With the rapid advancement of text-to-image (T2I) diffusion models, personalized generation has emerged as a prominent research focus. However, existing methods face significant challenges when attempting to disentangle multiple concepts from a single image, such as complex data preparation and severe overfitting. To address these issues, we propose ConceptCraft, an attention-based multi-concept disentanglement method. ConceptCraft leverages self-attention and cross-attention maps to automatically generate precise masks in a single initialization step, without relying on manual annotations or external segmentation models, thereby substantially reducing training overhead. These masks guide the activation of cross-attention during training to effectively separate multiple concepts, while a classspecific attention regularization technique is employed to mitigate overfitting. In addition, we introduce uniform sampling and re-weighted sampling strategies to further enhance generation quality. Experimental results demonstrate that ConceptCraft outperforms existing baselines in both image alignment and text alignment tasks, and exhibits superior robustness in handling complex scenarios. This work presents an efficient and scalable solution for multi-subject personalized generation from a single image.
With the development of the Internet of Things and intelligent sensing technologies, open-source Inertial Measurement Unit (IMU) datasets and open-source models trained on them have become core foundations for research and development in academia and industry. However, their copyright protection faces major challenges: open-source sharing convenience raises data and model theft risks, while methods like encryption and differential privacy clash with open-source principles, and traditional watermarking fails to fit open-source data reuse and model distribution needs. To address the copyright confirmation dilemma of open source IMU datasets and models, this study proposes IMUWatermark, the first frequency-domain backdoor watermarking framework for IMU systems, specifically designed for open source scenarios. It supports watermarking for both open source datasets and open source models, enabling fulllifecycle protection from data sharing to model distribution. Identifying challenges in IMU frequency domain watermarking, such as balancing stealth and effectiveness, adapting to dynamic scenarios, and resisting filter interference, IMUWatermark addresses them through mechanisms such as low-frequency embedding and dynamic parameter adjustment. Experiments on public datasets such as PAMAP2 and MHEALTH verify its effectiveness: Dataset watermarking achieves an attack success rate greater than 90 %, and model watermarking reaches an attack success rate greater than 85 %.
With the rapid advancement of generative forgery technologies, the detection of multi-modal deepfake audio-video content has become an urgent demand in cyber security and forensic analysis. However, detecting audio-video deepfakes remains challenging: forgery traces are often subtle, dispersed, and highly resolution-dependent; existing multimodal methods rely on simple concatenation or shallow interactions, leading to insufficient exploitation of cross-modal consistency. To address these issues, we propose a Cross-level Multi-modal Fusion (CLMF) framework that progressively integrates audio cues into visual representations through cross-level attention, adaptively enhancing complementary information while suppressing redundancy. In addition, we design an Adaptive Audio Feature Enhancement module (AAFE) to highlight subtle frequencydomain artifacts often masked by noise, and a Multi-scale Visual Feature Enhancement module (MVFE) to capture both local texture inconsistencies and global structural distortions. These components jointly achieve robust and consistent cross-modal alignment of forgery traces, leading to significant improvements in deepfake detection performance. On the FakeAVCeleb benchmark, AURORA achieves an accuracy (ACC) of 94.32% and an area under the curve (AUC) of 93.66%, demonstrating superior performance.
Aerospace simulation is a research area attracting increasing attention, as it contributes greatly to shortening experimental time, reducing risks during flight trials, and enhancing spacecraft performance.Recently, computing technologies have been widely adopted in analyzing, verifying, and optimizing spacecraft operational status. However, creating accurate and reliable aerospace simulation models still faces many challenges, such as logical oversimplification and external disruptions. To tackle these challenges, we propose to refine the parameters and assess the credibility of these simulation models. In particular, we first reveal the limited generalization capabilities of existing parameter refinement methods and propose a generalizable deep reinforcement learning approach for parameter optimization. This method exploits reinforcement learning's unique advantages to maximize the expected benefits, thereby enabling effective and accurate parameter refinement in aerospace simulation models. Second, to mitigate the high expenses incurred by multiple simulation experiments, we propose to create a high-precision agent model. We construct a Back-Propagation (BP) neural network with an improved discrete particle swarm algorithm, which enhances the efficiency of simulation experiments. Finally, in order to quantitatively and accurately assess the simulation model's reliability, we propose a novel evaluation method based on Hilbert-Huang algorithm. Our proposed methods have been evaluated in multiple key aerospace scenarios, such as primary and secondary rocket separation and grid rudder applications. The efficiency and effectiveness have been verified by these intensive experiments.
3D object detection is a critical problem in the field of computer vision, widely applied in autonomous driving, robotic navigation, and other domains. Although modern detectors have achieved success in singlesensor object detection, they remain vulnerable to complex environments due to the limitations of single-sensor modalities. We propose SE-D3FNet, a multi-modal fusion framework for 3D object detection that integrates Squeeze-and-Excitation (SE) channel attention and a dynamic 3D focal loss to significantly improve detection accuracy. We present an enhanced feature extraction network termed SE-ResBlock, which demonstrates superior capability in capturing global contextual information. The loss function is also optimized to more accurately capture targets with poor recognition rates. Experimental results on the KITTI benchmark demonstrate that our proposed 3D object detection algorithm achieves superior performance for the car category compared to existing methods.
The proliferation of spatial data applications (e.g., location-based services) has driven massive outsourcing of geospatial datasets to public clouds, yet exposing critical vulnerabilities to external hackers and internal adversaries. To address the dual challenges of securing fundamental geometric range queries while supporting dynamic updates, this paper proposes a dynamic searchable symmetric encryption scheme for geometric range search that complies with both forward privacy and backward privacy. Specifically, we build a multi-level balanced matching query index and constructthree layers of ciphertext to perform screeningand verification, enhancing query efficiencythrough three rounds of matching andscreening. Furthermore, we establish dual index directories deployed across two isolated servers, implementing cross-server query verification to guarantee single-dimensional range visibility per node. This architecture effectively conceals access patterns through spatial information segregation. Subsequently, we propose a ciphertext-update mechanism with embedded covert cryptographic operators, systematically detecting and neutralizing information leakage during update operations, thereby preserving both forward and backward privacy. Conclusively, comprehensive empirical evaluations demonstrate the scheme's operational efficiency and real-world applicability.
Workflow scheduling in cloud data centers faces challenges such as task dependencies, resource heterogeneity, and dynamic workloads. Heuristic and meta-heuristic algorithms lack environment awareness and fail to meet real-time scheduling requirements. Existing deep reinforcement learning algorithms exhibit specific environment-aware capabilities; however, their feature modeling capabilities are insufficient, making it challenging to effectively model the nonlinear relationships between tasks and resources. To address these issues, we design a dynamic priority-aware mechanism that computes task priority factors in real time to adjust scheduling strategies adaptively. In low-load scenarios, dependency release is accelerated based on the priority rule of the successor node quantity to improve resource utilization; in high-load scenarios, bottleneck subtasks are prioritized based on the priority rule of the weighted critical path to shorten overall task completion times. We propose a Dynamic Priority-Aware Transformer-based Advantage ActorCritic (DPA-TA2C) scheduling method. The proposed method models the task topology, resource states, and priority factors as a sequence of features via a Transformer encoder. It collaboratively optimizes scheduling decisions using the policy and value functions of the A2C network to achieve end-to-end scheduling. Experimental results show that DPA-TA2C improves resource utilization by 49.5 % on average in low-load scenarios and reduces makespan by 51.6 % in high-load scenarios, demonstrating the efficiency and robustness of the proposed method for scheduling in complex cloud environments.
Vehicular edge computing (VEC) leverages roadside units (RSUs) to provide low-latency and energy-efficient computation for vehicular networks. However, existing offloading schemes still face challenges under high vehicle mobility, dynamic network conditions, uneven RSUs loads, and different Quality of Service (QoS) requirements. To address this, we propose a mobility-aware partial task offloading (MAPRO) framework that partitions tasks for parallel processing across multiple RSUs while considering mobility, network dynamics, and QoS demands. A hybrid framework combining Proximal Policy Optimization (PPO) and numerical optimization efficiently solves offloading, resource allocation, and power control. Experiments on real trajectories show that MAPRO outperforms state-of-theart methods.
The Internet of Vehicles (IoV) has emerged as a key enabler of Intelligent Transportation Systems (ITS), IoV faces significant challenges in scalability and security. The growing volume of real-time vehicular data strains network resources, and traditional centralized architectures introduce single points of failure that compromise reliability and data integrity. This paper investigates the integration of IOTA-a Directed Acyclic Graph (DAG) based distributed ledger designed for feeless and scalable transactions-into the IoV system. In the proposed architecture, vehicles operate as lightweight nodes and roadside units function as full nodes, yielding a resource-efficient architecture for secure, tamper-resistant data interaction across vehicular networks. We formally model the latency of transaction propagation and consensus confirmation phases to gain insight into its performance characteristics under high concurrency. We conducted extensive simulations under diverse traffic conditions on a private IOTAenabled vehicular network. Experimental results demonstrate that the IOTA-based IoV system achieves low latency on the order of a few milliseconds and high throughput of hundreds of transactions per second even as vehicle density increases. The system exhibits strong scalability and operational stability, indicating considerable potential for deployment in real-world ITS environments.
Efficiently injecting knowledge into large language models (LLMs) continues to pose a significant challenge. Current approaches, such as supervised fine-tuning (SFT) and retrieval-augmented generation (RAG), bear their own limitations. SFT requires substantial computational resources, while RAG is heavily dependent on the availability of high-quality knowledge bases. This paper introduces an innovative method for efficient knowledge injection, providing a balanced trade-off between resource utilization and performance. It extracts knowledge from domain-specific LLMs by storing the context representations and their corresponding target tokens in a datastore and injects it into another LLM by neighbor interfering with the LLM's decoding process. Experimental results with four LLMs reveal that our proposed method yields average gains of 33.1 in ROUGE-L and 16.3 in BLEU-4 on manufacturing specification datasets, and 7.7 in ROUGE-L and 4.5 in BLEU-4 on medical consultation datasets. We believe this method offers a novel direction and insight for LLM knowledge injection.
Large-scale sparse linear programming (LP) underpins critical applications in logistics, manufacturing, and scientific computing. However, conventional CPU-only solvers often fail to meet real-time performance demands as problem sizes reach millions of variables and constraints. This paper presents GCP-LP, a GPU-CPU collaborative framework that accelerates the COIN-OR Linear Programming solver by offloading key computational bottlenecks-including sparse matrixvector multiplication (SpMV), pivot selection, and Cholesky factorization-to GPUs while retaining sequential control flow on the CPU. By combining asynchronous data transfer, GPUoptimized memory layouts, and adaptive load balancing, GCPLP achieves up to 10% module-level and 5% overall runtime reductions on NETLIB and Mittelmann benchmarks.
In modern urban traffic management, accurate traffic flow prediction contributes to travel decision optimization, signal control, emergency response and resource scheduling, which improves road efficiency, reduces congestion and promotes sustainable urban planning. However, traffic flow data are characterized by complex spatio-temporal dependence, multi-scale periodicity, and highly dynamic changes. Existing studies mostly focus on single spatio-temporal domain modeling and ignore frequency domain information fusion, which makes it difficult to comprehensively capture the potential laws of traffic flow. To this end, a multi-scale spatio-temporal fusion Transformer prediction model is proposed, which systematically integrates frequency-domain analysis with spatio-temporal dependent modeling. The model contains three parts: (1) spatial-temporal adaptive neighbor selection algorithm, which dynamically supplements topological information based on spatio-temporal correlation to enhance the efficiency of inter-subdivisional information transfer; (2) frequency-domain feature coupling module, which fuses the frequency-domain and spatial-domain features by fast Fourier transform to enhance the ability of temporal pattern sensing; and (3) spatio-temporal-frequency-domain dual-attention encoder, which combines the linear-attention mechanism to efficiently capture the longrange spatio-temporal dependencies. Experimental results on several real traffic datasets show that the model significantly outperforms existing methods in terms of mean absolute error, root mean square error and mean absolute percentage error, and demonstrates stronger robustness in complex spatio-temporal patterns and abnormal fluctuation scenarios.
Deep neural networks (DNNs) are increasingly deployed in real-time applications, yet their inference performance is often constrained by inefficient GPU utilization. While various inter-operator scheduling methods have been proposed to address this issue, many suffer from coarse-grained operator classifications, incomplete parallelism, unawareness of GPU resource contention, etc. To overcome these limitations, we propose HIOCS, a heuristic inter-operator co-scheduling framework that accelerates DNN inference on GPUs through fine-grained kernel-level scheduling. HIOCS employs a graph optimization technique at kernel granularity that reconstructs data dependencies based on memory access patterns, thereby uncovering latent parallelism. Building on this optimized graph, a heuristic co-scheduling algorithm is introduced to prioritize critical-path kernels and selectively co-locate non-critical kernels by jointly considering latency impact, resource contention, and kernel affinity to minimize makespan and maximize GPU utilization. Extensive experiments on popular DNN models across diverse domains demonstrate that HIOCS consistently achieves higher GPU utilization and better scalability. Compared to PyTorch default mode and the state-of-the-art method Opara, HIOCS achieves up to 21.19x and 1.13x speedup in inference latency, respectively, making it well-suited for latency-sensitive deployment scenarios.
Old photo restoration is a challenging task due to diverse and compound degradations, complex structural damage, and the lack of real paired training data. To address these issues, we present a novel diffusion-based framework that integrates contrastive learning and frequency-aware modeling to effectively restore old photographs. Specifically, we design a category-guided contrastive encoder to align the latent distributions of synthetic and real old photos, effectively bridging the domain gap. Based on this aligned latent space, a diffusion model is employed to iteratively generate high-quality restorations with strong semantic and structural consistency. Additionally, we introduce a Discrete Cosine Histogram Attention (DCH) block to jointly capture structured (e.g., holes, scratches, blotches) and unstructured (e.g., noise, blur, fading, low resolution) degradations in the frequency domain. Without relying on real paired data, our method generalizes well to real-world old photos. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art methods in both synthetic and real-world scenarios. The code is available at https://github.com/Marvel-LG/Code.
The Vehicular Edge Computing (VEC) paradigm significantly reduces task processing latency in Internet of Vehicles (IoV) and Intelligent Transportation Systems (ITS) by deploying computational resources at Roadside Units (RSUs). However, the high mobility of vehicles and dynamic task arrivals lead to uneven load distribution among RSUs, severely impacting system performance. Actually, load balancing as an important evaluation metric for VEC system greatly affects the performance of individual edge servers in terms of latency, energy consumption, and task completion rates. In view of this, we propose a Proximal Policy Optimization (PPO) based deep reinforcement learning (DRL) approach to determine the task offloading and migration decisions and incorporate the fairness into the constraint, aiming to achieve efficient load-balanced task offloading in VEC. Particularly, we introduce a metric named Load Balancing Metric (LBM) to optimize RSU resource allocation and employ dynamical task migration strategies to optimize the metric. Simulation results demonstrate that this approach significantly enhances load balancing performance, reduces average latency and energy consumption, and provides an efficient resource scheduling solution for VEC systems.
Vision Language Models (VLMs) can enhance Internet of Things (IoT) applications by efficiently extracting valuable information from excessively long videos captured by IoT cameras. Due to the large volume of video data and the high computation overhead of VLMs, a practical deployment strategy is to transmit the video to the cloud only on demand and also deploy the VLMs on the cloud for video analytics. Yet, the interaction experience between humans and VLMs is degraded by the high latency in such cloud-edge collaboration applications. The latency is caused by both the video transmission process and the heavy VLM inference process. We propose SwiftReTaKe, a two-round transmission framework coupled with a low-latency pre-pruning strategy to reduce both network and inference latency. By first sending keyframes for relevance estimation and then adaptively transmitting informative frames, SwiftReTaKe minimizes data transfer and LLM computation. Compared to the state-of-the-art (SOTA) long video processing method, SwiftReTaKe reduces the latency by 6 times with only 3.33% accuracy drop.
The rapid evolution of malware variants and imbalanced family distributions pose significant challenges to accurate classification and system security. This paper introduces an adaptive multimodal learning framework for robust malware family classification. The key innovation lies in a Modality-Aware Feature Selection (MAFS) mechanism that dynamically selects the most discriminative features from byte-level, structural, and semantic modalities. Coupled with an adaptive fusion strategy, our approach effectively addresses feature redundancy, modality heterogeneity, and missing data scenarios. Extensive experiments on 5,841 samples from 10 families demonstrate state-of-theart performance, with 99.87% accuracy, 99.82% macro-F1, and 0.0175 log loss, significantly outperforming existing methods in handling both feature loss and class imbalance.
Straggler nodes are a major bottleneck in large-scale distributed training, degrading efficiency and stability. However, current solutions, including In-Network Aggregation (INA), lack the adaptability to effectively manage these stragglers in dynamic environments. This paper proposes Straggler Dynamic Management (SDM), an adaptive method for large-scale distributed training that performs dynamic straggler management by coordinating the data and control planes to achieve accurate, time-based detection and efficient mitigation via a performanceaware redundancy strategy and semi-asynchronous aggregation. SDM manages stragglers through a coordinated architecture that decouples the data and control planes for efficient detection and response. It leverages the data plane to estimate each node's remaining completion time, ensuring accurate and low-overhead straggler identification. The control plane then mitigates their impact using two key strategies: a performance-aware redundancy scheme to reduce waiting delays, and a semi-asynchronous aggregation mechanism that dynamically adjusts synchronization to alleviate gradient staleness and improve model convergence. We implement and deploy SDM on a real-world hardware testbed and evaluate its performance under various straggler scenarios. Experimental results demonstrate that SDM significantly improves training efficiency and convergence stability in the presence of straggler nodes, particularly when multiple stragglers occur simultaneously, exhibiting greater robustness and adaptability than existing methods.
The security of vision-language pre-trained models (VLMs) has become an increasingly critical concern, particularly due to their vulnerability to adversarial attacks in open environments. Most existing methods focus on image perturbations, ignoring image-text structural relationships and the need for dynamic, semantics-aware attack strategies. To address these limitations, we propose a dynamically adaptive framework for generating multi-behavior adversarial patches. A topological neighborhood graph models cross-modal semantic structures, while a lightweight text classifier detects sensitive or jailbreak instructions to switch attack strategies. A two-stage optimization-semantic stripping and target binding-precisely controls perturbations for target-specific outputs. Experimental results demonstrate that the proposed method achieves high attack success rates across a variety of tasks, including image-text retrieval, image classification, and multimodal question answering. Moreover, the approach exhibits strong robustness and adaptability, exposing critical security vulnerabilities in current VLM systems and offering valuable insights for future research on the defense of multimodal models.
