Deep learning has shown great potential in assisting medical experts in diagnosing cerebral infarction through electroencephalogram (EEG) analysis. Classifying EEG abnormality levels is a critical component of this process. However, the inherent complexity of raw EEG signals makes reliable feature extraction challenging, while the scarcity of severe cases leads to substantial class imbalance. The limited size of available datasets further exacerbates these issues, as excessive dependence on generative data may cause classifier overfitting. To address these challenges, we propose a novel framework that integrates a diffusion-based generative model with a hybrid embedding-based few-shot learning strategy for EEG abnormality classification. The framework first extracts a comprehensive set of entropy-based and statistical features from EEG waveforms. The diffusion model incorporates class labels into the noise prediction mechanism, embedding label information into the timestep representation to generate high-fidelity, class-consistent synthetic feature vectors. The hybrid embedding-based few-shot classifier is trained using an episodic learning strategy that optimizes latent-space feature distances under limited data. During inference, the classifier employs a two-step decision procedure in which a preclassifier first estimates the two most probable candidate classes. The prototypical module then performs refined metric-based discrimination within this reduced candidate space, which improves both robustness and generalization. In addition, combining EEG abnormality levels with quantitative clinical scale scores enables more rapid and reliable auxiliary assessment of cerebral infarction. Experimental results demonstrate that the proposed framework yields substantial improvements in EEG abnormality classification accuracy and provides meaningful support for clinical decision-making.

Multimodal emotion recognition (MER) has been widely adopted in affective computing and human–computer interaction. However, real-world multimodal streams frequently suffer from missing or incomplete modalities due to sensor failures, privacy constraints, and heterogeneous acquisition costs, which often causes severe performance degradation for models trained under complete inputs. To address this issue, we propose hierarchical VAE–diffusion for emotion reconstruction (HVDER), a latent-space generative framework for incomplete MER. HVDER first employs modality-specific variational autoencoders (VAEs) to project language, visual, and audio features into a unified low-dimensional latent space with KL-regularized structure. Then, a two-stage coarse-to-fine conditional diffusion module completes missing modality latents by recovering global emotion semantics followed by refining local discriminative details. Finally, the completed and observed modality representations are fused for emotion prediction, optimized end-to-end with a multiobjective loss integrating reconstruction, diffusion matching, cross-modal alignment, and classification supervision. Extensive experiments on CMU-MOSEI and CMU-MOSI validate the effectiveness and robustness of HVDER. Under fixed modality-missing settings, HVDER achieves average ACC2/F1 of 76.8/75.4 on CMU-MOSEI and 73.4/72.9 on CMU-MOSI, outperforming representative baselines across all modality combinations. Under random missing with a high missing rate of MR = 0.7, HVDER maintains ACC2/F1 of 75.4/72.5 on CMU-MOSEI and 68.0/67.2 on CMU-MOSI, indicating a stronger performance lower bound in severely incomplete scenarios. Ablation studies further confirm that latent-space modeling and the coarse-to-fine diffusion design jointly contribute to the main performance gains.

With the popularity and development of smartphones, many mobile applications of various types have emerged. How to recommend mobile applications that match the user’s preferences and usage habits among the massive applications is a problem that needs to be solved. Traditional mobile application recommendation methods cannot dynamically track user behavior and preference changes in time and cannot timely correct the recommendation model, resulting in poor recommendation effects. The continual update of mobile applications will also invalidate the recommendation model based on traffic classification. To solve these problems, this article proposes A Deep Q-Network– (DQN) based traffic classification method for mobile application recommendation with continual learning, which embeds a DQN-based traffic classification model in the mobile terminal and sets up a reward and punishment mechanism to achieve self-supervised learning. By continuously adjusting and optimizing the model, the effectiveness of the traffic classification model is ensured, and the recommendation model is provided with accurate and reliable user behavior data support. Experiments on the ISCX and private datasets show that the proposed method performs better and can effectively guarantee the accuracy of the classification model.

In the framework of 6G-driven Symbiotic Internet of Things (Symbiotic IoT), vehicular edge computing systems enhance the inference efficiency of large-scale AI models through collaboration between road side units (RSUs) and in-vehicle terminals. As a core component of symbiotic systems, RSUs face intermittent challenges posed by renewable energy supply when hosting offloaded vehicular deep learning tasks. Existing energy consumption prediction methods rely heavily on massive sampled data and lack a priori evaluation, leading to significant difficulties and risks in acquiring large-scale measured data: when the remaining power of RSUs is insufficient to sustain the sampling period, the prediction process is forced to interrupt, wasting both the operational energy already consumed and the additional scheduling computation energy. This article proposes a two-stage energy consumption prediction method for symbiotic Internet of Vehicles (IoV) edge nodes, addressing the above challenges through a collaborative mechanism of “a priori rough estimation—dynamic precise prediction.” In the first stage, a nonlinear regression model is established based on limited measured data, combined with analysis of model architecture and hardware instruction-level energy consumption characteristics, to enable a priori energy estimation without actual model execution. This supports lightweight decision-making for RSUs during task allocation, reducing computational overhead compared to traditional sampling methods. In the second stage, real-time energy consumption monitoring is used to dynamically update prediction curves, achieving prediction calibration within extremely short operation cycles to balance prediction accuracy with the energy constraints of edge nodes.Experimental results in typical scenarios of IOV applications demonstrate that the proposed a priori evaluation method achieves an average error of 10% under different data volumes and batch sizes, while the real-time prediction method yields short-term and long-term prediction errors below 3% and 6%, respectively—outperforming existing methods in accuracy. This approach effectively assists RSUs in energy management, reducing energy waste caused by the absence of a priori evaluation and prediction errors, and contributes to the efficient utilization of renewable energy in low-latency scenarios, such as connected vehicles.

Encrypted Traffic Classification (ETC) using Deep Learning (DL) faces two bottlenecks: homogeneous network traffic representation and ineffective model updates. Currently, multimodal-based DL combined with the Continual Learning (CL) approaches mitigate the above problems but overlook silent applications, whose traffic is absent due to guideline violations leading developers to cease their operation and maintenance. Specifically, silent applications accelerate the decay of model stability, while new and active applications challenge model plasticity. This paper presents Multi-ARCL, a multimodal adaptive replay-based distributed CL framework for ETC. The framework prioritizes using crypto-semantic information from flows' payload and flows' statistical features to represent. Additionally, the framework proposes an adaptive relay-based continual learning method that effectively eliminates silent neurons and retrains new samples and a limited subset of old ones. Exemplars of silent applications are selectively removed during new task training. To enhance training efficiency, the framework uses distributed learning to quickly address the stability-plasticity dilemma and reduce the cost of storing silent applications. Experiments show that ARCL outperforms state-of-the-art methods, with an accuracy improvement of over 8.64% on the NJUPT2023 dataset.

Social media platforms such as Weibo and TikTok have become more influential than traditional media. Sentiment in social media comments reflects users’ attitudes and impacts society, making sentiment analysis (SA) crucial. AI driven models, especially deep-learning models, have achieved excellent results in SA tasks. However, most existing models are not interpretable enough. First, deep learning models have numerous parameters, and their transparency is insufficient. People cannot easily understand how the models extract features from input data and make sentiment judgments. Second, most models lack intuitive explanations. They cannot clearly indicate which words or phrases are key for emotion prediction. Moreover, extracting sentiment factors from comments is challenging because a comment often contains multiple sentiment characteristics. To address these issues, we propose a dual-branch combination network (DCN) for SA of social media short comments, achieving both word-level and sentence-level interpretability. The network includes a key word feature extraction network (KWFEN) and a key word order feature extraction network (KWOFEN). KWFEN uses popular emotional words and SHAP for word-level interpretability. KWOFEN employs position embedding and an attention layer to visualize attention weights for sentence-level interpretability. We validated our method on the public dataset weibo2018 and TSATC. The results show that our method effectively extracts positive and negative sentiment factors, establishing a clear mapping between model inputs and outputs, demonstrating good interpretability performance.

Traffic Classification (TC) has been applied to a wide range of applications, from security monitoring to quality of service (QoS) provisioning in network Internet Service Providers (ISPs). In recent years, many researchers have applied Machine Learning (ML) or Deep Learning (DL) to TC, namely AI-TC. However, AI-TC methods face significant challenges, including high data dependency, exhaustively costly traffic labeling, and network subscribers’ privacy. This paper proposes a TC framework for smart home networks using Federated Learning (FL) that protects traffic data privacy by performing local training and inference of TC models. Firstly, we design a DPI-based traffic labeling method on edge home gateways as FL nodes, which enables these nodes to have data labeling capability while protecting data privacy. Then, a semi-supervised TC model based on an autoencoder (AE) is proposed to reduce the dependence of the model on labeled traffic samples. Finally, an XAI-based method is utilized to interpret the model to ensure its explainability. We validate the proposed method on public and real datasets using benchmarking methods. The experimental results show that the method can achieve high performance using a small number of samples while protecting data privacy and improving the model’s credibility. Experimental code can be found in the following url: https://github.com/PrinceXuan12138/HGW-TC-Experimental-code.

The AI-based network traffic classification technology has played a vital role in Zero-touch network and Service Management (ZSM). Network Traffic Features are the critical input for AI-based network traffic classification. However, unlike images and natural language traffic, features have high dispersion and weak stability. On the one hand, part of the key numerical flow features have a wide range of values. By scaling the features, the model can be better trained, which allows the classifier to make more accurate predictions. On the other hand, edge networks cannot scale features as well as the training set in the cloud. This situation affects the model’s performance on edge devices. This paper proposes a novel processing method of flow features for the AI-based network traffic classification model using cloud-edge collaboration. The method leverages the strong computational power of cloud computing to scale the flow features and transmit parameters derived from feature scaling to edge networks. The edge network processes the real-time flow through parameters, significantly enhancing the devices’ accuracy. To validate its effectiveness, we conducted four experiments, comparing models trained with and without feature scaling. The experimental results demonstrate that the model trained with feature scaling outperforms the model trained without scaling.

With the rapid development of the Internet, people pay more and more attention to network security and data privacy. Using the characteristics of SDN data and control separation, it is easy to embed a traffic detection model in edge devices to achieve abnormal traffic detection. However, although the traditional intrusion detection model can provide good recognition accuracy, it requires many labeled samples for model training. Not only is it challenging to obtain labeled samples, but it also brings privacy issues. This paper combines federated learning and anomaly-based CAE model in the SDN network and realizes intrusion detection on encrypted traffic under the premise of effectively protecting data privacy and reducing the workload of data labeling. Furthermore, we design an aggregation model selection algorithm based on loss and data volume evaluation, which reduces the overall training time of the federation and improves the model's accuracy.

With the rapid development of communication net-work, the types and quantities of network traffic data have in-creased substantially. What followed was the frequent occurrence of versatile cyber attacks. As an important part of network security, the network-based intrusion detection system (NIDS) can monitor and protect the network equippments and terminals in real time. The traditional detection methods based on deep learning (DL) are always in supervised manners in NIDS, which can automatically build end-to-end detection model without man-ual feature extraction and selection by domain experts. However, supervised learning methods require large-scale labeled data, yet capturing large labeled datasets is a very cubersome, tedious and time-consuming manual task. Instead, unsupervised learning is an effective way to overcome this problem. Nonetheless, the ex-isting unsupervised methods are prone to low detection efficiency and are difficult to train. In this paper we propose a novel NIDS method called PGAN based on generative adversarial network (GAN) to detect the abnormal traffic from the perspective of Anomaly Detection, which leverage the competitive speciality of adversarial training to learn the normal traffic. Based on the public dataset CICIDS2017, three experimental results show that PGAN can significantly outperform other unsupervised methods like stacked autoencoder (SAE) and isolation forest (IF).

With the popularity of Deep Learning (DL), researchers have begun to apply DL to tackle with encrypted traffic classification problems. Although these methods can automatically extract traffic features to improve the ability of feature engineering of traditional methods like DPI, a large amount of data is still needed to learn the characteristics of various types of traffic. Therefore, the performance of classification model always significantly depends on the quality of datasets. Nonetheless, the building of datasets is a time-consuming and costly task, especially encrypted traffic. Apparently, it is often more difficult to collect a large amount of traffic samples of those unpopular applications than well-known ones, which often leads to the problem of class imbalance between major and minor encrypted applications in datasets. In this paper, we proposed a novel traffic data augmenting method called PacketCGAN using Conditional GAN, which can control the modes of data to be generated. PacketCGAN exploit the benefit of CGAN to generate specified samples with the input of applications’ types as conditional and thereby achieve data balancing. As a proof of concept, three classical DL models including CNN were adopted to classify four types of encrypted traffic datasets augmented by Random Over Sampling (ROS), SMOTE(Synthetic Minority Over-sampling Techinique), vanilla GAN and PacketCGAN respectively using public datasets. The experimental evaluation results demonstrate that DL based encrypted traffic classifier over our new dataset augmented by PacketCGAN can achieve better performance than the other three in terms of encrypted traffic classification.