This paper explores a device-to-device (D2D) assisted setting in the multi-access networks, where the users communicate with each other and each one can access multiple neighbouring cache nodes in a cyclic wrap-around fashion. For such networks, we introduce two new combinatorial arrays namely the caching array and the delivery array to characterize the coded caching process. Based on this, a novel construction of caching array and delivery array is derived through a smart transform over the existing placement delivery array (PDA). The resulting coded caching scheme can achieve a small transmission rate under the case of uncoded placement. It is shown that the proposed scheme has advantages in both the transmission rate and the subpacketization level over the benchmark scheme.

With the rapid development of technologies such as cloud computing, big data, and the Internet of Things (IoT), Software-Defined Networking (SDN) is emerging as a new network architecture for the modern Internet. SDN separates the control plane from the data plane, allowing a central controller, the SDN controller, to quickly direct the routing devices within the topology to forward data packets, thus providing flexible traffic management for communication between information sources. However, traditional Distributed Denial of Service (DDoS) attacks still significantly impact SDN systems. This paper proposes a novel dual-layer strategy capable of detecting and mitigating DDoS attacks in an SDN network environment. The first layer of the strategy enhances security by using blockchain technology to replace the SDN flow table storage container in the northbound interface of the SDN controller. Smart contracts are then used to process the stored flow table information. We employ the time window algorithm and the token bucket algorithm to construct the first layer strategy to defend against obvious DDoS attacks. To detect and mitigate less obvious DDoS attacks, we design a second-layer strategy that uses a composite data feature correlation coefficient calculation method and the Isolation Forest algorithm from unsupervised learning techniques to perform binary classification, thereby identifying abnormal traffic. We conduct experimental validation using the publicly available DDoS dataset CIC-DDoS2019. The results show that using this strategy in the SDN network reduces the average deviation of round-trip time (RTT) by approximately 38.86% compared with the original SDN network without this strategy. Furthermore, the accuracy of DDoS attack detection reaches 97.66% and an F1 score of 92.2%. Compared with other similar methods, under comparable detection accuracy, the deployment of our strategy in small-scale SDN network topologies provides faster detection speeds for DDoS attacks and exhibits less fluctuation in detection time. This indicates that implementing this strategy can effectively identify DDoS attacks without affecting the stability of data transmission in the SDN network environment.

Traditional image classification often misclassifies unknown samples as known classes during testing, degrading recognition accuracy. Open-set image recognition can simultaneously detect known classes (KCs) and unknown classes (UCs) but still struggles to improve recognition performance caused by open space risk. Therefore, we introduce a cosine distance loss function (CDLoss), which exploits the orthogonality of one-hot encoding vectors to align known samples with their corresponding one-hot encoder directions. This reduces the overlap between the feature spaces of KCs and UCs, mitigating open space risk. CDLoss was incorporated into both Softmax-based and prototype-learning-based frameworks to evaluate its effectiveness. Experimental results show that CDLoss improves AUROC, OSCR, and accuracy across both frameworks and different datasets. Furthermore, various weight combinations of the ARPL and CDLoss were explored, revealing optimal performance with a 1:2 ratio. T-SNE analysis confirms that CDLoss reduces the overlap between the feature spaces of KCs and UCs. These results demonstrate that CDLoss helps mitigate open space risk, enhancing recognition performance in open-set image classification tasks.

Computer network traffic refers to the total amount of data passing through the network in a certain period of time, and is an important parameter to measure the load and running status of the network. In the background of cloud service and network slicing, predicting network traffic demand is beneficial for network operators to adjust network resources and meet user demand. However, the actual network traffic demand is characterized by real-time and sudden, the traditional network traffic prediction model has the problems of less prediction time points and low prediction accuracy, and the prediction model based on neural network has the problems such as gradient and error accumulation. Therefore, improving the network prediction model and avoiding the above problems are the focus of the current research. In this study, based on the time series theory of network traffic data, we propose a network traffic prediction algorithm using a Self-Adaptive Attention Mechanism (AAM). This algorithm dynamically adapts to process input data in an adaptive manner, reducing computational complexity while minimizing information loss. The Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) metrics are employed to evaluate the prediction model's performance. Experimental results indicate that compared to classical models like the Auto-regressive Integrated Moving Average Model (ARIMA) and Long Short-Term Memory networks (LSTM), AAM exhibits higher accuracy and better predictive performance on datasets with significant fluctuations. For instance, when predicting 6 time points, AAM reduces RMSE by approximately 16.50% and 61.40% compared to ARIMA on two datasets, respectively. When predicting 36 time points, the reduction is about 13.12% and 70.51%, and for the maximum of 144 time points, reductions are approximately 51.77% and 73.40%.

In edge intelligence networks, various applications exist based on different users and datasets. To realize intelligence, many works design various edge learning models to process these data. However, different types of users and models have negative influence on the learning performance, even degrading the learning gain. To address these issues, we propose a learning-aware optimization model with probabilistic scheduling. First, we design users scheduling to reveal the relationships between resource allocation and learning performance. Then, we also develop a fast proximal linearized algorithm in the large-scale communication networks. Next, we derive tasks scheduling in asymptotic cases to improve resource utilization efficiency. At last, extensive experimental results corroborate that probabilistic scheduling and multi-objective learning models improve the distinct edge learning gain efficiently compared with the state-of-the-art benchmark algorithms.