Rain significantly degrades the performance of computer vision systems, particularly in applications like autonomous driving and video surveillance. While existing deraining methods have made considerable progress, they often struggle with fidelity of semantic and spatial details. To address these limitations, we propose the Multi-Prior Hierarchical Mamba (MPHM) network for image deraining. This novel architecture synergistically integrates macro-semantic textual priors (CLIP) for task-level semantic guidance and micro-structural visual priors (DINOv2) for scene-aware structural information. To alleviate potential conflicts between heterogeneous priors, we devise a progressive Priors Fusion Injection (PFI) that strategically injects complementary cues at different decoder levels. Meanwhile, we equip the backbone network with an elaborate Hierarchical Mamba Module (HMM) to facilitate robust feature representation, featuring a Fourier-enhanced dual-path design that concurrently addresses global context modeling and local detail recovery. Comprehensive experiments demonstrate MPHM's state-of-the-art performance, achieving a 0.57 dB PSNR gain on the Rain200H dataset while delivering superior generalization on real-world rainy scenarios.

In real-world scenarios, multimodal information continuously evolves, with new entity and relation types emerging, necessitating timely updates to multimodal knowledge graphs for supporting downstream tasks.However, existing methods struggle to balance real-time adaptability and computational efficiency in continual learning scenarios.To this end, this paper proposes the Continual Multimodal Entity and Relation Joint Extraction (CMERJE) task and a Multimodal Prompt-based Boundaryenhanced Continual (MPBoCo) framework.Specifically, MPBoCo incrementally stores task-specific knowledge via learnable multimodal prompts, dynamically matches relevant prompts for each instance, and fuses them into a frozen backbone model for task-specific reasoning.Subsequently, the boundary-enhanced dual-branch module leverages the auxiliary branch to preserve local syntactic continuity and provide boundary guidance.Experimental results demonstrate that MPBoCo achieves superior performance in real-world scenarios, significantly outperforming baseline methods by 5.5% and 7.2% in 10-task and 5-task settings, respectively.

Accurately identifying IoT devices connected to a network is crucial for improving network management and ensuring network security. As IoT devices increasingly move across institutional boundaries, identifying devices that migrate among institutions has become a significant challenge. However, existing studies assume that IoT devices remain stationary. As a result, they are inadequate for addressing the security risks introduced by device mobility across institutions. Therefore, we propose an IoT device identification framework named DFGLIoT. The framework adopts a decentralized fully connected architecture to enable knowledge sharing among institutions. This design avoids single points of failure while effectively supporting the identification of cross-institutional mobile IoT devices. We model the communication traffic between IoT devices and their gateways as communication interaction graphs. These graphs provide a comprehensive view of the interaction process between the communicating parties. Based on this representation, we design a graph classifier that integrates a dual-scale dependency modeling module with a spatial feature extraction module. The classifier captures interaction patterns in the communication traffic and constructs behavior fingerprints of IoT devices, enabling accurate device identification. Experimental results on three public datasets demonstrate the effectiveness of DFGLIoT in identifying IoT devices that move across institutions. The source code can be accessed at https://github.com/traveler-wang/DFGLIoT.

Multimodal named entity recognition (MNER) on social media is a challenging task that uses corresponding images to aid in recognizing entities in short and noisy texts. Most existing methods suffer from two major drawbacks: weak image-text semantic correlation and the hybridity of entity types in intermodality alignment. To this end, a general framework ProMNER is proposed for unimodal and multimodal NER tasks, which includes text-driven visual-modality enhancement (TVE) and prompt-guided type-related refinement and interaction (PTRI). Specifically, to alleviate the bias caused by irrelevant visual content, TVE employs cross-modal translation to generate visual representations that replace the original ones. Furthermore, the generated can serve directly as an additional input modality for NER, enabling the framework to be generalized to unimodal and multimodal NER tasks. Subsequently, to tackle the hybridity problem of intermodality entity alignment, type-filled question prompts are designed as guidance information to explicitly refine type-related visual representations. PTRI obtains type-specific multimodal representations by capturing semantic intermodality interactions for each type, and learns a comprehensive multimodal representation through dynamic weighting. Experimental results on five benchmark datasets demonstrate the superiority and generality of ProMNER in unimodal and multimodal NER tasks.

Federated learning is an emerging machine learning paradigm that effectively alleviates the data silo problem by distributing the model training process to multiple data holders. However, data from real-world mobile applications often has multi-label and presents a long-tailed distribution, where labels are generally non-independent and non-identically distributed, thereby increasing the challenges caused by data heterogeneity. To address the above problems, we propose a Federated Chain Context Optimization (FedCCO) for long-tailed multi-label image classification. Inspired by the success of Chain of Though (CoT) in enhancing the semantic expressive ability of models, this method fine-tunes the CLIP model using semantic descriptive vectors generated by the Chain Context Optimization (ChCoOp) to establish semantic correlations between head and tail classes across clients, which improves the ability of the model to recognize tail classes. The experimental results show that the FedCCO achieves satisfactory performance in long-tailed multi-label image classification in federated learning on VOC-LT and COCO-LT datasets.

Video question answering (VideoQA), a critical task in vision-language understanding and reasoning, encounters significant challenges in integrating visual concepts for compositional spatio-temporal reasoning, due to the inherent complexity of video data. Existing methods struggle to effectively capture both static and dynamic visual concepts and establish practical symbolic reasoning, which are essential for answering compositional questions. To address these challenges, we propose a neural-symbolic framework called Progressive Symbolic Reasoning based on Static-Dynamic Scene Parser Network (PSR-SDSP) for VideoQA tasks. PSR-SDSP features two key innovations: (1) the Scene Parser Network, which converts unstructured video data into static-dynamic symbolic representations, extracting structured representations of persons, objects, relationships, and action chronologies. (2) the Symbolic Reasoning Machine (SRM), which conducts top-down question decomposition and bottom-up compositional reasoning. Specifically, the SRM applies a program execution algorithm that integrates static-dynamic visual concepts with predicted programs to deduce the answers. Our results on AGQA Decomp show that PSR-SDSP outperforms existing models in compositional spatio-temporal reasoning. Experiments on the Novel Compositions and More Compositional Steps tasks highlight its potential generalization. Evaluations on datasets with varying degrees of domain differences, including STAR, ActivityNet-QA, and TVQA+, confirm PSR-SDSP's scalability.

By analyzing command-line arguments during process execution, Endpoint Detection and Response (EDR) and Security Information and Event Management (SIEM) tools can detect potential malicious commands and identify Advanced Persistent Threats (APT). In practice, most implementations rely on heuristics and regular expression matching. However, attackers can bypass detection by obfuscating commands using Unicode modifier letters. This paper investigates the mechanism behind this evasion technique. Through reverse engineering, we identify the root cause and discover an internationalization API vulnerability in Windows. The impact assessment reveals that 424 system programs are potentially at risk. We further examine command arguments containing internationalized domain names with modifier letters. Code analysis traces how these domains resolve across operating systems, exposing additional attack surfaces. Based on these findings, we formally define modifier letter evasion attacks. We then propose a threat model, which allows attackers to construct evasion commands that remain executable while bypassing command-line detection. Furthermore, we design evasion cases and validate them through proof-of-concept experiments. Results show that high-risk cases can bypass certain leading EDR and SIEM tools. To mitigate this attack, we propose a detection approach, develop a corresponding tool, and suggest defense measures.

Cross-device federated learning (FL) enables collaborative model training across heterogeneous edge devices while preserving data privacy. However, system heterogeneity remains a major challenge, especially under constrained computation and memory resources. Although model compression—particularly knowledge distillation—has been widely used to reduce overhead, it inevitably introduces model heterogeneity, leading to degraded performance. To address this overlooked issue, we propose AHFL, a resource-adaptive and data-heterogeneity-aware federated learning framework. AHFL employs three coordinated strategies: a data-driven client grouping mechanism to assess and exploit heterogeneity levels, adaptive model compression tailored to each group’s resource profile, and a novel group distribution representation module with theoretical convergence guarantees to mitigate performance degradation caused by model heterogeneity. Extensive experiments demonstrate that AHFL reduces computational cost by 1.7× while simultaneously improving global accuracy by 4.13%. In experiments using the same compressed architecture, AHFL narrows the global-local accuracy gap to under 2%, achieving accuracy improvements of +7.47% (local) and +3.29% (global). The code is available at: https://github.com/CST-FederatedLearning/AHFL.

Federatedlearning has garnered significant attention in the Internet of Things and healthcare applications due to its ability to train a shared global model across distributed clients. However, imbalanced data distribution leads to model discrepancies among clients. Most existing methods adopt implicit alignment strategies while overlooking explicit modeling of geometric and directional discrepancies in feature representations, which undermines local model optimization. To address this issue, we propose a method of representation deviation alignment in federated learning, which projects features onto the principal feature space to measure deviations between local and global feature representations explicitly. Specifically, Federated learning with Representation Deviation Alignment (FedRDA) employs a feature encoder to extract compact features and construct unbiased principal feature spaces for global and local models. Then, the residual projection in the feature space serves as a quantitative measure of the representation deviation, effectively capturing the latent direction differences between models. Besides, we introduce a representation consistency alignment strategy, which ensures that the distribution of local client features becomes more uniform within the global feature space. Extensive experiments on SVHN, CIFAR-10, CIFAR-100, Tiny-ImageNet, and GC10 demonstrate that FedRDA effectively reduces the classifier bias caused by representational differences.

Federated semi-supervised learning ( FSSL ) involves training a model in a federated environment using a few labeled samples and many unlabeled samples . Compared with semi-supervised learning, FSSL faces more complex data situations, especially when data are non-independently and identically distributed ( non-IID ), adding more challenges to the learning process. The previous method addresses the aforementioned issues by enlarging the training sample space through multiple client random sampling and reweighting the parameters. Although it achieves high accuracy, it sacrifices communication efficiency. In this study, we propose PDCFed, a two-stage sampling method that uses the P redicted D istribution C hanges of samples after different data augmentations . We evaluate the credibility of the samples based on the maximum probability predicted by weak augmentation. When the samples are in a less reliable space, they are further sampled after adjusting the predicted distribution changes using a Gaussian function . To enhance the model’s generalization ability , an entropy penalty term is incorporated after unsupervised training loss. Extensive experiments demonstrate that this method outperforms existing methods on three datasets with non-IID data and significantly improves communication efficiency.

The traditional rule-based congestion control algorithms cannot set congestion window size flexibly, resulting in the inadaptation of the dynamic networks. This article presents a method to model end-to-end TCP congestion control problem using classification techniques. The network status parameters as the input and the type of network status as the output are defined through the analysis of some existing congestion control algorithms. NewReno, CUBIC, and Compound are used as feedback to produce the training data of the XGBoost classifier. The experimental results show that the classifier effectively shapes the strategies of three outstanding congestion control algorithms and almost achieves the same throughput, delay, and fairness. The proposed method makes the congestion control algorithm able to learn from data produced by network.

Similarity graph learning is the most key technique for multiview spectral clustering. However, existing methods fail when applied to uncertain data contaminated with various types of noise in an open environment. Due to the damaged structure by noise, unreliable similar relationships are learned, which extends similarity inconsistency among views. Moreover, the high-order correlation hidden in graphs are ignored generally. To address these problems, we propose a reliable similarity learning scheme for multiview clustering on uncertain data. This method can significantly improve spectral clustering performance in a noisy environment, and the contributions of our scheme include the following three aspects: 1) Uncertain data subspace reconstruction and adaptive graph learning are combined to construct a view-specific graph from high-quality recovered data, thus improving robustness. 2) A low-rank tensor constraint is utilized to facilitate multiview fusion, where the latent high-order correlation among view graphs will be fully explored when learning the consensus graph structure. 3) Data recovery, view-specific graphs, and latent consensus tensor structure are assembled into a unified framework, to be optimized jointly for mutual benefit. Our study also develops an efficient algorithm for obtaining overall solutions. The experimental results on several datasets demonstrate that our proposed approach shows significant improvements in robustness and evaluation metrics over the comparison methods.

Classifying encrypted traffic is critical to effective network analysis and management. While traditional payload- based methods are powerless to deal with encrypted traffic, machine learning methods have been proposed to address this issue. However, these methods often bring heavy overhead into the system. In this paper, we propose a hybrid method that combines signature-based methods and statistical analysis methods to address this issue. We first identify SSL/TLS traffic with signature matching methods, and then apply statistical analysis to determine concrete application protocols. Our experimental results show that the proposed method is able to recognize over 99% of SSL/TLS traffic and achieve 94.52% in F-score for protocols identification.