The rapid adoption of Internet of Healthcare Things (IoHT) devices has significantly enhanced healthcare services but also exposed systems to cyber threats and data breaches, as medical data is highly sensitive and vulnerable. Current cybersecurity mechanisms often fail to ensure proper integrity, confidentiality, and availability of healthcare data, while noisy datasets and limited computational resources further complicate attack detection and prevention. To address these challenges, this study proposes a blockchain-oriented cybersecurity framework for IoHT, termed Nested Dense Graph Attention Capsule Networks with Circulatory System-Based Optimization for Robust Data Protection (NDGACN-CSBO). The framework leverages data from CICIoMT2024 and WUSTL EHMS 2020 datasets, which are preprocessed using the Adjusted Min–Max with Decimal Scaling and Statistical Column Normalization (AMM-DSSCN) technique to handle inconsistencies and scale features. Feature selection is performed via the Sharpbelly Fish Optimization (SFO) algorithm, ensuring that the most relevant features are retained. Classification of normal and attack patterns is conducted using Nested Dense Graph Attention Capsule Networks (NDGACN), whose performance is further enhanced through Circulatory System-Based Optimization (CSBO) for weight parameter tuning. To secure data sharing, a blockchain-integrated Generation Right Trading (GRT) model with digital smart contracts is employed, enabling tamper-resistant, authenticated, and reliable data exchange in cloud environments. Experimental results demonstrate that NDGACN-CSBO achieves 99.23 and 99.88
The K-Nearest Neighbor (kNN) search problem faces significant challenges in high-dimensional spaces due to the curse of dimensionality, which severely degrades the performance of traditional index structures and approximate methods. To address these limitations, we propose KLF (KNN Learned Filter), a neural filtering framework that reformulates kNN search as a dynamic probability-based filtering task. KLF replaces per-query distance computations with neural inference over a pre-trained model, shifting the nature of online computation. KLF employs a multi-layer perceptron to predict the likelihood that each data point is a true k-nearest neighbor to a given query, producing a probability score for every point in the index. A density-aware adaptive threshold then filters these scores to generate compact candidate sets. On the GloVe−200D dataset, KLF achieves 96.6
Digital asset markets are expanding but remain difficult to evaluate, particularly for retail investors who face information overload, unstructured disclosures, and limited regulatory safeguards. Traditional due diligence methods often fail to capture the complexity of blockchain-based projects, exposing investors to misrepresentation and fraud. At the same time, LLMs have shown remarkable results on analyzing large information quantities. Thus, this paper investigates whether large language models (LLMs) can enhance digital asset assessments by providing inspectable, automated, and low-cost decision support. Using a Design Science Research methodology, we developed a framework and a prototype tool to apply LLM-based evaluation to token-offering whitepapers, evaluated across 1,139 generated runs covering 29 token offerings. We compared different content processing approaches, from raw document analysis to aspect-focused summarization, and tested prompting strategies including Zero-Shot, Few-Shot, and Chain-of-Thought reasoning. Our findings show a nuanced pattern. On aggregate quantitative metrics, structured summarization combined with Chain-of-Thought prompting achieves the lowest rating error, although differences in rank correlation between configurations are not statistically resolvable given the number of projects evaluated; aspect-focused summarization structures criterion-specific evidence and candidate red flags explicitly, but its rating error is higher and the interpretability of this representation was not independently evaluated. To address the single-model limitation, we added a lightweight multi-model robustness check using two small open-source models on a matched subset of cases; the results suggest that the observed optimism bias is not specific to the original closed LLM, although performance still varies across model configurations. Overall agreement between model and expert ratings remains limited, which suggests that LLMs are better understood as decision-support tools for screening and explanation than as substitutes for expert judgment. The study therefore contributes a design-oriented due diligence artifact and a dependence-aware evaluation of how summarization and prompting choices affect calibration, output-level inspectability, and validity boundaries in digital asset assessment.
Link prediction (LP) algorithms play a critical role in complex networks by inferring potential future connections based on structural patterns and user interaction behavior. However, most existing LP methods struggle to capture hierarchical feature relationships and often fail to model complex part–whole interactions between nodes. These constraints decrease their generalization ability particularly in networks with sparse or noisy connectivity patterns. In order to address these limitations, this paper presents a new hybrid framework, the Graph-Lite Capsule Network (GLCN) that consists of Graph Convolutional Networks (GCNs) and Lite Capsule Network (LCN). GCN creates node embedding that is regularized by dropout and dimensionality reduction is performed using Principal Component analysis (PCA). These embeddings are processed by LCN, where a Rectified Linear Unit (ReLU) activation is used to better identify meaningful features. After this, the squash function is used to capture meaningful feature relationships, followed by an optimized routing process. At the last stage, the model uses the improved capsule representations to find possible links in the graph. This architecture is very effective in modeling the structural relationships within the graph to ensure greater efficiency. LCN reduces routing steps by eliminating discrete vote matrices and focuses on relevant interactions among the capsules. To understand the efficiency of the proposed method, a comparative evaluation has been performed on six different datasets, each one being a representation of different interaction and collaboration patterns. The findings show that there is an average improvement of about 9.1
Within the legal domain, named entity recognition (NER) is an essential pre-processing tool for important tasks such as information extraction and redaction. Research to date has focused on rule-based approaches or machine learning strategies that require significant time and effort to train through labelled corpora. Large, multi-purpose generative large language models (LLMs) have been successfully employed on natural language processing (NLP) tasks with no prior supervised training and may represent a way forward in addressing these existing bottlenecks. Despite promising results in certain NLP use cases, the Few-Shot LLM performance on unfamiliar, domain-specific NER tasks has, to date, remained uncertain. In this legal clause NER task, we evaluate a Few-Shot, Chain-of-Thought (CoT) prompting strategy for OpenAI’s GPT-4 LLM on a new CUAD-NER dataset that we have repurposed for the task of Person and Organisation entity identification. In a comparative study, we demonstrate that the GPT4 + Few-Shot CoT prompt strategy outperforms existing state-of-the-art models, achieving a 5
