Differential privacy is widely applied in deep learning to mitigate the risk of sensitive information leakage. However, additional privacy protection measures, such as per-example gradient clipping, introduces extra computational overhead and storage requirements, which impede the exploration of differentially private deep learning methods. To address this challenge, we propose a novel differentially private low-rank weight fusion approach. Our approach utilizes a model weight composed of a shared weight and low-rank weights. The rank of the low-rank weights determines the number of sub-networks involved in the fusion process. During the training phase, we fix the shared weight and solely train the low-rank weights. After each training batch, we add noise to safeguard the privacy of the low-rank weight fusion operation, thereby achieving privacy-preserving training. This approach significantly reduces storage overhead and eliminates the need for per-example gradient computations. Additionally, the low-rank weight can be applied as a plug-in, particularly in transfer learning scenarios. We evaluate the knowledge transfer capability of our method and conduct experiments in both self-learning and transfer learning scenarios. Our results demonstrate that our approach achieves comparable accuracy to the batchensemble method without introducing additional noise.
This paper presents a simple but efficient solution to broaden the impedance bandwidth, suppress the mutual coupling (MC), and extend the scanning range of a single-layer microstrip patch phased array. By simultaneously exciting and adjusting the electric fields (E-fields) of the transverse magnetic (TM)10 and TM20 modes, dual resonances are supported to enhance the impedance bandwidth, and weak fields are generated on coupled patches to achieve the MC reduction. Furthermore, the radiation fields of the TM10 and TM20 modes are combined inherently within the impedance band to achieve the wide-beam pattern. The proposed solution can also be applied to a linear phased array in which the above three functions are integrated for wideband wide-angle scan. A fabricated prototype of the phased array is measured for performance validation. Measured results confirm that the designed array can achieve its scanning range of ±60° with good impedance matching and high inter-port isolation across the −10 dB active reflection band from 5.56 to 5.98 GHz (7.3%).
Deploying deep neural networks on compute-in-memory (CIM) hardware is challenged by analog non-idealities such as device nonlinearity and circuit-level noise. This paper proposes a cross-layer framework that enables accurate and efficient deployment of a compressed sensing-based object detection (CSDet) model on a CIM platform built with programmable linear random-access memory. The framework incorporates device- and circuit-level error modeling into the training and deployment flow, enabling slope-aware compensation and statistical correction of residual errors. Experimental results show that the calibrated CSDet achieves up to 11.3% improvement in detection accuracy with minimal hardware overhead, demonstrating the practicality of robust inference on non-ideal CIM systems.
Group action quality assessment (GAQA) aims to assess the quality of coordinated movements performed by multiple individuals. Most existing methods often overlook structural formation patterns that are essential for understanding coordination in multi-person GAQA. In this paper, we propose a formation-based multimodal scoring network that incorporates formation transcripts and visual cues for comprehensive evaluation. Specifically, we first leverage the sequence reasoning capabilities of large language models to capture both the technical and artistic semantics from formation transitions. Then, we introduce a shared feature extraction module to learn modality-invariant representations across visual and formation modalities, and a modality-specific extraction module to retain unique and complementary modality-specific cues. Extensive experiments on three challenging datasets (i.e., LOng-form GrOup, OlympicFS, and FS1000) demonstrate the effectiveness of our method.
The rapid development of multimodal transportation poses dual challenges for secure cross-modal trajectory data management. Addressing the core issues of spatial structure distortion and spatiotemporal fragmentation in existing methods, which result in reduced usability and credibility of the generated trajectories. This paper proposes trajectory rule-aware privacy guardian (TRAP-Guard), a dual-layer road network-constrained privacy preservation model. The model utilizes sequential path variational autoencoder based on variational autoencoder architecture with path completion to effectively improve the spatial topological consistency of the generated trajectories. During the trajectory point-level perturbation phase, by integrating the global candidate set of the road network, the exponential mechanism, and temporal constraints, fine-grained completion satisfying graph-geo-indistinguishability is achieved. Experimental validation on real-world public transportation datasets confirms that TRAP-Guard significantly enhances spatial rationality, temporal consistency, and overall utility of privacy-preserved trajectories.
Graph searchable encryption (GSE) ensures privacy protection for graphs while maintaining search functionality. A key aspect of GSE is exact shortest distance search, which identifies the closest relation between nodes in an encrypted graph. However, even minimal leakage from updates and queries can be exploited by attackers to infer the database or query content. Current GSE schemes supporting exact shortest distance queries overlook the potential information leakage from frequent relationship changes, particularly edge deletions, which compromise data privacy. This paper introduces the concept of backward privacy in GSE and formalizes three types of backward privacy (type-I, type-II, and type-III), a novel contribution. We then propose two backward-private GSE schemes, Bayonet and Baccarat. Bayonet offers fine-grained control over data decryption with type-III backward privacy, while Baccarat enhances type-II backward privacy by delegating edge deletion to a non-collusive proxy. We provide formal security proofs for both schemes and evaluate their efficiency on ten real-world graph datasets. Experimental results show that Bayonet sacrifices server storage and computational overhead for type-III backward privacy, while Baccarat achieves higher security and matches the query efficiency of the state-of-the-art GSE scheme with only forward privacy, incurring minimal proxy storage and computational costs.
With the rapidly growing demand for high density integrated circuits, advanced packaging technologies such as chip stacking and heterogeneous integration have become mainstream. However, reliability issues related to interconnect structures under thermal cycling loads have become increasingly prominent. This study systematically investigates the structural deformation of internal interconnects in stacked SiP modules caused by thermomechanical stress under temperature cycling conditions, with a specific focus on fatigue failure induced by thermal loads. Stress strain analyses are performed on both the upper and lower layers of bumps and the through silicon via (TSV) array to identify critical failure locations and predict thermal fatigue life. The results indicate that deformation is most severe at the corners of the stacked SiP module throughout the temperature cycling process, and these corner regions are also the most prone to failure for both bumps and TSVs. The bottom layer bumps experience higher stress than the top layer bumps and are thus more likely to suffer thermal fatigue failure. The critical failure nodes for the bumps are located at the contact interface between the upper surface of the corner bumps and the lower surface of the redistribution layer (RDL), with a predicted thermal fatigue life of 4.63 × 104 cycles. For the TSV structures, the maximum equivalent stress of 226.41 MPa occurs at the corners of the interposer at the bottom of the stacked SiP module, corresponding to a predicted thermal fatigue life of 5.45 × 106 cycles. In addition, increasing the bump height, reducing the bump pitch, and using SAC405 solder with a higher silver content are found to effectively enhance the thermal reliability of the bumps.
Dynamic vehicle detection is of great significance for perceiving the dynamic development trend of the environment and plays an important role in construction of static environment and accurate localization. To detect vehicles robustly and quickly in various scenarios without training a large amount of data, a vehicle detection approach based on a novel pose estimation method is proposed using LiDAR remotely-sensed data. The distribution characteristics of point clouds on a vehicle surface are analyzed and described by normal vector histogram features. Considering the characteristics of vehicle surfaces, a vehicle measurement model consisting of four rectangles is proposed. Combined with the Kullback-Leibler divergence measure, the fitting degree between the point cloud clusters and a measurement model can be obtained, and the position and attitude of a candidate vehicle can be determined. With the estimated pose information, vehicle motion characteristics are utilized to determine the candidate vehicles. Finally, experiments on karlsruhe institute of technology and toyota technological institute datasets are implemented to verify the superiority of the proposed detection approach. Particularly, in distant areas, the F1 score is increased by 9.6% compared with the previous pose estimation based on the coherent point drift method.
Change detection in high-resolution (HR) remote sensing images has emerged as a fundamental task in Earth observation, aiming to accurately identify and delineate land surface changes over time. Nevertheless, its performance is often constrained by pseudo-changes arising from style discrepancies between bi-temporal images, as well as those introduced by local shadows. Recently, Mamba-based architectures have attracted significant interest due to their efficient modeling capabilities. Building upon this foundation, this paper proposes an efficient-Mamba feature interaction and multiscale refinement (EFMR) method for HR remote sensing image change detection, which enhances overall detection accuracy while effectively mitigating pseudo-changes. Firstly, EFMR employs the efficient Mamba feature interaction module for efficient spatiotemporal alignment, emphasizing structural differences and suppressing pseudo-changes caused by style discrepancies. To mitigate shadow-induced pseudo-changes, EFMR utilizes the multiscale feature refinement module to refine multi-level change features through contextual modeling and residual enhancement. Additionally, to recover spatial details lost during upsampling, EFMR integrates a decoder based on the enhanced vision state space block, combining two-dimensional state-space modeling with multi-kernel depth-wise convolutions. On three public datasets, qualitative and quantitative experiments show that the proposed method achieves superior performance over nine state-of-the-art techniques in both overall accuracy and detailed visualization.
While vision-language models (VLMs) demonstrate remarkable capabilities in autonomous driving applications, they face significant challenges in safety-critical environments due to their insufficient risk awareness, susceptibility to visual distractions, and challenges in effectively balancing safety requirements with learning efficiency. To mitigate these challenges, we propose safe-GRPO, a safe-aware group relative policy optimization framework specifically designed for safe decision-making in VLM-based driving systems. Safe-GRPO introduces three key innovations: i) an uncertainty-aware, safety-sensitive reward mechanism that dynamically adjusts reward and penalty scales according to both risk severity and model uncertainty; ii) an uncertainty-guided noise-enhanced training method that adaptively improves robustness against visual interferences; and iii) a risk-aware dynamic clipping strategy that controls policy updates based on scenario-specific risks. Extensive experiments on an enhanced real-world scenarios with driving theory tests dataset (3926 risk-annotated scenarios) show that safe-GRPO significantly outperforms existing approaches in decision accuracy, robustness, and risk prediction accuracy-achieving a 75.14% safe decision rate in high-risk scenarios, demonstrating notable advancements towards reliable real-world autonomous driving applications.
With the increasing bandwidth and frequency, the geometric complexity of packaging, board-level, and interconnect systems are continuously increased. Direct full-wave electromagnetic simulation of these passive systems consumes enormous computational resources. To simplify the analysis, macromodeling the electromagnetic responses of passive systems has become common practice. Although vector fitting is an excellent macro-modeling method, it typically requires users to manually specify the fitting order, which determines the accuracy. In this paper, a universal automatic pole determination method is proposed for vector fitting to achieve accurate macromodeling for single-input/multiple-output systems. This approach directly produces a rational function macromodel from the given electromagnetic response, and it performs robustly on both systems with noise and large-scale multiport interconnect systems.
Multimodal transport plays a vital role in modern logistics by integrating multiple transportation modes to improve efficiency and reduce costs. However, challenges such as fragmented information, lack of standardized electronic documentation, as well as critical privacy, security, and trust vulnerabilities hinder its development. This review offers a comprehensive analysis of privacy-preserving and secure data flow in multimodal transport, illustrating its architecture, limitations, and adaptive solutions in cross-domain logistic collaboration. We highlight the prospects of blockchain technology, particularly consortium blockchains, in enhancing trust, data regulation, and cross-domain collaboration, while identifying current limitations in ensuring privacy-preserving and secure interactions. Advanced solutions, including privacy-preserving sharing and robust authentication methods adapted to cross-domain multimodal transport scenarios that incorporate blockchain, are reviewed to address these gaps. Finally, we outline future research directions aimed at building secure, scalable, and interoperable multimodal transport systems.
Malware is malicious software invented to harm or exploit computer networks, systems, and devices. The detection of malware is a critical issue in today's hybrid environment, which consists of Windows and Android operating systems. Traditional malware detection practices, such as signature-based detection, often do not detect new and advanced malware. To address this concern, researchers have proposed various advanced approaches, including static and dynamic analysis, by dint of machine learning and ensemble learning, and deep learning for effective malware detection for both Android and Windows separately. However, still, they can't properly handle the issues of model interpretability, concept drifting and high false positive rate. To solve these issues, this research investigates a novel technique for malware identification in a hybrid environment, where both Windows and Android operating systems are analyzed together through a unified framework that integrates both static and dynamic evaluation techniques. Alternatively, if separate analysis is used: this research investigates a technique where separate analyses are conducted for Android and Windows environments, leveraging both static and dynamic evaluation techniques tailored to each platform. The performance of the proposed approach is evaluated through three experimental trials using a balanced, benchmarked, updated and large dataset from both the Android and Windows operating systems to create a hybrid environment. A model is proposed to evaluate the accuracy of malware detection which handles the issue of model interpretability, concept drifting and high false positive rate. It achieves 95.50% accuracy, 98.85% precision, 96.16% recall, 0.06% false positive rate (FPR), 96.60% F1-score and 96.4%. The results highlighted that the proposed approach sur state-of-the-art malware detection techniques.
Accurate network traffic prediction is crucial for modern network management. To overcome the limitations of traditional models in handling the nonlinear and multiscale dynamics of real traffic, this paper proposes deep spatiotemporal graph convolutional network (DeepSTGCN). Its core design integrates a dual-channel graph convolution framework, which captures linear and nonlinear spatial dependencies using Pearson and Spearman correlation-based graphs, and a temporal module with stacked gated recurrent units and attention. We further introduce a graph contrastive learning mechanism to enhance model robustness and generalizability by enforcing node embedding consistency under perturbations. Evaluated on multiple real-world datasets, DeepSTGCN outperforms state-of-the-art baselines in accuracy and reliability. Ablation studies confirm the contribution of each component. The model offers a robust solution for traffic forecasting, with strong potential to improve proactive resource allocation and quality of service.
With the increasing adoption of heterogeneous multi-chip integrated systems in high-performance computing, thermal coupling and non-uniform temperature distributions have become critical concerns, demanding fast and accurate steady-state thermal prediction methods. While recent approaches employing neural networks-such as convolutional neural networks and graph convolutional networks (GCNs)-offer improved computational speed, they often suffer from limitations in prediction accuracy and structural adaptability. To address these challenges, this paper proposes a multi-scale GCN-based method for rapid thermal field prediction. The proposed approach constructs graph structures by treating each pixel in the input maps as a node and dynamically aggregates node features across adjacency relationships defined at multiple spatial scales (scale 1, 2, and 3). This multi-scale, dynamic aggregation mechanism enables capturing thermal propagation patterns effectively at varying spatial resolutions. The network consists of a multi-scale encoder and a GCN-based decoder, trained with Softplus activation and dropout regularization. A simple linear iterative clustering-augmented power/material-discontinuity adaptive partition is developed to enhance boundary fidelity with fewer nodes, at a modest runtime cost. Furthermore, using a dataset we developed comprising over 200 chiplet configurations with diverse and flexible power-material layouts, the proposed method achieves a temperature prediction mean error within 0.68 °C. This approach delivers both high accuracy and computational efficiency, offering a viable solution for thermal-aware modeling and rapid optimization in complex multi-chip systems.
As remote sensing technology continues to evolve rapidly, multisource data collaboration has become a research hotspot in the field of Earth observation. Compared to single-source approaches in environmental monitoring, disaster assessment, precision agriculture, and other domains, multisource remote sensing demonstrates significant advantages in monitoring reliability and complex scenario interpretation capabilities. This paper systematically reviews the theoretical advances, typical applications, and future challenges of multi-source remote sensing technology. First, we analyze the complementary advantages of multisource remote sensing data in spatial-spectral resolution and terrain characterization capabilities, including hyperspectral, multispectral, and synthetic aperture radar (SAR). Next, we explore algorithmic frameworks for pixel-level, feature-level, and decision-level fusion, as well as current trends in multi-modal remote sensing foundational models. Furthermore, this paper identifies key technical challenges, such as the heterogeneity of multisource data, complexity of cross-modal feature association, insufficient model generalization, and computational efficiency bottlenecks. Finally, we provide an in-depth analysis of future research directions to advance theoretical foundations and technical solutions in remote sensing information science.
Change detection (CD) is crucial for a wide range of real-world applications. While numerous CD methods have demonstrated improved performance, many existing approaches over-rely on image modality information, which can lead to inaccurate representations of change information. To address this issue, we propose a novel approach, contrastive learning for improved image-text alignment (CLITA) in hyperspectral change detection. First, the difference features are fed into a three-dimensional skip connection module to extract both spatial and spectral features. Then, we introduce a novel coarse-grained and fine-grained text representation to better capture both change and unchanged information for CD tasks. Moreover, linguistic features are utilized as a shared cross-domain semantic space, with visual-linguistic alignment achieved through contrastive learning in this semantic space. Extensive experiments on several publicly available datasets demonstrate the superior performance of the proposed method.
A mixed-order tetrahedral spectral element method (TSEM) is implemented for nonlinear electro-thermal bidirectional coupling simulation of semiconductor devices. This method combines the high-order accuracy of the spectral method with the strong geometric adaptability of the finite element method. And it can alleviate the burden of multi-scale computation by using different interpolation orders based on the same mesh in different physical fields, and improves the accuracy and convergence of solving strongly nonlinear problems by increasing the interpolation order. In addition, the TSEM constructs system matrix equations directly through Warp&Blend node-optimized Vandermonde and differentiation matrices, improving the efficiency, generality, and flexibility of the multiphysics preprocessing. Then, we apply this method to study temperature-dependent effects in semiconductor devices, fully considering temperature-dependent drift and diffusion current terms. This provides a robust solution to the growing simulation and design challenges posed by the continuously increasing power density. Compared with the commercial software, several numerical experiments are conducted to demonstrate the superior stability, convergence, and efficiency of the proposed algorithm.
Silicon carbide metal-oxide-semiconductor field-effect transistors (SiC MOSFETs) show significant advantages in diverse application scenarios due to their superior performance. However, the existing simulation program with integrated circuit emphasis models developed for low-voltage devices cannot be effectively applied to high-voltage (> 3.3 kV) SiC MOSFETs in practical applications, which may cause unreliable circuit simulations. This paper presents an empirical model for 3.3 kV SiC MOSFETs that considers temperature effects. The model focuses on static and dynamic characteristics using non-segmented equations, ensuring smooth and accurate predictions of the device's output and transfer characteristics. Besides, temperature-dependent parameters are incorporated into the equations to correctly reflect the correlation between characteristics and temperature.For the capacitance model, the new compensation equation is added to the commonly used capacitance fitting equations, and the accuracy of this enhanced capacitance model is verified by fitting the capacitance experimental results of the device using the new capacitance equations. The body diode modeling adopts a combined approach of parameter extraction using the model editor and equation fitting. The adoption of non-segmented effectively avoids convergence problems inherent in traditional segmented models, while the integration of compensation equations brings simulation results closer to real operating conditions. Moreover, the incorporation of temperature-dependent parameters expands the model's applicability, enabling it to better support related design processes.
In this paper, a wideband composite two-dimensional (2D) beam steering loop magnetoelectric dipole array antenna with enhanced switching resolution and reduced mutual coupling is proposed, analyzed and measured. The enhancement of beam-switching resolution is achieved through the spatial transformation and integration of the multi-port beam-steering elements. Specifically, the proposed eight-beam magnetoelectric (ME) dipole array is constructed by integrating two ME dipole arrays, where each array can generate four beams in different azimuthal directions. Furthermore, a characterization method is proposed in this paper for considering the coupling relationship between multi-port integrated antenna arrays. The slotted ground structure is adopted to reduce the mutual coupling. The design guidelines and parameter analysis of the proposed antenna are provided, and the antenna is fabricated and measured. The measured results show that the peak gain of proposed antenna is 8.02 dBi, the overlapped bandwidth for all eight beam states is 29.9% (4.52 GHz to 6.11 GHz) and an angular resolution of 45° for 2D beam switching. By feeding each port respectively, the proposed antenna can realize 2D wideband beam switching with an angular resolution of 45°. The proposed antenna possesses advantages in low-cost, wideband eight-beam switching capability and easy processing, thereby providing a promising candidate for wideband wireless beam-steering applications.
