Chun Wang

Multi-energy load forecasting in integrated energy systems (IES) is a critical component for ensuring the stability of energy supply and the reliable operation of economic activities. The energy load of IES, along with Meteorological variable, exhibits spatio-temporal coupling, random fluctuations, and irregular periodic trends, posing significant challenges to the development of accurate and stable prediction models. To this end, this paper proposed a Kolmogorov-Arnold Networks Meet Science (KAN2.0) algorithm with Spatio-temporal and trend (ST3-KAN2.0) feature for multi-energy load prediction. Firstly, the proposed algorithm utilizes a Spatio-temporal Graph Convolutional Network (STGCN) based on a Gaussian kernel-constructed spatial matrix to extract features from multi-energy loads and associated meteorological variables. Subsequently, a Convolutional Neural Network (CNN) is employed to capture trend information decomposed by Seasonal-Trend decomposition procedure (STL). Then, the extracted spatio-temporal features and future trend information are integrated using a feature fusion method. Finally, an structure-aware KAN2.0 algorithm is applied to build a multi-energy joint prediction model based on the extracted features. The accuracy and robustness of the proposed algorithm are validated using real-world datasets. Verification results demonstrate that the proposed algorithm outperforms the STGCN-CNN and Informer algorithms in prediction accuracy by notable margins. Additionally, the model exhibits superior computational efficiency, with low memory consumption during training and fast processing speed on the test set. Furthermore, robustness tests indicate that the algorithm maintains stable performance even under varying levels of noise interference.

In the Internet of Vehicles (IoV), certificateless-based authentication has emerged as a promising approach to secure federated learning (FL) participation. However, existing certificateless-based authentication schemes still exhibit three notable limitations: 1) some schemes cannot provide security against public-key replacement (PKR) attacks and key-leakage threats; 2) most existing schemes do not take realistic coalition attacks into account; and 3) while some schemes applicable to FL settings can offer relatively comprehensive security guarantees, they often rely on costly cryptographic operations. These limitations make it challenging for certificateless authentication frameworks to simultaneously ensure high security and low latency, particularly in resource-constrained FL deployments within IoV environments. To address these challenges, we propose FedAuth, a certificateless, lightweight, and privacy-enhanced authentication scheme tailored for FL in IoVs. First, FedAuth establishes a structured pseudonymization mechanism that enables vehicles to anonymously participate in the FL process while maintaining accountability. Afterward, FedAuth ensures robust authentication security by cryptographically anchoring key generation to each vehicle’s pseudonym, enforcing strict key validation, and introducing hash-bound multilayered signature verification that prevents signature recombination, thereby resisting PKR attacks, coalition attacks, and key-leakage threats. Second, FedAuth guarantees the authenticity and integrity of gradients through lightweight signature construction, which supports batch authentication while avoiding expensive operations, such as bilinear pairings, thereby reducing latency. Security analysis shows that FedAuth preserves identity privacy and ensures model integrity in the face of adversarial conditions. The experimental results confirm that FedAuth significantly reduces computational overhead by at least 43.42% and saves not less than 19.8 kB of communication cost per 100 vehicles, outperforming state-of-the-art schemes in both efficiency and robustness.

In Vehicular Ad-hoc Networks (VANETs), federated learning (FL) enables collaborative training a global intelligent transportation model without sharing vehicles’ raw data. Achieving model convergence in VANETs requires multiple rounds of FL for aggregation and updates. To improve the efficiency of model convergence, researchers explore methods of multi-round aggregation. The latest work, Flamingo (S&P 2023), proposes a multi-round aggregation scheme based on a reusable key mechanism. Its application in VANETs with dynamic network structures enhances the efficiency of model training. However, this scheme has a defect: vehicles cannot verify the correctness of the aggregation result. Once a semi-trusted server returns incorrect aggregation results due to computational errors, it may reduce the model’s accuracy or even cause training failure. In this scheme, we propose a verifiable multi-round aggregation scheme for FL in VANETs (VMFL), enabling vehicles to verify the aggregation results. Firstly, a multi-round verification mechanism is designed to reduce the number of interactions for vehicles by using reusable keys. Additionally, we propose a lightweight proof scheme that allows vehicles to verify the results with minimal computation, reducing the computational overhead of the verification process. Finally, a security analysis of VMFL is performed to demonstrate its security in cases of vehicle dropout. We validated the efficiency of VMFL across different datasets through experiments, showing no significant increase in time overhead compared to Flamingo, and demonstrating that its verification time is reduced by about 90% compared to the state-of-the-art scheme, thereby demonstrating the scheme’s usability.

Session-based recommendation (SBR) aims to predict the next item a user may interact with based on an anonymous session, playing a crucial role in real-time recommendation scenarios. However, existing SBR models struggle to effectively capture local session dependencies and global item relationships, while also facing challenges such as data sparsity and noisy information interference. To address these challenges, this paper proposes a novel Multi-View Graph Contrastive Learning Neural Network (MVGCL-GNN), which enhances recommendation performance through multi-view graph modeling and contrastive learning. Specifically, we construct three key graph structures: a session graph for modeling short-term item dependencies, a global item graph for capturing cross-session item transitions, and a global category graph for learning category-level relationships. In addition, we introduce simple graph contrastive learning to improve embedding quality and reduce noise interference. Furthermore, a soft attention mechanism is employed to effectively integrate session-level and global-level information representations. Extensive experiments conducted on two real-world datasets demonstrate that MVGCL-GNN consistently outperforms state-of-the-art baselines. MVGCL-GNN achieves 34.96% in P@20 and 16.50% in MRR@20 on the Tmall dataset, and 22.59% in P@20 and 8.60% in MRR@20 on the Nowplaying dataset. These results validate the effectiveness of multi-view graphs and contrastive learning in improving both accuracy and robustness for session-based recommendation tasks.

Traditional node classification is typically conducted in a closed-world setting, where all labels are known during training, enabling graph neural network methods to achieve high performance. However, in real-world scenarios, the constant emergence of new categories and updates to existing labels can result in some nodes no longer fitting into any known category, rendering closed-world classification methods inadequate. Thus, open-world classification becomes essential for graph data. Due to the inherent diversity of graph data in the open-world setting, it is common for the number of nodes with different labels to be imbalanced, yet current models are ineffective at handling such imbalance. Additionally, when there are too many or too few nodes from unseen classes, classification performance typically declines. Motivated by these observations, we propose a solution to address the challenges of open-world node classification and introduce a model named OWNC. This model incorporates a dual-embedding interaction training framework and a generator–discriminator architecture. The dual-embedding interaction training framework reduces label loss and enhances the distinction between known and unseen samples, while the generator–discriminator structure enhances the model’s ability to identify nodes from unseen classes. Experimental results on three benchmark datasets demonstrate the superior performance of our model compared to various baseline algorithms, while ablation studies validate the underlying mechanisms and robustness of our approach.