Network traffic prediction is an important tool for the management and control of IoT, and timely and accurate traffic prediction models play a crucial role in improving the IoT service quality. The degree of burstiness in intelligent network traffic is high, which creates problems for prediction. To address the problem faced by traditional statistical models, which cannot effectively extract traffic features when dealing with inadequate sample data, in addition to the poor interpretability of deep models, this paper proposes a prediction model (fusion prior knowledge network) that incorporates prior knowledge into the neural network training process. The model takes the self-similarity of network traffic as a priori knowledge, incorporates it into the gating mechanism of the long short-term memory neural network, and combines a one-dimensional convolutional neural network with an attention mechanism to extract the temporal features of the traffic sequence. The experiments show that the model can better recover the characteristics of the original data. Compared with the traditional prediction model, the proposed model can better describe the trend of network traffic. In addition, the model produces an interpretable prediction result with an absolute correction factor of 76.4%, which is at least 10% better than the traditional statistical model.

The rapid development of information technology has promoted the transformation of traditional networks into intelligent networks. Huge data traffic is generated by various types of traffic services in the intelligent networks, which can easily lead to network congestion, system instability, and other problems. These problems may incur great requirements and pose challenges for queue management algorithms. Most traditional active queue management (AQM) algorithms judge the congestion level of the network based only on the size of the average queue length while ignoring the network traffic variations. This makes these algorithms difficult to achieve effective improvement of congestion control efficiency. To address this problem, a novel network congestion control algorithm, namely the average queue length and change rate-RED (AC-RED) is proposed in this paper. AC-RED can better relieve network congestion by reconfiguring the packet loss strategy model based on the average queue length change rate. The simulation results of NS2 show that in complex and dynamic network environments, the performance of AC-RED algorithm in the average queue length, packet loss rate, delay, and delay jitter is improved in most load conditions compared with the other five algorithms.

Timely and accurate network traffic prediction is a necessary means to realize network intelligent management and control. However, this work is still challenging considering the complex temporal and spatial dependence between network traffic. In terms of spatial dimension, links connect different nodes, and the network traffic flowing through different nodes has a specific correlation. In terms of spatial dimension, not only the network traffic at adjacent time points is correlated, but also the importance of distant time points is not necessarily less than the nearest time point. In this paper, we propose a novel intelligent network traffic prediction method based on joint attention and GCN-GRU (AGG). The AGG model uses GCN to capture the spatial features of traffic, GRU to capture the temporal features of traffic, and attention mechanism to capture the importance of different temporal features, so as to realize the comprehensive consideration of the spatial-temporal correlation of network traffic. The experimental results on an actual dataset show that, compared with other baseline models, the AGG model has the best performance in experimental indicators, such as root mean square error (RMSE), mean absolute error (MAE), accuracy (ACC), determination coefficient (), and explained variance score (EVS), and has the ability of long-term prediction.

In wireless sensor networks (WSNs), base stations (BS) or sink nodes (SNs) play a crucial role in receiving, processing, and transmitting data collected from sensors. Therefore, their traffic prediction performance significantly impacts the efficiency of sensor data transmission. While deep learning (DL) models have been widely utilized in WSNs due to their impressive feature extraction capabilities, they still face challenges related to sample size dependency and interpretability. Motivated by these challenges, this paper proposes a reinforcement learning (RL)-based grey prediction method to enhance the intelligence of traffic prediction in BS or SNs. In environment construction, we initially apply the Savitzky-Golay (SG) filter to mitigate the strong burstiness of network traffic. Subsequently, we introduce an enhanced exponential weakening buffer operator (EEWBO) to improve the small sample prediction performance of grey model GM(1,1) by introducing a new parameter, namely the correction factor (CF). Finally, we theoretically reveal the relationship between the fine-tuning strategy of CF and the buffer effect strength of EEWBO, providing interpretability for the agent’s action selection. During agent training, a sliding window is employed to update the state space, and the CF fine-tuning strategy is mapped to the action space. Additionally, a reward function based on prediction accuracy is designed. Experiments on public datasets show that the proposed RL-EEWBO-GM(1,1) exhibits advantages in sample size dependence, interpretability, and prediction performance when compared with classical statistical and DL models.

The proliferation of edge Internet of Things (IoT) devices intensifies congestion challenges in edge IoT networks. Adjusting edge router buffer sizes is a common and effective solution to reduce queuing delay and mitigate congestion. However, designing an optimal buffer sizing strategy that effectively balances low queuing delay with high-throughput and minimal packet loss remains a key challenge. To address this, this article proposes deep reinforcement learning-based adaptive buffer sizing (DRL-ABS), an adaptive buffer sizing (ABS) method based on deep reinforcement learning. DRL-ABS employs a Dueling DQN model to implement its strategy, effectively controlling buffer size to reduce queuing delay. First, we introduce a novel congestion determination mechanism to enhance decision stability against transient traffic fluctuations. Then, we design a reward function based on the congestion status to balance packet loss rate and throughput. Finally, we construct a NS3-PyTorch joint simulation framework to bridge the gap between NS-3 and artificial intelligence (AI) algorithms. Simulation results demonstrate that, compared to static buffer sizing, traditional ABS, and AI-dreiven ABS, DRL-ABS significantly reduces the average buffer size, average queuing delay, and jitter while maintaining a high-average throughput and a comparable total packet loss rate. Furthermore, DRL-ABS exhibits good stability, maintaining a high average buffer utilization rate across different congestion control systems.

The transition from 5G to 6G is set to transform telecommunications, offering higher data rates, ultra-low latency, and the ability to connect billions of devices simultaneously. Artificial intelligence (AI) is crucial for this evolution, particularly through federated learning (FL), which enables secure model training while preserving data privacy. However, the decentralized nature of FL introduces security challenges, such as model poisoning and inference attacks, especially in sensitive 6G scenarios like smart healthcare and Vehicle to Everything (V2X) systems. To address these issues, we introduce FLGuard, a novel framework that combines FL as the aggregation component with Variational Auto-Encoder (VAE), Generative Adversarial Network (GAN), and Diffusion Model as generative modules, incorporating Rényi Differential Privacy (RDP) as its security mechanism. This architecture aims to improve the accuracy, responsiveness, and security of FL models. Additionally, we provide a case study that highlights FLGuard’s capability to withstand poisoning attacks, demonstrating its effectiveness in safeguarding privacy while enhancing the performance of FL models. Simulation results show that FLGuard consistently achieves high robustness and accuracy under various attack scenarios.