Key node identification in attributed heterogeneous networks aims to find a set of core nodes that can maximize the scope of information propagation within complex networks comprising multiple node types and rich attribute information. However, existing methods typically employ early feature fusion or symmetric contrastive learning when integrating topological structure and node attributes, where semantic noise can readily propagate through gradient backpropagation and induce negative transfer contamination. Meanwhile, node selection mechanisms based on local structural similarity tend to cause seed nodes to over-concentrate within a few communities, resulting in influence overlap that limits the propagation reach. To address the above challenges, this paper proposes a heterogeneous network with cross-community alignment and late fusion (HCF-Net). The proposed method adopts non-contrastive self-supervised learning with gradient stopping and a late score fusion strategy, enabling unidirectional alignment from the semantic branch to the topological branch, where the two branches independently generate node importance scores and then perform weighted fusion, thereby suppressing the interference of semantic noise on topological representation learning. It constructs a cross-community semantic neighbor graph to establish long-range semantic associations among nodes from different communities, and introduces a neighborhood discount re-ranking strategy to enhance the spatial dispersion and propagation coverage of the seed node set. Experimental results demonstrate that HCF-Net outperforms comparison methods in terms of ranking accuracy, spatial dispersion, and propagation effectiveness.
Obfuscation techniques are widely employed in malware to conceal program logic by transforming code structure, control flow, or symbolic information while preserving original functionality, thereby evading static analysis and automated detection. Existing image-based approaches for Android malware classification typically convert bytecode directly into grayscale images, neglecting the heterogeneous effects of obfuscation across different code abstraction levels. Consequently, their performance degrades significantly when applied to obfuscated samples. A systematic analysis of obfuscation mechanisms reveals that opcode sequences exhibit high invariance under symbol-level obfuscation, and their transition patterns remain largely preserved even under structural-level obfuscation—properties that can be leveraged to construct robust image representations. To address this challenge, MRODroid (Markov RGB-based obfuscation-resilient Droid), a novel method for obfuscation-resilient Android malware classification, is proposed. MRODroid models opcodes at three granularities, i.e. class-level, method-level, and instruction-level, as Markov transition probability matrices, which are encoded into the red, green, and blue channels of a single RGB image. This formulation integrates multi-scale contextual dependencies into a unified color representation, effectively preserving semantic features resilient to obfuscation. A lightweight convolutional neural network is further designed to accommodate the sparse yet hierarchical nature of such image representations. The architecture incorporates channel attention to enhance discriminative features and residual connections to facilitate effective deep feature propagation with minimal computational overhead. Experimental results demonstrate that MRODroid consistently outperforms state-of-the-art image-based methods under multiple obfuscation scenarios, confirming its efficacy for robust Android malware classification.
While large language models (LLMs) can process extensive input contexts, they face critical challenges in long-form generation: uncontrollable output length, and a tendency for content to deviate from initial input instructions as the text extends. To address these limitations, this paper introduces TreeWriter, a novel framework designed to generate content of arbitrary length while ensuring consistent alignment with the initial writing instruction. Specifically, the TreeWriter framework transforms a complex writing task into a structured “writing tree” via recursive task decomposition. In this structure, each leaf node represents a specific writing subtask, thereby breaking down a macroscopic writing objective into a series of manageable, executable steps. Within the generation process at each leaf node, the framework employs a dual-model approach. First, a reasoning model constructs a detailed writing outline based on the current node’s objective and the preceding text. Then, a writing model generates the corresponding text segment conditioned on this outline. The final document is then composed by integrating the text segments from all leaf nodes in the sequence dictated by the tree??s topology. To streamline the framework, this paper also introduces the “text continuation dataset” to train a single LLM to perform the dual roles of both reasoning and writing, thus unifying the core functions. On the proposed LoFT benchmark, the method not only boosts the overall score of large-scale models like DeepSeek-V3 from 79.53 to 81.18, but more importantly, provides critical stability for smaller-scale models like Qwen2.5-14B, lifting its overall score from 10.79 to 49.85. These experimental results underscore the robustness and reliability of the proposed method for long-form text generation.
Research focuses on the dual-robot evacuation problem in an equilateral triangular area. On the basis of the original communication models (wireless and face-to-face), the dual-robot evacuation problem in an equilateral triangular area with the S/R (sender-receiver) communication model is proposed. In the S/R communication model, one robot is called the Sender, which can only send messages wirelessly, while the other robot is called the Receiver, which can only receive messages wirelessly. Additionally, the two robots can also communicate face-to-face, and the communication state of each robot is known to both itself and the counterpart. It is assumed that the two robots with the S/R communication have a maximum speed of 1, with their initial positions set at the centroid of the equilateral triangle and the exit located on the boundary of the triangle. The task of the robots is to search for the exit collaboratively and complete the evacuation process. The evacuation cost of the algorithm is defined as the time when the last robot reaches the exit, which is adopted to evaluate the algorithm efficiency. By analyzing the algorithm design ideas and robot trajectory characteristics under the wireless and face-to-face communication models, combined with the S/R communication model and the geometric properties of the equilateral triangle, the Sender-Receiver asymmetric cooperative evacuation algorithm is proposed, with its efficiency analyzed and the lower bound of the evacuation cost for this problem proved. The relationship between the overhead of S/R communication algorithms and that of face-to-face communication and wireless communication algorithms is analyzed. A prominent feature of the algorithm lies in the inherent asymmetry in the trajectories of the two robots. Furthermore, the correctness of the theoretical analysis is verified through simulation experiments.
Image dehazing is an important research direction in image restoration for computer vision. Publicly available datasets are mainly dominated by large-scale synthetic hazy image datasets, while real-world hazy image datasets remain relatively limited. In recent years, supervised image dehazing methods have achieved promising performance on synthetic hazy image benchmarks. However, these methods generally exhibit limited restoration performance on real-world hazy images. Therefore, constrained by the availability of suitable datasets, researchers have increasingly focused on unsupervised image dehazing methods to improve the generalization ability of models in real-world scenarios. This review focuses on single-image dehazing based on unsupervised learning and systematically surveys, classifies, and summarizes representative studies from different sources published during 2019 to 2025. The reviewed methods are comparatively analyzed in terms of their core techniques, network architectures, supervision mechanisms, advantages, limitations, and applicable scenarios. According to their primary learning mechanisms, existing unsupervised image dehazing methods are categorized into four groups: physical-prior-based methods, generative adversarial network-based methods, fusion-based methods, and other emerging methods. Their technical evolution and major development trends are also summarized. Furthermore, based on reported experimental results, the quantitative performance of representative models on commonly used datasets, including SOTS-Indoor, SOTS-Outdoor, HSTS, NH-HAZE, I-HAZE, O-HAZE, and RTTS, is systematically compiled and compared. The relationship between dehazing performance and computational complexity is further analyzed in terms of model parameters, computational cost, and inference time. Based on these analyses, this review summarizes the major challenges faced by unsupervised image dehazing, including real-world generalization, training stability, model complexity, and evaluation criteria. Finally, future research directions are discussed from the perspectives of test-time adaptation, adaptive physical priors, lightweight model design, and diffusion models, providing a reference for further research and practical applications of unsupervised single-image dehazing.
