出了基于深度网络模型压缩的广告点击率预估模型.该模型无需特征工程,利用多头注意力机制对全局特征的有效采样与预组合,结合深度神经网络(DNN),更好地捕捉全局高阶特征组合的非线性关系;结合残差网络对复杂网络的优势,提高深度神经网络的收敛效率,避免模型过拟合;经模型蒸馏得到轻量化模型,可以更好地应用于实际场景.在Avazu数据集上实验表明,该模型的预估准确率和运行速度均有优异表现.

The current maximum fault-tolerant tile mining has the following problems: 1) the mining speed is slow 2) the mining speed is greatly affected by the tolerance. To solve these problems, a maximum fault-tolerant tile-mining algorithm based on parallel PSO is proposed in this paper. PSO algorithm is used to find the maximum fault-tolerant tile quickly and accurately, and the Spark framework is combined to further improve the calculation speed. Compared with the maximum fault-tolerant tile mining algorithm of integer linear programming, experimental results are superior to traditional algorithms in speed and stability. Then the proposed algorithm was applied to wind power generation system, and the experiment outcome shows that the algorithm is accurate and eddective for the dateset of real system.

Time sensitive network requires ultra-low delay time services. It is necessary to know the state of the transmission paths, especially to be able to quickly detect and locate fault point when a fault occurs. The fault detection and location method based on segment routing is studied in this paper. A candidate probe paths set generation algorithm based on segment routing is proposed, and an optimal probe matrix construction algorithm is designed. According to the information about the probe paths in the probe matrix, the fault node can be detected and located quickly. Due to the minimization of the number of the probe paths, the detection process does not occupy too much bandwidth resources, and the time is short, which can meet the needs of time sensitive network.