Insider threats come from people who have authorized access, hence they are harder to detect using traditional rule-based systems. This makes them one of the toughest cybersecurity challenges. This study investigates artificial intelligence (AI) and machine learning (ML) to create a more flexible framework. The study also discusses Explainable AI (XAI) methods, which helps analysts understand anomalies better by giving clear human-readable explanations. The suggested hybrid model combines anomaly detection and behavioral analytics to instantly find suspicious user activity and to learn typical behavioral patterns, it uses supervised and unsupervised techniques like auto-encoders and Isolation Forests classifiers. Hybrid learning models improve detection accuracy, lower false positives, and adjust to changing conditions. The proposed model is evaluated on the CERT Insider Threat Dataset (CMU-SEI), which simulates realistic insider behaviours. This research contributes to creating proactive, transparent, and intelligent defense mechanisms, shifting cybersecurity from a reactive approach to a preventive and adaptive strategy for protecting critical digital infrastructures.
A random or pseudorandom sequence of bits is important and used in cryptography, simulations, and mathematics. These values are created using a random or pseudorandom binary generator. It is essential that these random strings have a high period and high linear complexity, and that they pass statistical randomness tests to ensure their effectiveness. A generator must be developed taking all these factors into account. Each step of the process must be carefully controlled to guarantee a good final result. To ensure that the pseudorandom binary generator is sufficiently robust, cryptographic components must be combined correctly. This project demonstrates how to create a reliable pseudorandom binary generator. The most important parts of the generator described in this presentation are the New Self-Shrinking Generators (NSSGs). These are combined using Threshold Generators in one case and Geffe Generators in another.
Non-Fungible Tokens (NFTs) are widely adopted across digital art, gaming, and decentralized applications. However, existing NFT marketplaces largely treat NFTs as static assets with limited interactivity and narrative depth. This paper presents Demon Forge Hub, a decentralized NFT marketplace designed to address these limitations through on-chain gamification, structured storytelling, and improved accessibility. Built using ERC-721 smart contracts deployed on the Polygon Mumbai network, the platform enables NFTs to function as evolving digital characters with persistent attributes such as levels, rarity, and battle outcomes. Metadata is stored using NFT.Storage and IPFS to ensure verifiable persistence. Additionally, an AI-assisted lore generation module produces structured narrative identities for each NFT. The proposed system demonstrates how NFT marketplaces can transition toward interactive, narrative-driven digital ecosystems.
As the demand for edge computing grows, there is a critical need for learning mechanisms that offer both energy efficiency and temporal predictability. In this study, we present a single-board implementation of a Coupled Learning Loop Network (CLLN), a nonlinear analogue architecture based on self-adjusting resistive elements. Unlike traditional von Neumann systems, this architecture learns via physical relaxation processes rather than gradient descent algorithms. We constructed a 16-edge network integrated with an FPGA to manage reconfigurable inputs and verified its operation through a sine waveform regression task, demonstrating rapid convergence. To validate scalability and performance on high-dimensional tasks, we developed a physics-informed digital twin of the system and benchmarked it against the MNIST dataset using CPU, GPU, and TPU architectures. While specialised accelerators (TPUs) running TensorFlow Lite demonstrated highly reduced latency, our analysis reveals that digital frameworks remain subject to latency jitter (up to 110 s standard deviation on standard processors) due to discrete memory access and operating system overheads. In contrast, the analogue network exhibits high temporal determinism, with settling times governed strictly by RC time constants. Combined with a measured power envelope of approximately 0.45 W, this intrinsic stability and efficiency suggest that FPGA-driven nonlinear analogue networks offer a robust solution for latency-critical and energy-constrained edge applications.
Residue Class (RC)-based peer-to-peer (P2P) networks employ structured group-head overlays to achieve scalable, locality-aware routing; however, the reliance on intermediate group heads exposes the forwarding process to insider attacks. A malicious group head can modify, replace, or forge forwarded data while remaining undetected, undermining end-to-end integrity. This paper proposes a Non-Interactive Zero-Knowledge Proof (NIZKP)–based forwarding mechanism that enforces cryptographically verifiable hop-by-hop integrity in RC-based P2P networks. Each group head attaches a NIZKP proving that the outgoing ciphertext is a correct re-encryption of the same plaintext received, while preserving a Pedersen commitment to the file hash. We formally define the forwarding relation, construct a Sigma protocol and its Fiat–Shamir-based non-interactive instantiation, and analyze the resulting completeness, soundness, and zero-knowledge guarantees. We evaluate the proposed mechanism using lightweight simulations over a ten–group-head backbone with forwarding paths of varying lengths. Results show that proof generation incurs an average overhead of approximately 0.12 ms per hop, while proof verification averages 0.10 ms per hop, leading to near-linear end-to-end forwarding latency of 0.19 ms (2 hops), 0.29 ms (3 hops), and 0.37 ms (4 hops). In adversarial scenarios, all tampering attempts are detected at the next honest hop, yielding a 0
Suppose a rooted tree T is given which has n nodes. A level ancestor query, denoted LA(v, k), asks for the k-th node on the path from v to the root in T. The level ancestor problem is the problem of preprocessing T into some data structure which can answer these queries. We present the second ever O(n) preprocessing, O(1) query algorithm which does not use Method of Four Russians; the first ever was given by Hagerup [14]. We describe how to parallelize both our and Hagerup’s [14] preprocessing algorithm to run in O(n/P) time on P, P ≤n/log(n) , processors on a CREW PRAM. We give ideas for calculating parallel tree accumulations over one binary operator which does not rely on tree contraction.
Credit card fraud poses a serious challenge to financial institutions due to the increasing volume and complexity of digital transactions. This study investigates how machine learning models—specifically Random Forest and Logistic Regression—can be optimized to minimize false negative predictions, where fraudulent transactions are mistakenly labeled as legitimate. Using a highly imbalanced credit card dataset, the research applied the SMOTE technique to oversample the minority class and conducted experiments by tuning key hyperparameters for each model. Logistic Regression was tested using variations of class weight and probability threshold, while Random Forest was optimized by adjusting the number of trees and maximum tree depth. Model performance was evaluated using accuracy, precision, recall, and F1-score, with particular focus on recall and F1 to reflect the real-world priority of catching fraud without excessive false alarms. Results showed that both models could be effectively tuned to reduce false negatives, with Random Forest yielding higher overall recall and F1-scores. Logistic Regression, though simpler, showed promising adaptability through class weight adjustment. The findings demonstrate that targeted hyperparameter tuning can significantly improve the reliability of machine learning systems in fraud detection tasks. Future work includes exploring additional algorithms and developing automated tuning techniques.
As generative artificial intelligence (AI) becomes embedded in coursework, AI literacy initiatives often emphasize operational competence while underweighting how students narrate AI’s role in learning, authorship, and value. This paper introduces an interpretable natural language processing (NLP) approach for modeling student orientations toward AI in education from reflective essays. Using 48 de-identified undergraduate essays collected across four academic terms, the study combines qualitative coding with term frequency–inverse document frequency (TF–IDF) text representations and unsupervised clustering (k-means clustering and Ward hierarchical clustering). Four recurring narrative orientations—AI Symbiote, Prompt Craftsman, Ethical Guardian, and Resistance Leader—correspond to stable lexical clusters. Results include cluster descriptors and a set of visual diagnostics (principal component analysis (PCA) projection, dendrogram, silhouette analysis, stability across feature caps, and a cluster–archetype overlap heatmap). The paper demonstrates how interpretable text mining can bridge computational modeling and educational research, offering a lightweight taxonomy for AI literacy design and human-centered educational technologies.
In recent years, as the application of blockchain technology has expanded, there has been a growing need for appropriate modeling to support its design and implementation. However, traditional modeling methods have struggled to comprehensively represent the unique dynamic characteristics, decentralization, and concurrency of blockchain, leading to challenges in design efficiency and versatility. In this paper, we propose BSCUM (Blockchain System Construction Unified Model), a unified model based on the primitive elements that constitute a blockchain, taking into account software engineering perspectives. This model effectively represents dynamic characteristics and decentralization, which existing modeling methods could not fully handle, by clearly defining elements such as transactions, hash values, smart contracts, and inter-node communication, and organizing their relationships.
In this study, we examine how well Robust Policy Optimization (RPO) can generalize in comparison to Proximal Policy Optimization (PPO) under environmental perturbations in deep reinforcement learning (deep RL), specifically in continuous control environments simulated using MuJoCo. RL policies often perform poorly when exposed to even minor changes in the dynamics of the environment during testing and are known to overfit to training environments. To address this, we systematically perturb the environment’s gravitational force and agent body mass during both training and testing across MuJoCo environments, namely Hopper, HalfCheetah, and Walker2d environments. Our results show that RPO displays greater robustness compared to PPO across most of our experiments when exposed to test-time perturbations. Additionally, we learn that training with perturbations introduced improved generalization across both algorithms when testing on unseen conditions. Our findings demonstrate that training with unseen conditions in mind can lead to robust RL policies that can better handle real-world variability.
Corneal ectasia represents a group of ocular disorders characterized by progressive biomechanical weakening and thinning of the cornea. While advanced imaging like Optical Coherence Tomography (OCT) has improved diagnostic capabilities, identifying subclinical cases such as forme fruste keratoconus (FFKC) remains a significant challenge in refractive surgery screening. This study aims to develop and evaluate machine learning (ML) models to enhance the detection of corneal ectasia using clinical data from the Anterion® system. A clinical dataset comprising 186 examinations and 88 shared features was used to train and test six supervised learning algorithms, including Decision Tree (DT), Random Forest (RF), Naïve Bayes (NB), Support Vector Classifier (SVC), Logistic Regression (LR) and K-Nearest Neighbors (KNN). To address the limitations of a small clinical sample, a synthetic dataset of 400 cases was generated to assess model generalization. Results on the clinical data showed high performance across all models, with the DT achieving perfect baseline accuracy, though suggesting potential overfitting. Evaluation on the synthetic dataset highlighted the robustness of the SVC (97.50
Overlapping community detection in heterogeneous information networks (HINs) has attracted increasing attention due to the inherent multiplicity of relations and semantic contexts. While recent studies have successfully combined self-supervised heterogeneous graph neural networks (HGNNs) with fuzzy clustering, how the degree and structure of fuzziness vary across different network views remains insufficiently understood. In this paper, we present a systematic analysis of view-dependent fuzziness in overlapping community detection. Using a unified self-supervised HGNN framework, we derive node representations from multiple network views, including schema-based, meta-path-based, and concatenated representations, and apply Fuzzy C-Means clustering under identical experimental conditions. We evaluate fuzziness using complementary metrics, including the fuzzy partition coefficient, membership entropy, and the max–second membership gap. Our results reveal that different views induce distinct and non-trivial fuzziness patterns. The meta-path view achieves lower entropy and larger gap values, indicating well-separated memberships at the node level, whereas the schema view exhibits moderate uncertainty despite comparable global partition quality. In contrast, naive representation integration leads to increased entropy and reduced membership separation. These findings demonstrate that fuzziness is a structured, view-dependent property rather than a clustering artifact. By clarifying how representation choices shape overlapping memberships, this study provides a principled foundation for analyzing and interpreting fuzzy community structures in heterogeneous information networks.
This paper details the creation of a code obfuscation framework centered on Non-Linear Mixed Boolean Arithmetic (MBA) and augmented by additional techniques to maximize mitigation against reverse engineering while preserving runtime correctness. Each obfuscation step is applied incrementally to ensure valid code output, with each step building off of the previous. The obfuscation process involves debug-info stripping, identifier renaming, constant and string obfuscation with runtime reconstruction, MBA algebraic rewrites, junk-code generation, control-flow obfuscation, and optional compilation to bytecode executed on a custom virtual machine. Results indicate that combining MBA with traditional obfuscation methods substantially increases the difficulty for static analysis and de-obfuscation while maintaining acceptable runtime overhead.
Reliable forecasting of water demand plays a key role in promoting sustainable urban development and optimizing utility operations. However, in newly developed or data-sparse districts, the lack of historical consumption data makes model training challenging. This paper explores spatial transfer learning to address this issue, enabling the use of models trained in data-rich regions for forecasting in target areas with no labeled data. Using a real-world data set from ten districts in North-East Italy, we evaluate transfer learning approaches based on data aggregation and ensemble methods. We propose a knowledge-based selection approach for filtering out unsuitable data sources with respect to a given target region in order to reduce the prediction error of the underlying aggregation and ensembling forecasting methods. The results show that the knowledge-based filtering mechanism, which considers basic information such as average annual water inflow, the number of households, or regional information can reduce transfer errors in the aggregation- and ensemble-based models by two-thirds.
The rise of fake accounts on social media platforms, particularly Instagram, poses a significant threat to user safety, platform integrity, and the reliability of social media metrics. This research focuses on detecting fake Instagram accounts using supervised machine learning techniques. Specifically, the research compares the performance of two ensemble supervised learning models, Random Forest and CatBoost, using a balanced dataset of Instagram profiles labeled as real or fake. The dataset included features such as follower count, number of posts, bio length, profile picture status, and external link presence. The research process included data preprocessing, model training, and performance evaluation using metrics such as accuracy, precision, recall, F1-score, and confusion matrices. Both models achieved high performance. While Random Forest showed slightly higher precision, CatBoost outperformed Random Forest in overall accuracy, recall, and F1-score, making it more effective for high-risk applications where identifying fake profiles is critical. Feature correlation analysis revealed that factors such as the presence of a profile picture and the completeness of a profile’s bio were highly indicative of account authenticity. The findings suggest that CatBoost’s boosting-based approach may offer improved effectiveness in real-world cybersecurity applications. This study contributes to ongoing efforts to enhance online safety by providing insight into which machine learning methods are most effective for social media moderation. Future work may include exploring cross-platform generalizability, feature engineering improvements, and evaluating training efficiency.
Artificial Intelligence (AI) is evolving rapidly and modern AI requires devices that recognize and interpret human emotion based on facial expression. One of the first recognized algorithms that makes this possible is called ’Facial Expression Recognition’ (FER). The algorithms used to identify emotions based on facial expressions have been a research focus in the field of psychology for many years. The potential uses of FER are to monitor health, protect people, improve customer experience, and improve the accuracy of affective computing. Unfortunately, even with all of the technological advancements, FER algorithms still suffer from the limitations of the datasets available, the wide variety of ways humans express emotion resources needed to calculate the largest number of classifications possible. As a result, there is a big gap in the need for greater accessibility, efficiency, and generalization of FER algorithms. This research proposal attempts to fill that gap by creating an easily trainable and highly accurate deep learning model using a lightweight convolutional neural network to identify multiple classes of emotions based on the analysis of a single frame of video data or a single image input. Using a methodology that includes systematic image gathering, preprocessing of the dataset, and creating adaptable strategies to handle the differences between different datasets has allowed the creation of a computationally efficient and lightweight model. With this approach, emotion-aware systems can be developed that correctly identify human facial expression emotions, while also being sensitive enough to be deployed in real-world situations.
In this study, we proposed a three-dimensional point-cloud reconstruction method using PointNet autoencoder models guided by semantic segmentation. The objective was to automatically construct virtual spaces that reflected the geometry and object arrangement in real-world environments. Utilizing the large-scale ScanNet++ dataset, scenes were partitioned into semantic classes, and feature extraction was optimized for each specific class to enable high-precision reconstruction of complex indoor scenes. The experimental results demonstrated that the proposed method reproduced the shapes and colors of major structural elements with high fidelity. Notably, this method was particularly effective in completing occluded regions, such as floors hidden by furniture. Although quantitative analysis indicated a performance drop for classes with complex geometries compared to simple structural elements, visualization using uniform manifold approximation and projection and a clustering purity score of 0.972 confirmed that the model successfully learned class-specific geometric features.
This paper investigates the impact of temporal context modeling on sentiment polarity classification of user-generated product reviews using the Microwave Oven Reviews corpus. We first provide a concise linguistic and statistical characterization of the dataset, highlighting substantial variability in review length, lexical diversity, and frequent use of negation and contrastive discourse constructions that challenge order-agnostic lexical models. Building on strong Bag-of-Words (BoW) baselines established in prior work, we evaluate sequence-based sentiment classifiers that combine pretrained fastText embeddings with bidirectional long short-term memory (BiLSTM) encoders. Two embedding sources (Wiki-News and Common Crawl) and multiple embedding adaptation strategies, including frozen, fine-tuned, and freeze–unfreeze training, are examined. Experimental results show that explicit temporal context modeling yields consistent and substantial performance gains over the strongest BoW baselines. While BoW systems achieve equal error rates (EER) in the 11–12
Implementation of hybrid model architecture in Retrieval-Augmented Generation (RAG) systems has increased in recent years where it adopts knowledge graph to retrieve entities, relationships and path alongside vector retrieval which increase the quality of retrieval compared to traditional RAG. The knowledge graphs used in these systems are constructed automatically from data after the appearance of Large Language Models (LLMs). However, in many of these automatically constructed knowledge graphs lack explicit conceptual structure. Most nodes are labelled with generic, surface level names, especially the ones constructed on unstructured texts. This limits the type-aware retrieval of nodes and narrows the potential of the query system. Simultaneously, recent works has shown that ontologies can be generated from unstructured text using Large Language Models and prompt engineering. Yet, their integration with GraphRAG pipelines remained unexplored. This paper proposes a unified framework in which a knowledge graph and an ontology are automatically generated from the same source text and then aligned so that the ontology is used to relabel and enrich the knowledge graph. The enriched graph supports type-aware retrieval, enabling GraphRAG to answer queries even when surface terms do not explicitly appear in the text. We have proposed the method and discussed with a case demonstration how ontology-guided relabeling increases semantic coverage, improves retrieval generality, and strengthens the interpretability of graph-based results.
We study overlapping community detection in heterogeneous information networks (HINs) by combining self-supervised heterogeneous graph neural networks (HGNNs) with fuzzy clustering. Existing self-supervised HGNNs, such as HeCo, are typically evaluated on single-label tasks (node classification, link prediction) and rely on hard clustering (e.g., K-means), which forces each node into exactly one community. This setting is misaligned with real networks, where nodes often belong to multiple communities. We propose a framework that uses a HeCo-style front-end to learn two views of node representations on HINs—schema view and meta-path view—and then applies Fuzzy C-means (FCM) to obtain fuzzy community memberships. To better exploit multi-view information, we standardize each view separately before concatenation, preventing one view from dominating the distance computation in clustering. We analyze not only standard clustering metrics (NMI, ARI) but also the fuzzy partition coefficient (FPC) and the distributions of maximum and second-largest memberships. Experiments on the DBLP heterogeneous bibliographic network show that schema-view embeddings significantly outperform raw features and meta-path-view embeddings in clustering accuracy, and that view-wise standardized concatenation further improves NMI while yielding more overlapping communities (lower FPC, lower max membership, higher second-largest membership). Our results demonstrate that combining self-supervised HGNN embeddings with FCM provides an effective and interpretable way to uncover overlapping community structure in HINs.
