National Computer Network Emergency Response Technical Team/Coordination Center of Chinar

Pre-trained Vision-Language Models (VLMs), e.g. CLIP, have become essential tools in multimodal transfer learning. However, fine-tuning VLMs in few-shot scenarios poses significant challenges in balancing task-specific adaptation and generalization in the obtained model. Meanwhile, current researches have predominantly focused on prompt-based adaptation methods, leaving adapter-based approaches underexplored and revealing notable performance gaps. To address these challenges, we introduce a novel Reconstruction-based Multimodal Adapter (RMAdapter), which leverages a dual-branch architecture. Unlike conventional single-branch adapters, RMAdapter consists of: (1) an adaptation branch that injects task-specific knowledge through parameter-efficient fine-tuning, and (2) a reconstruction branch that preserves general knowledge by reconstructing latent space features back into the original feature space. This design facilitates a dynamic balance between general and task-specific knowledge. Importantly, although RMAdapter introduces an additional reconstruction branch, it is carefully optimized to remain lightweight. By computing reconstruction loss locally at each layer and sharing projection modules, the overall computational overhead is kept minimal. A consistency constraint is also incorporated to better regulate the trade-off between discriminability and generalization. We comprehensively evaluate the effectiveness of RMAdapter on three representative tasks: generalization to new categories, generalization to new target datasets, and domain generalization. Without relying on data augmentation or duplicate prompt designs, our RMAdapter consistently outperforms state-of-the-art approaches across all evaluation metrics.

ABSTRACT In recent years, artificial intelligence technologies have been widely used, bringing security, convenience and certain risks. Deepfake techniques raise significant security concerns by manipulating facial images to create convincing but false representations. We introduce a novel mask‐supervision‐based deepfake detection method to improve detection performance in this context. Our approach corrects the model's focus on irrelevant regions through mask supervision, using pixel‐level labels to guide the model towards synthetic facial regions and ensure more accurate extraction of spatial features. In addition, we incorporate a frequency‐domain feature extraction module that exploits the robustness of frequency‐domain cues to compression artefacts. We first preprocess the input image and then feed it into the mask supervision and frequency‐domain feature extraction modules. The mask supervision module extracts intermediate features using the High‐Resolution Network (HRNet) and refines spatial features by guiding the prediction mask with the ground‐truth mask. The frequency‐domain module extracts features via the Discrete Cosine Transform (DCT) and filtering across different frequency bands. Finally, spatial and frequency‐domain features are concatenated and fed into a classification network to output the final prediction. Experimental results show that our method maintains good robustness in compressed scenarios.

Abstract With the continuous growth of threat intelligence data and the increasing complexity of attack chains, attributing Advanced Persistent Threat (APT) attacks to specific organizations and validating attack behaviors have become challenging due to the limited interpretability and credibility of prediction results. Although graph-based APT organization prediction methods are capable of modeling the relationships among attack entities, their outputs still require further verification and analysis in the context of real-world attack scenarios. To address this issue, this paper builds upon the AARGS model for APT organization prediction and introduces a large language model to perform semantic reasoning and high-level relationship completion on the predicted results. By automatically parsing and reasoning over attack behaviors described in real threat intelligence texts, a comprehensive semantic representation of the attack chain is constructed and systematically compared with the model’s prediction outcomes. Furthermore, three-dimensional reconstruction of attack chains, together with temporal evolution and geographic propagation visualizations, is employed to intuitively present and analyze the dynamic evolution of attack paths. Experimental results demonstrate that the proposed approach effectively enhances the interpretability and reliability of APT attribution results, providing strong support for APT traceability analysis and informed defense decision-making.

Intent recognition, as a core technology of human-computer interaction systems, directly affects the quality of intelligent services. Traditional single-modal intent recognition is limited by the constraints of a single information source, resulting in low accuracy in complex scenarios. This paper proposes a Cross-Modal Attention Fusion (CMAF) model for multimodal intent recognition, which achieves accurate intent recognition through deep fusion of text, audio, and visual multimodal information. The model adopts a hierarchical feature extraction architecture and combines intra-modal attention with cross-modal attention mechanisms to effectively capture semantic correlations and complementary information between modalities. Experimental results on three public datasets, namely MIntRec, SIMMC, and the newly added medical consultation dataset (MedDialog), show that the CMAF model improves the accuracy by 4.2%, 3.8%, and 5.1% respectively, and the macro-F1 score by 3.9%, 3.5%, and 4.7% respectively compared with baseline models. Particularly, the performance degradation of CMAF is only 5.2% under 0dB noise, verifying the significant advantages of multimodal fusion in intent recognition tasks.

Computer vulnerability is the core risk sources threatening the security of information systems. They widely exist in multiple layers such as hardware, firmware, software, and the misconfiguration of system. Their exploitation may lead to severe consequences including data leakage, system out-of-control, and network paralysis. Based on multiple vulnerability research literatures, this work systematically sorts out the definitions, typical types, and harmful cases of vulnerabilities from three dimensions: hardware vulnerabilities, software vulnerabilities, software vulnerabilities, and the misconfiguration of system. Then we proposes targeted defense strategies, providing comprehensive theoretical references and practical guidance for vulnerability management and security protection.

In this correspondence, we propose a novel Ordered Directed Acyclic Graph (O-DAG) blockchain to improve model parameter selection for Blockchained Federated Learning (BFL) in Intelligent Transportation Systems (ITS). Our ODAG blockchain composes an innovative Ordered Transaction Topology (OTT) and a Deep reinforcement learning-enhanced Consensus Mechanism (DCM). The OTT ensures the traceability of BFL model parameter selection by using blockchain to sequentially record local and global model parameters. By establishing traceable iteration relationships among these parameters, the OTT facilitates the efficient selection of updated model parameters for BFL. Building upon the OTT, our DCM accelerates BFL convergence by enabling adaptive and fast model parameter selection. The DCM optimizes parameter selection policies for both local training and global aggregation by getting evaluations from heterogeneous ITS clients. Simulations show that our ODAG blockchain outperforms state-of-the-art baselines, improving BFL convergence speed by 38.64% and model accuracy by 7.47%, while maintaining strong scalability in ITS networks

The integration of Large Language Models (LLMs) with external tools enables intelligent automation far beyond text generation, yet coordinating multiple LLM-based agents remains difficult due to interaction overhead, resource waste, and fragile information flow. This paper introduces OmniNova, a modular and hierarchical multi-agent framework that unifies language models with capabilities for web search, browser automation, and code execution. OmniNova advances the state of the art through a hierarchical architecture that separates coordination, planning, supervision, and specialization; a dynamic routing mechanism that activates agents according to task complexity and state; and a multi-layered LLM integration strategy that allocates high-capability reasoning models only where they are cognitively necessary while assigning routine work to lighter models. Across 50 complex tasks in research, data analysis, and web interaction, OmniNova improves task completion (87% versus a 62% baseline), reduces token usage by 41%, and delivers higher human-rated quality (4.2/5 versus 3.1/5). The contribution includes both a principled system design and an open-source implementation intended to support research and practical deployment. Code is available at https://github.com/Superagentsys/OmniNoval.git.