In hard-label black-box adversarial attacks, where only the top-1 predicted label is accessible, the prohibitive query complexity poses a major obstacle to practical deployment. In this paper, we focus on optimizing a representative class of attacks that search for the optimal ray direction yielding the minimum ℓ₂-norm perturbation required to move a benign image into the adversarial region. Inspired by Nesterov's Accelerated Gradient (NAG), we propose a momentum-based algorithm, ARS-OPT, which proactively estimates the gradient with respect to a future ray direction inferred from accumulated momentum. We provide a theoretical analysis of its convergence behavior, showing that ARS-OPT enables more accurate directional updates and achieves faster, more stable optimization. To further accelerate convergence, we incorporate surrogate-model priors into ARS-OPT's gradient estimation, resulting in PARS-OPT with enhanced performance. The superiority of our approach is supported by theoretical guarantees under standard assumptions. Extensive experiments on ImageNet and CIFAR-10 demonstrate that our method surpasses 13 state-of-the-art approaches in query efficiency.

The exponential growth of deep learning models imposes severe memory constraints on GPUs, significantly increasing training costs. The prevailing solution for addressing memory constraints is tensor offloading, exemplified by ZeRO Infinity, which leverages GPUs, CPUs, and NVMe SSDs to enable large-scale model training. However, ZeRO-Infinity faces significant performance bottlenecks due to memory access imbalance, coarse-grained tensor transfers, NVMe bandwidth and latency limits, and software complexity. Compute Express Link (CXL) emerges as a promising technology for building disaggregated memory pools, yet its integration into large-scale training remains underexplored.This paper introduces an efficient CXL memory pool into the system for tensor offloading and leveraging CXL protocol features for hardware acceleration. The proposed design incorporates NUMA-aware memory allocation and a Dynamic Adaptive Pipelining (DAP) strategy to enhance communication–computation overlap, along with a hybrid object-based memory management scheme to mitigate fragmentation and improve allocation efficiency. Experimental evaluation on an 8- GPU system with CXL Type-3 expansion cards demonstrates up to 72.7% throughput improvement, 62.9% latency reduction, and support for training 1.14× larger models compared with SSD based ZeRO-Infinity. This is the first work to integrate CXL with ZeRO-Infinity for large-scale training, offering practical insights for future CXL-based heterogeneous systems.

Automatic modulation recognition is essential for large-scale wireless communications, but traditional methods often ignore unknown signals in open-set conditions, leading to their incorrect classification as known types and thereby compromising system reliability and communication security. To handle this challenge, a novel automatic modulation open set recognition (AMOSR) model based on deep metric learning with embedding enhancement is proposed in this article. First, for each known example, deep neural model is employed to separately extract the original in-phase and quadrature (IQ) signal and its instantaneous features, which are then fused to obtain embedding. Second, random erasing is employed to the original IQ signal and instantaneous features separately to obtain an augmented example, which has similar structure but different semantics with known example, and the embedding of this example is obtained by using first step. Then, the embedding space, in which tuplet loss and margin loss are combined with the embeddings of known and augmented examples, is trained to improve the overall performance of the model. Finally, after training, AMOSR is implemented using the class centers of known classes. Experiments on three automatic modulation datasets show that our model has better average performance than several mainstream methods in the field of computer vision.

Road traffic state estimation is an essential component of intelligent transportation systems (ITSs). However, some road sections lack fixed detectors, making it difficult to obtain complete traffic state data for the entire city. Thus, through the fusion of data collected from both the fixed and mobile detectors, we propose a novel model framework, a reasoning model based on the multisource causal interpretation graph (MS-CIG), to infer traffic state data for the road sections without detectors. First, a cross-layer random walk strategy is proposed to achieve fusion and embedding of road network topology graphs (constructed by the fixed detector data) and road network logic graphs (constructed by the mobile detector data). Second, the spatial similarity between road sections can be calculated by combining road sections static feature data, Point of Interest (PoI) distribution and embedding features; thereby, a weighted spatiotemporal graph of the road network is constructed. Finally, an interpretive module is combined to provide interpretability analysis for graph-based semi-supervised inference, thereby making the inferred data generated by graph-based semi-supervised more reasonable and accurate. Through the above processing, we are able to obtain the complete traffic state data for the entire city. Experimental evaluations on a real traffic data set demonstrate the superiority of the proposed method.

Automatic Modulation Open Set Recognition (AMOSR) has practical significance in detecting unknown classes. However, a challenge arises when unknown samples closely resemble known samples, posing a formidable task for accurate detection. A novel AMOSR framework based on multi-view discriminators’ joint judgment is proposed to handle this challenge. Firstly, utilizing signal domain knowledge, multi-dimensional features are extracted through varied signal time-frequency transforms and encoders, baesd on which multiple discriminators are created. Secondly, Constrained Clustering Prototype Loss and Geodesic Contrastive Loss are introduced to pretrain these discriminators, providing more space for unknown signals. Then, collaborative learning is employed to further fine-tune the aforementioned discriminators, enhancing information sharing between modalities. Furthermore, a set of indicators is constructed, and multi-criteria fusion is performed using the TOPSIS algorithm to evaluate the discrimination capabilities of different classifiers in both closed-set and open-set scenarios. Furthermore, a decision tree is constructed to segregate test signals into known and unknown classes, in which discriminators with higher confidence levels are given precedence. Finally, TOPSIS hierarchical ensemble pruning algorithm that considers diversity and open-set recognition capabilities is adopted to reduce model complexity while maintaining original performance. Extensive experiments conducted on modulation datasets demonstrate the superiority of this framework over state-of-the-art AMOSR results.

Automatic modulation recognition (AMR) of radio signal is an important research topic in the area of non-cooperative communication and cognitive radio. Recently deep learning (DL) techniques enable significant progress in AMR. However, the techniques of adversarial machine learning cause the threats of adversarial attacks in DL-based AMR. In this paper, we aim to make AMR model robust, accurate and lightweight, thus propose a multi-distillation mechanism for robust training of DL-based AMR models, namely Adversarial Multi-Distillation (AMD). In the framework of AMD, by knowledge distillation, two powerful teacher models transfer the learned classification knowledge and defense knowledge, respectively, to the student model to form robust training. Our experiments with public dataset RML2016.10a show that the proposed method can significantly improve the defense of AMR models to against adversarial perturbations and keep relatively high classification accuracy, which enables robust decision making with lightweight models under adversarial attacks.

With the growing traffic congestion problem, more and more deep reinforcement learning (DRL) methods have been applied in traffic signals control(TSC). But researches show that DRL is very fragile with abnormal data. In this paper, special traffic state abnormal data (TSAD) are simulated, based on which the robustness of DRL is analyzed and improved for traffic signals control. Firstly, the perturbation noise is generated based on the Discrete Carlin&Wagner attack, which is then added to the normal data to simulate the TSAD. Secondly, under different type of TSAD, the robustness of DRL models for traffic signals control is explored, which are demonstrated to have certain vulnerability, especially with high traffic flows. Finally, induction model based on reward detection(IMR) and mask the activation values of decision neurons (MVN) are proposed to effectively improve the robustness of DRL models for traffic signals control.

The future trajectory prediction of heterogeneous traffic-agents for autonomous vehicles in mixed traffic scene is of great significance for safe and reliable driving. Thus, we propose the Multi-View Adaptive Hierarchical Spatial Graph Convolution Network (MVHGN) to predict the future trajectories of heterogeneous traffic-agents. Firstly, multiple logical correlations are obtained based on the time series data of traffic-agents and a multi-view logical network is constructed. The multi-view logical feature extraction is realized based on the graph convolution module. Then, combining the multi-view logical features and the adaptive spatial topology network, the logical-physical features at the micro level are mined through the graph convolution module; based on the logical-physical features at the micro level and the regional clustering network at the macro level, the global logical-physical features are obtained. Finally, the model predicts the future trajectories of traffic-agents based on the encoder-decoder structure of the GRU. For the Apolloscape trajectory data set, the performance of our proposed method MVHGN is better than that of the comparison models.

Deep learning based approaches have achieved great success on the automatic cardiac image segmentation task. However, the achieved segmentation performance remains limited due to the significant difference across image domains, which is referred to as domain shift. Unsupervised domain adaptation (UDA), as a promising method to mitigate this effect, trains a model to reduce the domain discrepancy between the source (with labels) and the target (without labels) domains in a common latent feature space. In this work, we propose a novel framework, named Partial Unbalanced Feature Transport (PUFT), for cross-modality cardiac image segmentation. Our model facilities UDA leveraging two Continuous Normalizing Flow-based Variational Auto-Encoders (CNF-VAE) and a Partial Unbalanced Optimal Transport (PUOT) strategy. Instead of directly using VAE for UDA in previous works where the latent features from both domains are approximated by a parameterized variational form, we introduce continuous normalizing flows (CNF) into the extended VAE to estimate the probabilistic posterior and alleviate the inference bias. To remove the remaining domain shift, PUOT exploits the label information in the source domain to constrain the OT plan and extracts structural information of both domains, which are often neglected in classical OT for UDA. We evaluate our proposed model on two cardiac datasets and an abdominal dataset. The experimental results demonstrate that PUFT achieves superior performance compared with state-of-the-art segmentation methods for most structural segmentation.

Road traffic state estimation is an essential component of intelligent transportation systems (ITSs). However, road traffic state data collected by traffic detectors are often incomplete, which can cause problems across a variety of transportation applications, such as traffic state prediction and pattern recognition. We present GA-GAN (Graph Aggregate Generative Adversarial Network), consisting of graph sample and aggregate (GraphSAGE) and a generative adversarial network (GAN), to impute missing road traffic state data. Instead of using the original road network structure, which presents the spatial information to process a graph operation, we reconstruct the road network according to the correlation coefficients of road historical data. We utilize GraphSAGE to aggregate the temporal-spatial information from the neighbors of each road in the reconstructed road network. GAN is used to generate complete traffic state data from the extracted temporal-spatial information to achieve traffic state data imputation. To illustrate the efficient performance of the model, experiments are conducted on traffic data collected from California and Seattle, Washington, showing that the proposed model outperforms state-of-the-art methods.

With the improvement of living standards, people pay more attention to health, which is significant to analyze people's hospitalizing behaviors. The wide use of mobile devices generates a great deal of data, which contains a lot of travel information about residents. Many people would like to see a doctor through calling an online car hailing for its convenience. Thus, based on big trajectory data generated by the online car hailing, the hospitalizing behaviors of residents are analyzed in this paper. The hospitalizing behaviors are analyzed from two aspects. One is performed from the temporal aspect, in which the daily numbers of trips of hospitalizing behaviors under different modes are analyzed. The other one is performed from the spatial aspect, in which the hot hospitals, popularity, and gravity distribution of hospitals are analyzed. Based on the spatial analysis, the network constructed by the hot hospitals is also analyzed. The results show that the hospitalizing behavior analysis can reflect the hospitalizing behaviors in detail, which can make contributions to the decision-making of infrastructure configuration for institutions, such as urban planning departments and hospitals.

The realization of road traffic prediction not only provides real-time and effective information for travelers, but also helps them select the optimal route to reduce travel time. Road traffic prediction offers traffic guidance for travelers and relieves traffic jams. In this paper, a real-time road traffic state prediction based on autoregressive integrated moving average (ARIMA) and the Kalman filter is proposed. First, an ARIMA model of road traffic data in a time series is built on the basis of historical road traffic data. Second, this ARIMA model is combined with the Kalman filter to construct a road traffic state prediction algorithm, which can acquire the state, measurement, and updating equations of the Kalman filter. Third, the optimal parameters of the algorithm are discussed on the basis of historical road traffic data. Finally, four road segments in Beijing are adopted for case studies. Experimental results show that the real-time road traffic state prediction based on ARIMA and the Kalman filter is feasible and can achieve high accuracy.

Differential evolution is a fast, robust, and simple population-based stochastic search algorithm for global optimization, which has been widely applied in various fields. However, there are many mutation strategies in DE, which have their own characteristics. Therefore, choosing a best mutation strategy is not easy for a specific problem. Different mutation strategies may be appropriate during different stages of the evolution. In this paper, we propose a DE with multi-stage strategies (DEMS). In DEMS, the evolution process of DE is divided into multiple stages according to the average distance between each individual in the initial population. Each stage has its own strategy candidate pool which includes multiple effective strategies. At the beginning of each generation, the average distance between each individual is first calculated to determine the evolution stage. Then for each target vector in the current population, a mutation strategy is randomly selected from the strategy candidate pool with respect to the stage to produce a offspring vector. Numerical experiments on 15 well-known benchmark functions and the CEC 2015 benchmark sets show that the proposed DEMS is significantly better than, or at least comparable to several state-of-the-art DE variants, in terms of the quality of the final solutions and the convergence rate.