The implementation of an effective intelligent reflecting surface (IRS) assisted millimeter-wave (mmWave) vehicular system critically depends on the proper beamforming design exploiting accurate channel state information (CSI). However, it is difficult to obtain accurate CSI due to the passive characteristics of reflective elements and the high-speed movement of vehicle user equipment (VUE). Furthermore, the channel estimation requires a tremendous amount of pilot overhead. In this paper, an instantaneous CSI aware graph neural network (IA-GNN) model based on unsupervised deep learning (DL) was proposed to implicitly learn the channel features and capture the relationship between the IRS and VUEs for IRS-assisted vehicle-to-infrastructure (V2I) links. The graph attention (GAT) mechanism is introduced to better integrate the correlation of node features into the IA-GNN model and improve the accuracy of mapping relationships. By using the IA-GNN model, different utility functions are taken as system objectives for joint active and passive beamforming optimization. Numerical results demonstrate that the proposed beamforming optimization scheme using the IA-GNN model has significant advantages over the existing schemes in terms of generalization and low pilot overhead.

The density peak clustering (DPC) algorithm is sensitive to the selection of the truncation distance and prone to chaining effects during point assignment. To address these issues, a hierarchical assignment-based adaptive density peak clustering (HADPC) algorithm was proposed in this paper. This algorithm first introduces the concept of natural neighbors to adaptively compute local density, thereby reducing reliance on global parameters. Next, it reveals potential cluster structures by distinguishing between core points and boundary points, forming core clusters within a localized scope based on data distribution. Finally, boundary points are assigned through a weighted voting mechanism, resulting in complete clusters. To evaluate the effectiveness of the HADPC algorithm, comparative experiments are conducted against five baseline algorithms on the synthetic and University of California, Irvine (UCI) datasets. Experimental results indicate that the HADPC algorithm can achieve higher clustering accuracy than the comparison algorithms.

Although the Aquila optimization (AO) algorithm demonstrates promising results in various optimization tasks, it is still easy to fall into local optima, particularly when addressing more intricate problems. To tackle this issue, a reinforced tent-enhanced AO (RTEAO) algorithm that incorporates multiple strategic integrations was proposed in this paper. During the initialization phase, a tent chaotic mapping method is employed to enhance population diversity. To update individual positions iteratively, an elite reverse learning strategy and an optimal retention mechanism are employed to circumvent local optima. Experimental evaluations on the CEC2005 test functions demonstrate significant improvements in both convergence speed and accuracy. Furthermore, the RTEAO algorithm showcases remarkable optimization performance in practical applications, specifically in the calibration of a back propagation (BP) neural network for regression analysis on the Boston Housing Price Dataset and Wisconsin Breast Cancer Dataset, thereby affirming its efficacy and versatility in confronting real-world optimization challenges.

To overcome the sensitivity of the least square twin extreme learning machine (LSTELM) to noise and outliers, a robust fuzzy least square twin extreme learning machine based on fuzzy hyperplane (RFHLSTELM) is proposed. The RFHLSTELM algorithm assigns appropriate weights to each sample according to the fuzzy memberships, which reduces the influence of noise effectively. Additionally, the hyperplane parameters are replaced by symmetric triangular fuzzy numbers, and the fuzzy decision boundary further improves the reliability of classification. Experimental results on synthetic, benchmark, and United States Postal Service (USPS) datasets with noise demonstrate that the RFHLSTELM algorithm outperforms existing algorithms on most datasets.

Deep learning (DL) exhibits great potential in addressing the nonconvex optimization challenges for secure precoding in multiple-input multiple-output (MIMO) communication systems. However, existing studies rarely consider the internal eavesdropping threat where legitimate users may intercept confidential messages intended for other users. To fill this gap, a mixed secure autoprecoder (MSAP) approach is proposed for two-user MIMO broadcast channels with confidential messages. The MSAP jointly designs a data-driven autoencoder (AE) and a model-driven zero-forcing (ZF) constraint to optimize the transmitter and receivers simultaneously, ensuring that each user’s signal is nulled at the unintended receiver to guarantee confidentiality. Extensive simulation results demonstrate that the MSAP achieves stronger security against legitimate eavesdroppers and lower symbol error rate (SER) for reliable transmission, and outperforms existing schemes.

In disaster areas, rescue vehicles (RVs) need to exchange large amounts of data, such as rescue commands and operational experience, to ensure driving safety and rescue efficiency. However, the communication infrastructure in these areas is often damaged, severely affecting normal communication and reducing rescue efficiency. In addition, vehicular networks in disaster scenarios are often untrustworthy and struggle to track malicious behavior, posing potential security risks and inefficiencies in data sharing among vehicles. To address these issues, a secure and efficient data sharing scheme for disaster rescue based on blockchain ( SEDS-BDR) is proposed. First, a blockchain-assisted vehicular network framework tailored for disaster relief is designed. Second, a reputation-based consensus algorithm is developed, which securely and immutably tracks improper behavior and records shared data through the blockchain, thereby improving consensus efficiency and system security. Moreover, since blockchain transactions are typically packaged based on time or transaction fees, different weights are assigned to data based on geographic location and information type in the proposed SEDS-BDR. This ensures that urgent data in disaster areas is prioritized for processing, thereby enhancing rescue efficiency. Experimental results show that the reputation- based consensus algorithm effectively shortens consensus time and improves system security. Additionally, the prioritized transaction packaging mechanism optimizes rescue routes, saving time.

To deploy convolutional neural networks (CNNs) on resource-constrained edge devices and accelerate inference, a hardware-aware CNN optimization algorithm named linear association of weights (LAW) was proposed in this paper. The algorithm reconstructs integer dot-product computations in quantized models, significantly reducing multiplication operations while maintaining accuracy. The implementation works at two granularities. At the single-kernel level, it transforms multiplications into efficient shift-and-add operations by exploiting linear relationships among weights. At the multi-channel level, a cross-channel weight association computation model ensures algorithm generality across different convolutional layers. Based on LAW, a configurable CNN hardware accelerator is designed, featuring an acceleration core that decodes weight association patterns and replaces conventional multipliers with shift-and-add combinations for hardware-level optimization. Experimental results demonstrate that compared with traditional 8-bit integer (int8) quantization, the LAW algorithm reduces 54% of multiplications for 7 × 7 convolution kernels and achieves 63% multiplication reduction for multi-channel 3 × 3 convolutional layers. When evaluated with residual network-18 (ResNet-18) on the ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC2012) dataset, the proposed accelerator achieves 40.21 ms inference latency per image with 70.16% top-1 accuracy and 90.12 giga operations per second (GOPS) peak computation amount performance. As an innovative algorithm specifically designed for vision tasks, LAW provides effective hardware acceleration support for vision-oriented deep learning networks, making it particularly suitable for unmanned systems and industrial inspection.

A novel uncertainty-aware forecasting model, termed Monte Carlo (MC) dropout-convolutional neural network-Kolmogorov-Arnold network-long short-term memory (MCKLNet), is proposed for high-volatility time series forecasting. The model combines a convolutional neural network (CNN) for local feature extraction, a convolutional Kolmogorov-Arnold network (KAN) for feature enhancement, long short-term memory (LSTM) networks for temporal dependency modeling, and MC dropout for Bayesian uncertainty quantification. It aims to address three major challenges in volatile time series prediction: insufficient local feature extraction, inadequate long-term dependency modeling, and limited reliability of probabilistic uncertainty estimation. To evaluate its effectiveness, pork price forecasting is employed as a representative case study due to its high volatility and sensitivity to multiple influencing factors. Experimental results show that MCKLNet consistently outperforms conventional baseline models in both point prediction accuracy and uncertainty interval calibration, demonstrating strong robustness, interpretability, and potential for broader application in economic forecasting, energy demand prediction, financial market analysis, and other complex sequential prediction tasks.