Fake news video detection aims to analyze and verify the authenticity of video-based news content using multimodal data, including visual, audio, and textual cues. Fake news videos typically contain subtle, misleading alterations, often limited to specific frames, or feature genuine video coupled with fabricated narratives, complicating authenticity assessment through video content alone. Existing approaches predominantly depend on intrinsic features within the video or utilize external knowledge derived from a single modality, limiting their capability to exploit multi-source cross-modal information effectively and rendering them vulnerable to substantial content noise. To address these issues, we propose an Enhanced Method for Fake Video Detection based on Cross-modal Retrieval utilizing Large Models (FVDLM). We first collect relevant video and text news to augment external knowledge. Given the sparsity of informative features in video content, the information bottleneck theory is employed to denoise irrelevant information. Furthermore, to effectively integrate cross-modal knowledge and enrich external context, we introduce a prompt learning approach utilizing large models to generate contextual knowledge. Three specialized prompts are crafted to assess video authenticity from multiple viewpoints. Comprehensive experiments validate the effectiveness and superiority of our proposed model.

Large Language Models (LLMs) demonstrate impressive capabilities but face a critical challenge: generating toxic content due to pre-training biases. Existing detoxification methods, such as fine-tuning and prompting, incur high computational costs or suffer from instability. Recently, Representation Engineering (RepE) emerged as a promising internal steering approach; however, current methods treat toxicity as a monolithic, single-dimensional concept. Steering along a single direction inadvertently suppresses harmless concepts, degrading general model capabilities. We hypothesize that the toxicity concept in LLMs comprises multiple linearly separable sub-concepts in the activation space. Through geometric visualization and quantitative analysis, we empirically validate this, demonstrating that distinct sub-concepts (e.g., Insult, Threat) reside in largely orthogonal subspaces. Building on this insight, we propose MARS (Multi-dimensional Adaptive Representation Steering), a framework that decouples toxicity into sub-concepts and applies iterative steering with adaptive magnitude during inference. This allows precise detoxification without compromising semantic quality. Comprehensive experiments on Llama3.1-8B, Mistral-7B, and Gemma-7B show MARS significantly outperforms baselines, reducing toxicity scores by up to 79

Large language model-based multi-agent systems (LLM-MAS) effectively accomplish complex and dynamic tasks through inter-agent communication, but this reliance introduces substantial safety vulnerabilities. Existing attack methods targeting LLM-MAS either compromise agent internals or rely on direct and overt persuasion, which limit their effectiveness, adaptability, and stealthiness. In this paper, we propose MAST, a Multi-round Adaptive Stealthy Tampering framework designed to exploit communication vulnerabilities within the system. MAST integrates Monte Carlo Tree Search with Direct Preference Optimization to train an attack policy model that adaptively generates effective multi-round tampering strategies. Furthermore, to preserve stealthiness, we impose dual semantic and embedding similarity constraints during the tampering process. Comprehensive experiments across diverse tasks, communication architectures, and LLMs demonstrate that MAST consistently achieves high attack success rates while significantly enhancing stealthiness compared to baselines. These findings highlight the effectiveness, stealthiness, and adaptability of MAST, underscoring the need for robust communication safeguards in LLM-MAS.

Detecting malicious encrypted Web traffic through communication behavior modeling has become a mainstream research paradigm. A burst is defined as a sequence of packets transmitted consecutively in the same direction. Under encryption, bursts serve as the smallest behavioral units that still preserve interaction semantics. However, existing burst-based modeling methods remain limited. For behavior entity modeling, current approaches often ignore fine-grained features within bursts, such as temporal rhythms and sending frequencies. As a result, the intrinsic temporal characteristics of bursts cannot be effectively captured. For interaction modeling, existing methods assume that dependencies between bursts are only sequential and are conveyed through the start and end packets. This assumption limits the representation of multi-dimensional cross-burst dependencies. To address these problems, we propose a Web malicious encrypted traffic detection framework based on an EXtended Traffic Interaction Graph (XTIG). First, temporal statistical features including burst duration, sending frequency, and inter-packet intervals are integrated to enhance the discriminability of behavioral entities. Second, a cross-burst fully connected structure is introduced to strengthen packet-level associations between adjacent bursts. In addition, a lightweight anomaly filtering mechanism is adopted to reduce computational overhead. GraphSAGE is employed to perform multi-class encrypted Web attack detection, as it aligns well with the structural characteristics of XTIG. Experimental results show that the proposed framework outperforms existing baselines, achieving an average multi-class accuracy improvement of approximately 3.75%. Moreover, the framework demonstrates efficient detection performance in real high-throughput network environments.

Logs serve as an effective data source for recording and judging system states and abnormal events in complex systems. Current deep learning-based methods have proven effective in detecting anomalies in these system logs. However, existing anomaly detection methods, which predominantly rely on template-based and global window-based approaches, still face challenges in terms of flexibility and practicality. Template-based methods, while widely adopted for their simplicity and efficiency, overlook parameter information and fail to capture the true execution semantics. And global window-based methods, despite their effectiveness in modeling global dependencies, cannot simultaneously capture both global and local dependencies, leading to the obscuration of important local features. To address these issues, we propose a Log anomaly detection method based on MUlti-scale SEmantic representation, LogMUSE. Specifically, LogMUSE obtains template and parameter information through log parsing, employs a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model for template semantic embedding, and enhances log entry representations via cross-attention mechanisms to effectively capture different parameter features under the same template. Additionally, we design the multi-scale Transformer model to capture global and local anomaly patterns, which enable fixed-length log sequences to focus on features at different scales. Extensive experiments on real-world benchmark datasets, including BGL, Thunderbird and Spirit, show that LogMUSE outperforms existing methods in log anomaly detection, achieving F1-scores of 98.62%, 94.32%, and 99.20% respectively. These results surpass the performance of current state-of-the-art methods and demonstrate the strong generalization across different system scenarios.

The increasing traffic pressure and growing demand for public safety have spurred research into intelligent X-ray prohibited item detection. However, these studies concentrate on horizontal bounding boxes, failing to effectively suppress interference from excessive background or neighboring items when dealing with prohibited items in various orientations. In addition to the lack of X-ray benchmarks for oriented prohibited item detection, traditional oriented detection methods also experience a significant performance drop when applied to X-ray scenarios. Existing oriented detection methods primarily address the planar rotation on rigid surfaces, while X-rays passing through passenger luggage to visualize internal structures exhibit both planar and stereoscopic rotation phenomena. The challenging stereoscopic rotation will cause increasing intra-class diversity and severe appearance deformation, thereby aggravating the difficulty of prohibited item detection. In this paper, we contribute the first oriented X-ray prohibited detection benchmark ORXray, featuring large-scale prohibited items, high-quality images, informative annotations, and real-world scenarios. Moreover, we first delve into the differences in appearance deformation between planar rotation and stereoscopic rotation, and propose the stereoscopic-aware oriented feature calibration (SWEAR) framework, integrating the novel stereoscopic perception guidance (SPG) module and progressive deformation calibration (PDC) module. To address the increasing intra-class diversity, SPG first augments stereoscopic rotation semantics and empowers the network with stereoscopic perception ability. Based on the perceived stereoscopic semantics, PDC further progressively calibrates the distorted features to restore the visual characteristics that are vanished by stereoscopic rotation. Comprehensive experiments demonstrate that SWEAR outperforms other state-of-the-art methods by a margin of at least 2.7 https://github.com/wytbwytb/SWEAR .

Hallucination has emerged as a pivotal challenge of Large Language Models (LLMs) that generate plausible yet non‑factual content, significantly impeding the trustworthy AI applications in real-world scenarios like medical diagnosis and autonomous driving. Editing the internal activations of LLMs during inference has shown promising effectiveness in mitigating hallucinations with minimal cost. However, previous editing approaches neglect the query‑specific inference pathways that require tailored truthful steering vectors, resulting in suboptimal hallucination mitigation. To address these issues, we propose the Query-Routed Activation Editing (QRAE) framework, which comprises Divergence-sensitive Head Routing (DHR) and Truth-hierarchical Preference Steering (TPS), to fully leverage query-specific semantics for adaptive activation editing. Specifically, DHR is proposed to establish a query-aware head selection criterion, thereby dynamically routing to truth-critical attention heads. Subsequently, TPS introduces a query-specific steering vector calibration policy with the guidance of progressive truth-preferred optimization, enabling precise and adaptive editing for each distinct query. Extensive experiments on the widely recognized TruthfulQA benchmark demonstrate that QRAE outperforms SOTA methods by up to 13.2% in MC1. Meanwhile, QRAE demonstrates strong generalization to out-of-distribution TriviaQA and Natural Questions benchmarks.

Federated learning based on homomorphic encryption has attracted widespread attention for its strong security and enhanced protection of user data privacy. However, the nature of encrypted computation introduces three major challenges: computation efficiency, attack defense, and contribution assessment. The first concerns the efficiency of encrypted computation during model aggregation, the second involves defense malicious attacks under encryption, and the third addresses the fairness of contribution assessment for encrypted local models. This paper presents an Efficient and Secure Federated Learning Model with Homomorphic Encryption (ESFLM) to protect model privacy and tackle the aforementioned challenges. First, we leverage multiple nodes to perform parallel aggregation of local models, thereby improving the efficiency of encrypted model aggregation. Second, we introduce trusted supervise nodes to inspect local models when the global model is under attack, enabling effective defense of malicious behavior under homomorphic encryption. Finally, we fairly reward local training nodes based on their verified training time, even when local models remain encrypted. Experiments on three real-world datasets demonstrate that our model significantly outperforms baseline approaches in terms of both efficiency and security.

Toxicity in large language model (LLM) outputs refers to undesirable attributes in generated text, such as aggression, insults, discrimination, violent tendencies, and gender or racial bias. Researchers have investigated prompt-based explainability methods to identify features in prompts that are likely to elicit toxic outputs, with the goal of informing future efforts to mitigate toxicity. However, as prompt attacks grow more sophisticated, malicious intent is increasingly obfuscated and no longer explicitly manifested at the word level, resulting in the ineffectiveness of traditional word-level explainability approaches. Furthermore, existing semantic-level explainability methods focus primarily on relational analyses of isolated semantic features and lack a quantitative framework that integrates multiple semantic features. As a result, they are unable to accurately identify the key drivers of toxic outputs. To address this challenge, we propose a cascaded attention-based framework for semantic explainability of toxicity in large language models (CASET), which maps toxic outputs to the multi-semantic features of the input prompt. We first construct a semantic representation model that extracts features across four distinct semantic dimensions. Based on the inherent hierarchical structure of natural language, we propose a hierarchy-aligned cascading scheme that organizes attention modules into a unified toxicity prediction model. Finally, we extract the attention weights to assess the relative importance of each semantic feature. We evaluate CASET on the RealToxicityPrompts and PolygloToxicityPrompts datasets using both an open-source LLM (Llama-3-8B) and a closed-source LLM (GPT-4.1-nano). Experimental results show that CASET effectively identifies key semantic features responsible for toxicity and outperforms baseline explainability methods on the ℒ_top_3 and ℒ_top_5 metrics. Furthermore, case studies and human evaluation validate the practical utility of our framework, demonstrating that its optimization suggestions can significantly reduce toxicity in LLM outputs.

Federated learning has recently garnered significant attention owing to its extensive application scenarios. Nevertheless, existing unprotected communication mechanisms give rise to two critical challenges: “parameter leakage” and “low communication efficiency.” To address these issues, this paper introduces a Blockchain-based Federated Learning framework, termed FBChain, which is specifically designed to secure and optimize parameter transmission in federated learning. First, FBChain leverages the immutability of blockchain to store the global model and the hash values (hv) of local model parameters, thereby ensuring that the transmitted data cannot be tampered with. Meanwhile, parameter encryption is adopted to safeguard privacy, and data consistency is guaranteed by verifying the correspondence between local parameter hashes and the stored records. Through this design, the problem of parameter leakage is effectively mitigated. Second, a novel consensus mechanism, Proof of Weighted Link Speed (PoWLS), is developed to dynamically select nodes with higher weighted link speeds for global aggregation and block generation. By doing so, the model alleviates the inefficiency of communication that commonly arises in federated learning systems. Finally, experimental evaluations validate the proposed FBChain model, demonstrating its capacity to enhance communication security while significantly improving efficiency in federated learning environments.

Deep learning-based models demonstrate a remarkable level of accuracy in network traffic identification. However, the black-box nature of neural networks often makes the identification results difficult to explain. Although some eXplainable Artificial Intelligence (XAI) methods have been applied to network traffic identification, most of them focus on model explainability and do not provide sufficient credibility. In emerging network systems that use proprietary protocols, low-credibility malicious traffic detection can result in severe consequences. Therefore, it is imperative to deeply understand network traffic features and trust the detection results. In this paper, we propose an explainable architecture for emerging network systems. This architecture enhances the explainability of malicious traffic detection from both input and output perspectives, aiming to understand network traffic data and improve the reliability of the results. The effectiveness of explaining inputs and outputs is verified through experimental analysis in case studies. Furthermore, we review the research on explainable models in the field of network traffic identification and summarize research opportunities.

Android malware poses a significant challenge for mobile platforms. To evade detection, contemporary malware variants use API substitution or obfuscation techniques to hide malicious activities and mask their shallow semantic characteristics. However, existing research lacks analysis of the hierarchical semantic associated with Android apps. To address this problem, we propose ArchSentry, an enhanced Android malware detection via hierarchical semantic extraction. First, we select entities and their relationships relevant to Android software behavior through the software architecture and represent them using a heterogeneous graph. Then, we structure meta-paths to represent rich semantic information to achieve semantic enhancement and improve efficiency. Next, we design a meta-path semantic selection method based on KL Divergence to identify and eliminate redundant features. To achieve a comprehensive representation of the overall software semantics and improve performance, we construct a feature fusion approach based on Restricted Boltzmann Machines (RBM) and AutoEncoder (AE) during the pre-training phase, while preserving the probability distribution characteristics of various meta-paths. Finally, Deep Neural Networks (DNN) process fusion features for comprehensive feature sets. Experimental results on real-world application samples indicate that ArchSentry achieves a remarkable 99.2% detection rate for Android malware, with a low false positive rate below 1%. These results surpass the performance of current state-of-the-art approaches.

Retrieval-Augmented Generation (RAG) has proven effective in enhancing the factuality of LLMs’ generation, making them a focal point of research. However, previous RAG approaches overlook the lexical diversity of queries, hindering their ability to achieve a granular relevance assessment between queries and retrieved documents, resulting in suboptimal performance. In this paper, we introduce a Lexical Diversity-aware RAG (DRAG) method to address the biases in relevant information retrieval and utilization induced by lexical diversity. Specifically, a Diversity-sensitive Relevance Analyzer is proposed to decouple and assess the relevance of different query components (words, phrases) based on their levels of lexical diversity, ensuring precise and comprehensive document retrieval. Moreover, a Risk-guided Sparse Calibration strategy is further introduced to calibrate the generated tokens that is heavily affected by irrelevant content. Through these modules, DRAG is capable of effectively retrieving relevant documents and leverages their pertinent knowledge to refine the original results and generate meaningful outcomes. Extensive experiments on widely used benchmarks demonstrate the efficacy of our approach, yielding a 10.6% accuracy improvement on HotpotQA.

In addition to the autonomous driving technology of single vehicles, the inter-group control algorithm serving the data sharing of multi-vehicle cooperative driving has also attracted widespread attention. To ensure secure communication, many encryption schemes have been proposed to protect the interaction data between vehicles. Nevertheless, traditional public key encryption schemes hinder the sharing of encrypted data. Based on the premise of ensuring the confidentiality of encrypted information, in order to facilitate efficient data sharing, conduct data searches across extensive cloud-based datasets, and authorize access under specified conditions, we introduce Fine-Grained Access Control with Privacy-Preserving Data Retrieval (FGAC-PPDR) for the Internet of Vehicles. This scheme offers a secure, flexible, and privacy-centric approach to data sharing for groups of vehicles in the IoV. Our proposed scheme enables encrypted data to be retrieved at the group level by the cloud server, and prevents vehicles outside the group from performing equality tests on the ciphertext. Furthermore, the data owner can create an authorization token with defined conditions to specify how the data is shared. During the process of data search and sharing, intensive computing tasks are undertaken by cloud servers with abundant computational resources. We also demonstrate that our scheme is secure against chosen ciphertext attacks (CCA). Finally, we provide security and performance analyses that verify the feasibility and effectiveness of our proposal.

Large Language Models (LLMs) have demonstrated broad applications but suffer from issues like hallucinations, erroneous outputs and outdated knowledge. Model editing emerges as an effective solution to refine knowledge in LLMs, yet existing methods typically depend on structured knowledge representations. However, real-world knowledge is primarily embedded within complex, unstructured text. Existing structured knowledge editing approaches face significant challenges when handling the entangled and intricate knowledge present in unstructured text, resulting in issues such as representation ambiguity and editing conflicts. To address these challenges, we propose a Conflict-Aware Knowledge Editing in the Wild (CAKE) framework, the first framework explicitly designed for editing knowledge extracted from wild unstructured text. CAKE comprises two core components: a Semantic-augmented Graph Representation module and a Conflict-aware Knowledge Editing strategy. The Semantic-augmented Graph Representation module enhances knowledge encoding through structural disambiguation, relational enrichment, and semantic diversification. Meanwhile, the Conflict-aware Knowledge Editing strategy utilizes a graph-theoretic coloring algorithm to disentangle conflicted edits by allocating them to orthogonal parameter subspaces, thereby effectively mitigating editing conflicts. Experimental results on the AKEW benchmark demonstrate that CAKE significantly outperforms existing methods, achieving a 15.43\% improvement in accuracy on llama3 editing tasks. Our framework successfully bridges the gap between unstructured textual knowledge and reliable model editing, enabling more robust and scalable updates for practical LLM applications.

Federated learning based on homomorphic encryption has received widespread attention due to its high security and enhanced protection of user data privacy. However, the characteristics of encrypted computation lead to three challenging problems: “computation-efficiency", “attack-tracing" and “contribution-assessment". The first refers to the efficiency of encrypted computation during model aggregation, the second refers to tracing malicious attacks in an encrypted state, and the third refers to the fairness of contribution assessment for local models after encryption. This paper proposes a federated learning storage security model with homomorphic encryption (FLSSM) to protect federated learning model privacy and address the three issues mentioned above. First, we utilize different nodes to aggregate local models in parallel, thereby improving encrypted models' aggregation efficiency. Second, we introduce trusted supervise nodes to examine local models when the global model is attacked, enabling the tracing of malicious attacks under homomorphic encryption. Finally, we fairly reward local training nodes with encrypted local models based on trusted training time. Experiments on multiple real-world datasets show that our model significantly outperforms baseline models in terms of both efficiency and security metrics.

The widespread adoption of encryption protocols has increased the complexity of detecting malicious Android traffic. By randomizing payload content, encryption obscures semantically explicit features in network traffic, thereby concealing its behavioral intent. Although existing methods mitigate this issue by expanding feature sets or extracting spatiotemporal patterns, they do not fundamentally reconstruct the original payload semantics. In this paper, we propose RATD, a detection model that enhances encrypted traffic representation by introducing semantics of source-APIs. This approach leverages the correlation between system API calls made prior to traffic transmission (referred to as source APIs) and the behavioral intent within encrypted traffic, thereby compensating for semantic loss. First, we construct API-traffic association samples by monitoring network connection APIs. Then, we transform the API sequences into graphs and apply a Graph Convolutional Network (GCN) to learn their structural and semantic representations. These features are fused with corresponding traffic features through a multi-source encoder module. Finally, to address the challenges of limited data availability in real-world deployment, we introduce a representation enhancement module to improve model’s robustness in scenarios with missing data. Experimental results show that RATD is significantly better than the state-of-the-art models across multiple datasets. In particular, in scenarios with missing API data, the accuracy of our model decreases by at most 2.9%, showing a stronger environmental adaptability.

The rapid proliferation of internet data, particularly through social media, has amplified the need for effective sentiment analysis, including the complex task of sarcasm detection. This paper presents a novel multi-modal sarcasm detection model leveraging cue learning techniques to address the challenges posed by data scarcity, especially in low-resource languages. The proposed model builds upon the CLIP architecture, integrating text and image modalities to co-learn sarcasm cues. The methodology encompasses discrete prompt generation, learnable continuous vectors, and multi-modal fusion to enhance detection accuracy. The multi-modal fusion process demonstrates a symmetric integration of text and image data, leading to improved performance. Experimental results on the Twitter Multi-modal Sarcasm Detection Dataset (MSD) demonstrate significant performance improvements over traditional models, highlighting the model’s robustness and adaptability in small-sample scenarios. This research contributes a practical solution for nuanced sentiment analysis, paving the way for advanced applications in public opinion monitoring and AI-driven decision-making processes.

Detecting Twitter bots is essential for combating misinformation and maintaining the integrity of online social networks. Existing methods often overlook the high-order interactions and heterogeneous relationships among users and tweets, limiting their effectiveness in addressing sophisticated bot behaviors. This article introduces HHG-Bot, a novel hyper-heterogeneous graph-based framework for Twitter bot detection. The proposed approach integrates heterogeneous graph convolutional networks with a trainable hypergraph aggregation model to capture complex, high-order interactions. To overcome the challenge of labeled data scarcity, HHG-Bot employs a meta-learning paradigm that enhances the model’s generalization capability across different bot types. Experiments conducted on the Twibot-20 benchmark dataset demonstrate that HHG-Bot achieves state-of-the-art performance, surpassing existing methods in terms of accuracy (86.17%), F1-score (87.51%), and Matthews correlation coefficient (MCC) (71.75%). The results validate the effectiveness of leveraging hypergraphs and meta-learning for detecting Twitter bots, particularly in scenarios with limited labeled data.