Recently, the cyberattack surface has expanded significantly due to the proliferation of Internet of Things (IoT) devices, and intrusion detection remains a vital component of modern network security. Traditional Intrusion Detection Systems (IDSs) often cannot detect complex and newly generated threats due to the heterogeneous characteristics and constrained computing capabilities of IoT devices. In this paper, we utilise ensemble learning techniques, specifically Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM), in combination with a Feed-Forward Neural Network (FFNN) to develop a hybrid IDS. To address class imbalance and complex feature interactions that can be managed by Gradient Boosting (GB) techniques, the proposed model combines the robustness of gradient boosting with that of deep learning in learning complex nonlinear patterns in network traffic. The experimental results show the hybrid model achieves strong performance, such as an accuracy of 99.42%, precision of 98.43%, recall of 99.65%, F1-score of 99.04%, and Area Under the receiver operating characteristic Curve (AUC) of 0.9994. These findings indicate the effectiveness of the proposed hybrid IDS in securing the IoT from diverse online threats.
The automation of penetration testing has long been constrained by the semantic-operational divide between human expert cognition and machine-executable workflows. We present COAPT, a novel architecture that bridges this gap through large language model (LLM) driven cognitive planning, enabling autonomous execution of full-cycle penetration testing aligned with the penetration testing execution standard (PTES) and MITRE adversarial tactics, techniques, and common knowledge (ATT&CK) framework. COAPT introduces three key innovations: (1) A cognitive planning architecture combining retrieval-augmented generation with chain-of-thought reasoning, grounding decisions in verified cybersecurity knowledge; (2) Formal semantic interfaces that translate strategic intent into executable commands for security tools through machine-actionable contracts; (3) Hierarchical rnulti-agent coordination that maintains tactical consistency across reconnaissance, exploitation, privilege escalation, and defense recommendation phases. Evaluated across diverse penetration testing scenarios, COAPT demonstrates superior performance over state-of-the-art tools in vulnerability exploitation success rates and operational efficiency, while significantly reducing LLM hallucination risks through its knowledge-grounded reasoning approach. The architecture's ability to autonomously chain multi-stage attacks and generate ATT&CK-mapped defense strategies establishes a new paradigm for cybersecurity automation that preserves human expert workflows at machine execution scales.
Deep learning offers notable promise for computational pathology, but its performance is constrained by the need for extensively annotated datasets, which are costly and laborious to produce. Self-Supervised Learning (SSL) provides an effective paradigm for learning discriminative representations from unannotated pathological images. However, existing SSL methods often overlook domain-specific characteristics of pathological images and suffer from the adverse effects of low-quality negative samples, leading to sub-optimal feature representations for downstream tasks. To overcome these limitations, we propose a novel Domain-Specific Self-supervised Contrastive Learning (DSSCL) framework, which incorporates two novel components: (1) a Stain-Separation Based Data Augmentation (SSDA) module that enhances stain-aware representation learning by fusing stain-separated components with original hematoxylin and eosin images, and (2) a Contrast-Aware Pair Refinement (CAPR) module that improves feature discriminability by filtering potential positives and mining hard negatives, thereby mitigating the influence of low-quality negatives. Extensive experiments demonstrate that DSSCL achieves comparable accuracy in classification tasks using only 0.1% labeled data compared to a network fine-tuned from ImageNet with 10% labeled data, while also delivering competitive performance in detection and segmentation tasks, underscoring its effectiveness in learning transferable and robust feature representations across diverse downstream tasks. The code is available at https://github.com/junjianli106/DSSCL.
Artificial Intelligence of Things (AIoT)-enabled Edge Collaboration systems (AEC) are characterized by openness, heterogeneity, and constrained device resources, which pose severe challenges to achieving trusted collaboration within AEC. While blockchain provides a trusted foundation for multi-party coordination, establishing robust inter-device trust relationships at low cost and promoting the collaborative participation of heterogeneous devices remain critical challenges that need to be addressed. To address the aforementioned challenges, this paper introduces DPRP, a lightweight and scalable multilayer blockchain model based on delegated proof of swarm reputation with hierarchical Practical Byzantine Fault Tolerance (PBFT). The model introduces a lightweight consensus protocol, DPoRP, which constructs swarm reputation by evaluating node contributions within collaborative swarms, thereby establishing a trust foundation. This enables a swarm-reputation-driven multilayer consensus mechanism, structured around meta-nodes functioning as autonomous units. To effectively incentivize swarm collaboration and constrain node behavior, DPoRP incorporates a hierarchical reward and penalty scheme tailored to swarm collaboration. Moreover, to reduce the participation threshold for heterogeneous devices, DPRP designs a dynamic, redundancy-aware hierarchical storage mechanism that adjusts redundancy factors based on actual storage demands, thus improving storage efficiency and supporting multilayer deployment. Simulation results and analysis demonstrate that DPRP significantly enhances system robustness and scalability while reducing communication complexity.
The objective of multidimensional graph fraud detection is to identify fraudulent entities within a graph. Graph Neural Network (GNN) models leverage graph structures to propagate messages from neighbors to target nodes, thereby obtaining precise representations of the target nodes. However, anomaly detection methods based on GNNs face challenges, such as structural inconsistency and over-smoothing, which reduce the suspicion scores of fraudulent nodes and hinder the effectiveness of anomaly detection models. To address these issues, we propose an end-to-end anomaly node detection model named Feature Decomposition Graph Neural Network (FD-GNN). In FD-GNN, a feature amplification module is first employed to enhance the differences between node representations. Then, a feature decomposition module is used to distinguish the natural attributes of nodes from their neighborhood attributes, with constraints applied to both types of attributes. Finally, a representation aggregation module utilizes differential aggregation operations to further distinguish normal nodes from anomaly ones. Experiments conducted on two real-world datasets, Amazon and YelpChi, demonstrate that FD-GNN achieves Area Under Curve (AUC) scores of 94.45% and 91.62%, respectively, outperforming existing multidimensional graph anomaly detection models.
The advancement of the Internet of Medical Things (IoMT) has rendered the clustering method pivotal in medical data analysis. Nevertheless, the exposure of patient privacy during this process constitutes a significant security concern. To achieve privacy protection for cluster analysis while ensuring the accuracy of clustering results in IoMT, this paper proposes a Differential Privacy-based availability-enhanced k-modes Clustering scheme, named DPKC. First, we introduce partition entropy to quantify attribute weights and calculate the initial center based on density with distance. This can mitigate the impact of random initialization, thus improving the accuracy of the clustering results. Second, we introduce cluster weights to calculate the distance between data points and centers during allocation. This adjustment aims to reduce the difference in compactness between clusters. Third, we employ a geometric mechanism to inject noise into the frequency of attribute values, ensuring data privacy during the iteration process. Theoretical analysis proves that DPKC satisfies differential privacy and prevents information disclosure. Experimental results show that DPKC improves the F-measure metric by 7.36% compared to existing algorithms.
Internet of things (IoT) technology is important in the era of big data, and in most IoT scenarios, data are constantly arriving in the form of streams. Streaming data have a fast frequency and a large volume, which, over time, has no upper limit. Streaming data contain much valuable information, such as the behaviors of IoT objects, which can be mined through frequent pattern mining algorithms. Due to high spatial complexity, such methods must adopt window mechanisms to reduce spatial and temporal costs. However, the statistics of frequent patterns based on window patterns are local, and for global data, some important patterns may be overlooked. Therefore, we propose a frequent pattern mining method for streaming data based on a new data structure, the short-pattern graph (SP-graph), aiming to significantly improve the upper limit of the data volume. The proposed algorithm does not need to set windows, or set sufficiently large windows to avoid missing important patterns. We first define the SP-graph data structure, and propose a frequent pattern mining algorithm based on it, as well as a real-time object behavior prediction algorithm and pattern-containing mining algorithm. We experimentally compare our method with classical methods on a real streaming dataset, and verify its much lower spatial complexity, greater data capacity, ability to flexibly adjust the frequency threshold, and offline and online mining efficiency, which enable real-time performance.
In the context of big data enabling e-commerce, content platforms, and social networks, Recommendation Systems (RSs) play a crucial role in providing personalized items and services suggestions to users. Among the various RSs approaches, Collaborative Filtering (CF) approaches, particularly Singular Value Decomposition++ (SVD++) algorithm, have gained widespread adoption due to their ability to leverage both explicit and implicit feedback derived from user history interaction data. This data includes user-item status, purchase history, and user interaction. However, traditional matrix-based methods, such as SVD++, often struggle to capture the multidimensional features and temporal dynamics inherent in real-world user-item data. To address this limitation, we propose a novel tensor-based High-Order SVD++ (HO-SVD++) recommendation method. This approach employs tensors to model multi-feature interaction data, with a particular emphasis on temporal dynamics. Specifically, we introduce a novel method for segmenting recommendation data based on user rating periods and construct a tensor to encapsulate these temporal features. Additionally, we propose a user-item correlation CF method that extracts implicit feature relationships between users and items. Building on this, we present the HO-SVD++ method, which is specifically designed for recommendation tasks involving multiple latent factors. Furthermore, we introduce a comprehensive recommendation framework based on the HO-SVD++ method. Experimental results on three classical recommendation datasets demonstrate that the proposed HO-SVD++ algorithm outperforms several classical and neural recommendation baselines in terms of recommendation accuracy.
Recently, Few-shot Knowledge Graph Completion (FKGC) has emerged as a significant research area, yet it encounters challenges stemming from the complexity of multi-semantic relationships in few-shot scenarios. To address these challenges, we propose Task-guided Diffusion Prototype network for FKGC (TDP-FKGC), a method that generates high-quality prototype representations via a task-guided diffusion process. Initially, we analyze the semantics of entity pairs, leveraging attention mechanisms to select pertinent reference pairs from the support set for the creation of a preliminary prototype. Subsequently, a task-guided diffusion process is formulated within this prototype space, and a conditional denoising model is employed to produce task-specific prototype representations. Experimental results demonstrate that TDP-FKGC outperforms current state-of-the-art FKGC methods on three widely used datasets. Furthermore, ablation experiments and analysis of different relationship types confirm the effectiveness and multi-semantic handling ability of our proposed model.
Anomaly detection plays a critical role in ensuring the security of data and systems across diverse real-world applications. Traditional anomaly detection relies on collecting large datasets on a central server, but in reality, data are often spread across different clients and cannot be directly shared due to privacy concerns. Federated learning (FL) has thus emerged as a promising framework for privacy-preserving anomaly detection by enabling collaborative model training without exposing raw data. However, a major challenge in FL-based anomaly detection (FLAD) is heterogeneity, i.e., clients often have data with different distributions, feature spaces, and resource constraints. This makes it difficult to build accurate and reliable anomaly detection models. While some surveys have explored aspects, such as privacy protection, anomaly detection techniques, or FL methodologies, a comprehensive review of privacy-preserving anomaly detection within heterogeneous FL settings remains lacking. This paper systematically review and categorize anomaly detection methods designed for secure and reliable use in heterogeneous FL, considering both data and client heterogeneity. Finally, we highlight future research directions, aiming to guide further progress and support the wider adoption of FLAD in real-world scenarios.
Mobile Edge Computing (MEC) has emerged as a viable means to improve processing efficiency by reducing distance between computation and Internet of Things (IoT) devices, results in lower latency and energy expenditure. Although MEC appears to be a good candidate for improving task distribution efficiency, there remain issues surrounding task offloading, high computational latencies, and the heterogeneity of IoT devices that make centralized resource allocation seamless. This research presents an improved task offloading paradigm, with an emphasis on reducing latency and improving efficiency in an IoT-MEC context. Incoming tasks are classified into four categories based on levels of complexity. Type 1 tasks can be executed on the IoT device, while Type 4 task are fully offloaded into a cloud. Type 2 and Type 3 tasks utilize a parallel execution approach, where tasks are partially executed on the IoT device and partially satisfies the MEC server. The model in this study incorporates matching and queuing theory to improve task allocation from the device to the server. An additional branch based heuristic method is also introduced to further reduce task processing efficiency. Algorithm used was implemented using Python libraries through simulations and real-world cases. Comparative analysis with baseline task offloading methods demonstrates substantial performance improvements, achieving reductions in task computation latency of 40%-70%. These results demonstrate that the proposed methodology is efficient and effective in optimizing the overall execution of task in MEC-enabled IoT networks.
Accurate diagnosis of Alzheimer's Disease (AD) is essential for early intervention. Traditional methods relying on single-modality data often fail to capture the complexity of the disease, limiting diagnostic accuracy. Integrating multimodal data, such as structural Magnetic Resonance Imaging (sMRI) and Single Nucleotide Polymorphism (SNP) data, can provide a more comprehensive understanding of AD. However, existing multimodal fusion methods often overlook the intricate relationships among different data types, resulting in suboptimal performance. To address these challenges, we propose a novel graph-based multimodal fusion framework for AD prediction. The framework constructs brain and gene ontology networks using domain-specific prior knowledge from sMRI and SNP data. It leverages Graph Convolutional Networks (GCN) to extract deep features from each modality and employs a cross-attention mechanism to dynamically weigh feature importance across modalities. Additionally, a Correlation-Aware Learning (CAL) module explicitly models inter-modal correlations, enhancing the interpretability and robustness of the fusion. We validate the effectiveness of our framework using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Results show that our framework significantly outperforms traditional methods in classification accuracy and feature representation. Our method enables accurate AD diagnosis by integrating multimodal data and explicitly modeling inter-modal correlations. It enhances the interpretability of multimodal integration and provides new insights into the genetic and structural mechanisms underlying AD, serving as a valuable tool for clinical diagnosis and research in neurodegenerative diseases.
Data synthesis under Local Differential Privacy (LDP) presents a promising approach for private data analysis and sharing, as it enables the execution of all analysis tasks on raw data without the need for a trusted aggregator. The select-measure-generate paradigm of data synthesis under Differential Privacy (DP) introduces specific challenges in the context of LDP, particularly because the noise inherent to LDP is significantly greater than that of DP, especially in high-dimensional datasets. The “select” step involves calculating the correlations between attributes to identify important marginal measurements (attribute pairs), while the “measure” step aims to estimate the frequency distribution of each selected marginal under LDP. However, the utility of both the correlation and frequency estimation for multidimensional data is often unsatisfactory under LDP, as the utility of data analysis tasks typically declines with an increasing number of dimensions. To address these issues, we propose a two-stage method, named FilterLDPSyn. In Stage 1, it filters out ineffective measurements based on one-dimensional frequency and entropy estimations under LDP. In Stage 2, it enhances the utility of the distribution by iteratively collecting two-dimensional values and restoring consistency between one- and two-dimensional distributions. Experimental results demonstrate the superiority of our proposed method over existing approaches.
Existing Two-Dimensional (2D) rectal tumor segmentation methods struggle to fully exploit the relationships between slices due to the absence of Three-Dimensional (3D) spatial information, while traditional 3D segmentation techniques often suffer from poor performance and low rotational robustness. To address these limitations, we propose a 3D Magnetic Resonance (MR) segmentation model based on a deep supervised residual capsule network (namely DRCU-Net). This model introduces a capsule module built on the 3D U-Net architecture, allowing it to capture the spatial hierarchical features of tumor tissue through a dynamic routing mechanism, thereby enhancing the model's robustness to rotation. Additionally, we employ deep supervision mechanisms to improve model performance and modify the LayerNorm function to extend layer normalization to 3D, facilitating its adaptation to the processing requirements of 3D data. We evaluate our model on the dataset from Shanxi Cancer Hospital (China), it achieves a Dice score of 0.7580 and a Mean Intersection over Union (MIoU) score of 0.6258, which demonstrating its superiority.
Generally, there are two popular ways to protect image copyright, i.e., proactive protection (preventing illegal use via adversarial perturbation) and passive protection (verifying ownership by digital watermarking). However, since the perturbation and watermark embedded into an image will interfere with each other, directly embedding them into the image cannot achieve the proactive protection and passive protection, simultaneously. To address this issue, we propose an image copyright dual-protection approach, which embeds an Extractable and Imperceptible Adversarial Watermark (EIAW) in the image frequency-domain. Specifically, the adversarial watermark is automatically embedded and optimized in the manner of allowing for effectively attacking the Deep Neural Networks (DNNs) and accurately extracting the embedded watermark, simultaneously. Moreover, instead of using the pixel-domain constraints, i.e., Lp norms, we introduce a frequency-domain constraint to optimize the watermark embedding locations. Experiments on ImageNet and CIFAR-10 demonstrate that the proposed EIAW achieves high attack effectiveness (up to 100%) and extraction accuracy (up to 93%), while maintaining good watermark imperceptibility.
The use of Large Language Models (LLMs) in data cleaning tasks has demonstrated impressive capabilities. However, the high inference costs associated with LLMs pose significant challenges, particularly when managing large-scale datasets within constrained budgets. While many studies focus on direct methods to reduce inference costs, we propose a novel framework to alleviate the high inference costs of LLMs by transforming the task into a multi-objective optimization problem. This framework begins by decomposing the complex data cleaning task into smaller, well-defined sub-tasks. For each sub-task, the most appropriate method is selected from a range of options, such as rule-based tools, code generation methods, smaller pretrained language models, or LLMs, depending on the tradeoff between cost and effectiveness. This allows for a systematic balance between cost and quality, enabling the completion of high-quality data cleaning tasks within budget constraints. Experimental results validate the effectiveness of this approach. The framework significantly reduces inference costs while maintaining high-quality data processing. This framework offers a practical pathway to optimizing LLM-based data cleaning methods, balancing computational efficiency and data processing quality. Future work could explore the dynamic adaptations for evolving sub-tasks or deeper integrations with explainable AI and human-in-the-loop approaches to enhance trust and interpretability in data cleaning pipelines.
Imbalanced data distribution is a common challenge that hinders model performance across various industries, particularly due to the scarcity of faulty samples. Sample generation can help address this issue, but existing methods often overlook causal structures and dependencies, resulting in low-quality synthetic samples. Hence, a novel generative method, Causal-Enhanced Latent Space Wasserstein Generative Adversarial Network (CE-LSWGAN) with gradient penalty, is proposed for imbalanced discrete integer data. First, a causal directed acyclic graph is extracted using the Peter-Clark (PC) algorithm. The causal dependencies are then embedded into the latent space of a Variational Autoencoder (VAE) through graph propagation, creating a constrained structural causal prior. Next, a Generative Adversarial Network (GAN) is introduced in this structured latent space to train the latent variables of encoder. A gradient penalty is applied to ensure stable training. To further enhance the generation of minority class samples, a dynamic resampling strategy based on class weights is employed. Finally, within the VAE decoder, the GAN-generated latent variables are integrated with causal constraints, ensuring that generated samples adhere to causal logic. Experimental results demonstrate that CE-LSWGAN achieves an average causal consistency of 0.6088 and improves similarity between generated and original data by 18.27% compared to six state-of-the-art baselines. Downstream classifiers show substantial performance gains, with the F1-score increasing by up to 60% compared to the baseline models without data augmentation. The model attains precision and recall scores of 0.9351 and 0.9352, respectively. Overall, CE-LSWGAN demonstrates superior performance in generation quality, causalfidelity, and task adaptability, providing a reliable data augmentation framework for the causal-based applications.
Retrieval-Augmented Generation(RAG)represents a transformative advancement for Large Language Models(LLMs)by integrating external knowledge to substantially improve accuracy and mitigate hallucinations.As a pivotal technology in the contemporary generative Artificial Intelligence(AI)landscape,RAG addresses fundamental challenges in knowledge-intensive tasks.
Knowledge Graphs (KGs) are used to organize and understand vast amounts of information. However, they usually suffer from incompleteness which limits their applicability. Knowledge Graph Completion (KGC) is a critical process for augmenting these graphs and is typically achieved through embedding-based and pre-trained language model based methods. Although Large Language Models (LLMs) have demonstrated significant potential for knowledge extraction and reasoning, they face challenges such as fact knowledge hallucinations which adversely affect their KGC performance. We propose Knit, a novel KG-integrated instruction tuning framework to alleviate LLM fact knowledge Hallucinations in KGC. The proposed framework comprises three key components: (1) KG-integrated information adapter, (2) knowledge prompts, and (3) KG-integrated instruction tuning strategy. These components enhance LLMs' ability to recognize entity relationships, improve embedding interpretability, and ensure response consistency. The experimental results across four public datasets (WN11, FB13, WN18RR, and YAGO3-10) demonstrate that Knit achieves state-of-the-art performance in KGC with significant improvements in all KGC subtasks.
Recently, with the rapid advancements in Large Language Models (LLMs), LLM-based Open-domain Question Answering (OpenQA) methods have reaped the benefits of emergent understanding and answering capabilities enabled by massive parameters compared to traditional methods. However, most of these methods encounter two critical challenges: how to integrate knowledge into LLMs effectively and how to adaptively generate results with specific answer formats. To address these challenges, we propose a novel framework, which aims to improve the OpenQA performance by exploring knowledge integration and controllable generation on LLMs simultaneously, namely GenKI. Specifically, we first train a dense passage retrieval model to retrieve associated knowledge from a given knowledge base. Subsequently, we introduce a novel knowledge integration model that incorporates the retrieval knowledge into instructions during fine-tuning to intensify the model. Furthermore, to enable controllable generation in LLMs, we leverage a certain fine-tuned LLM and an ensemble framework based on text consistency incorporating all coherence, fluency, and answer format assurance. Finally, extensive experiments conducted on three datasets with diverse answer formats demonstrate the effectiveness of GenKI with comparison of state-of-the-art baselines. Moreover, ablation studies have disclosed a linear relationship between the frequency of retrieved knowledge and the model's ability to recall knowledge accurately with the ground truth. Tests focusing on the out-of-domain scenario and knowledge base independence scenario have further affirmed the robustness and controllable capability of GenKI. Our code of GenKI is available at https://github.com/USTC-StarTeam/GenKI.
