Maintaining solution diversity and adopting an appropriate number of cooperating populations are critical for solving constrained multi-objective optimization problems (CMOPs). Under complex constraints and conflicting objectives, many existing algorithms suffer from premature convergence, especially when feasible regions are sparse or narrowly distributed. Insufficient diversity restricts exploration and hinders the approximation of a well-distributed Pareto front. Moreover, single-population designs often struggle to balance exploration and exploitation in the presence of large infeasible regions, leading to stagnation and poor convergence. To address these challenges, this work proposes a novel evolutionary algorithm termed CMOEA-TSD, which adopts a three-population, two-stage co-evolutionary framework to enhance diversity and coordinate multiple search behaviors. The algorithm integrates three subpopulations: a global search component that approximates the unconstrained front, a diversity-enhanced search that reformulates the problem into a relaxed multi-objective form using reference vectors and a dynamically shrinking constraint threshold, and a feasibility-oriented search focused on constraint satisfaction. A stage-switching mechanism based on the overall population convergence rate adaptively adjusts mating strategies, emphasizing exploration in early stages and exploitation in later ones. Extensive experiments on four benchmark test suites and nine real-world engineering problems demonstrate that CMOEA-TSD significantly outperforms seven state-of-the-art algorithms. Ablation studies further verify the effectiveness of both the co-evolutionary structure and the diversity enhancement mechanisms.
The rapid expansion of Internet of Things (IoT) environments has introduced significant security challenges, particularly due to limited labeled data and increasing vulnerability to adversarial attacks. Traditional intrusion detection systems (IDS) struggle to generalize under such constraints, while existing few-shot learning (FSL) approaches remain sensitive to adversarial perturbations. This work proposes a hybrid intrusion detection framework that integrates prototypical few-shot learning with adversarial training to address both data scarcity and robustness. The proposed model employs a Prototypical Network to learn discriminative embeddings from limited samples and incorporates adversarial optimization into the episodic training process to enhance resilience against gradient-based attacks, including FGSM, C W, DeepFool, and PGD. Experimental evaluation on the IDSIoT2024 dataset demonstrates that the proposed framework achieves high classification performance with only three samples per class, attaining over 93% accuracy under benign conditions. In adversarial settings, performance initially degrades significantly but is restored to 80-90% accuracy after adversarial training under white-box gradient-based attacks. The results confirm that integrating adversarial training within the few-shot learning paradigm improves both generalization and robustness, making the proposed approach suitable for secure and resource-constrained IoT environments.
Skin cancer is one of the most prevalent and lethal cancers globally, making early and reliable diagnosis essential. Typical CNN models struggles to handle the high dimensional dermoscopic data because of its duplicated features, which result in overfitting and diagnosis uncertainty. We introduce an optimization-based embedded feature selection framework driven by a Hybrid (ADAMW + SGD) optimizer that progressively optimizes gradient stability and represses noisy activations for directly incorporating such feature selection into the training of networks. This method is evaluated on ISIC-2020, Malignant–Benign, and PH2 datasets, across EfficientNet-B0, ResNet-50, InceptionV3, and ConvNeXt-Tiny architectures with 5-fold cross-validation. The Hybrid (ADAMW + SGD) optimizer significantly outperforms ADAMW, ADAM, and SGD, achieving 95.4
Time series anomaly detection is crucial for industrial monitoring and cybersecurity. However, existing approaches still struggle with multi-structured signals such as trend drifts, high-frequency perturbations, and periodic anomalies, owing to feature coupling and limited joint modeling across time and frequency domains. To overcome these challenges, we propose Deployment Oriented Dual Path Multi View Anomaly Detection (DP-MVAD), a deployment-efficient framework designed for complex non-stationary conditions. DP-MVAD achieves trend-perturbation disentanglement through a conditionally identifiable dual-path approximation, while a multi-view gated fusion mechanism across temporal, spectral, and wavelet domains ensures robust cross-scenario representations. Furthermore, an asymmetric local band attention module combined with a sliding-window strategy enhances the early detection of weak short-term anomalies. Experiments on multiple public benchmarks show that DP-MVAD surpasses state-of-the-art methods by 2.45
Graph partitioning is a critical enabler of scalability in distributed graph databases, impacting load balancing, communication overhead, and query performance. This paper presents a comparative analysis of four partitioning strategies Random Vertex, Metis-based, Random Edge, and HDRF spanning both edge-cut and vertex-cut paradigms. Each approach is evaluated on synthetic and real-world datasets for metrics such as edge cut, replication factor, load balance, latency, and throughput. Our evaluation framework simulates distributed OLTP/OLAP query workloads on a prototype graph database cluster. Results show that while Metis achieves optimal partition quality for static, balanced graphs, HDRF provides superior performance and scalability for dynamic, power-law networks. The experiments reveal important trade-offs between partitioning complexity and runtime performance across different graph structures. This work contributes a scalable benchmarking environment, a set of representative partitioning algorithms, and statistically grounded insights to guide practitioners in choosing the appropriate strategy for specific workload and graph characteristics.
With the increasing penetration of renewable energy, the complexity and uncertainty of power grid operation have significantly increased. Traditional transformer fault diagnosis techniques can no longer meet the stringent reliability requirements of modern power systems. To address this issue, this paper establishes a mapping relationship between power grid stability indicators and transformer fault states through physical mechanism analysis, and constructs an efficient and accurate transformer fault diagnosis and prediction model to meet the challenges of power grid security in complex environments. Based on a bidirectional long short-term memory network (BiLSTM), this paper integrates a convolutional neural network (CNN) to enhance local spatial feature extraction and introduces an attention mechanism to strengthen the information response at key time steps. A CNN-BiLSTM-Attention hybrid structure algorithm is designed to achieve deep representation and modeling of the multidimensional dynamic characteristics of the power grid. To further improve model performance, this paper proposes an Improved Antlion Optimization (IALO) algorithm. This algorithm introduces elite retention, adaptive step size, local perturbation, and diversity protection mechanisms, and has a clear parameter basis, constructing an efficient network parameter optimization strategy. Repeated experiments showed that, compared with the baseline model, the IALO-optimized CNN-BiLSTM-Attention model improved accuracy by 2.96
Wind turbine blade defects — including fine cracks, erosion, and delamination — pose critical safety and efficiency risks. However, existing segmentation methods struggle with multi-scale damage variability, irregular morphologies, and indistinct boundaries. To address these challenges, we propose CHS-Net, a Collaborative Hybrid Segmentation Network, as a task-driven segmentation framework that redesigns the conventional U-Net pipeline into a stage-wise collaborative architecture for wind turbine blade defect segmentation. Specifically, Contextual Multi-Scale Semantic Modeling (COSM) module is embedded in the encoder for depth-adaptive multi-scale perception, CNN-ViT Interaction (CIT) module is placed at the bottleneck for computationally efficient global context modeling, Twin Channel-Spatial Squeeze and Excitation (TCSE) module is introduced in the decoder for concurrent channel-spatial refinement, and Dual-path Attention Topology (DAT) module replaces the conventional output head for boundary-aware topological regularization. When evaluated on 2,250 annotated UAV images, CHS-Net achieves an IoU of 69.94 ± 0.11
In the digital age, online businesses often rely on advertising as a primary source of revenue. Many services and platforms depend on advertisement income to remain operational, allowing free services to be sustained through ad funding. Advertisers typically pay per click, with fees determined by click-through rates. However, a major challenge in this system is click fraud, a deceptive practice in which advertisements are repeatedly clicked without genuine interest or intent to make a purchase. This behavior leads to financial losses for advertisers, damages platform reputations, and distorts the effectiveness of online advertising. Detecting fraudulent clicks remains a complex challenge due to the constantly evolving tactics employed by fraudsters. This study explores the potential of artificial intelligence (AI) classifiers for identifying instances of click fraud. Because the performance of AI-based classifiers strongly depends on proper hyperparameter selection, a modified optimization metaheuristic is introduced to improve classification accuracy. A comparative analysis is conducted using real-world data, demonstrating promising results from the best-performing models with accuracy as high as 0.786496. Finally, the legal and policy implications of such a system are examined within the context of an evolving digital landscape.
This paper propose a secure and robust color-image watermarking framework for open-network environments in the Discrete Quaternionic Quadratic Phase Fourier Transform (DQQPFT) domain. Unlike RGB-separable schemes with fixed kernels, the method exploits a controllable non-rigid DQQPFT kernel to preserve inter-channel correlation and provide flexible quaternion frequency representations. The kernel parameters are incorporated as cryptographic keys to enhance key sensitivity and resist brute-force guessing. To reduce computational cost, DQQPFT is implemented via a real-matrix equivalent representation of quaternions while remaining consistent with the original transform. For adaptive watermark embedding in Quaternion Singular Value Decomposition (QSVD), we formulate a bi-objective optimization that jointly considers imperceptibility and robustness. An Enhanced Multi-Objective Ant Colony Optimization algorithm (E-MOACO) is developed by introducing dynamic weight scheduling and adaptive Gaussian perturbation, and a knee-point selection strategy is used to obtain well-balanced solutions. Extensive experiments demonstrate high visual quality and reliable detection: average PSNR exceeds 39 dB without attacks; average normalized correlation (NC) remains above 0.99 under common single attacks and above 0.90 under mixed attacks.
In cross-enterprise supply chain collaboration scenarios, transaction data such as transaction amounts and cross-enterprise exchange rates all possess rational number properties. However, existing approaches to processing rational numbers in secure computing involve scaling by the same multiple, which may result in loss of rational number precision. Any loss of precision or simplistic conversion of rational numbers to integers may pose risks of privacy leaks and erroneous judgments. To address the aforementioned issues, this paper studies the application of secure computation based on the intersection and union of rational number sets in cross-enterprise supply chain collaboration. this paper transforms rational numbers into matrices and designs secure computation protocols for set intersection and union under semi-honest and malicious models using the Elgamal homomorphic encryption algorithm. The security of the protocol under the malicious model is proven using an ideal-practical example. Finally, this paper analyzes the efficiency of existing protocols versus the proposed protocol, and proves its practicality.
Collaborative editing, as observed in projects such as Google Docs, often relies on costly central services. CRDTs are proposed as decentralized data structures for consistent collaborative editing. Nonetheless, to date no CRDT proposal supports a closed group of peers while both preventing interleavings and rejecting operations from distrusted nodes outside that group. We present the Syncordian CRDT, which guarantees strong eventual consistency and intention preservation across nodes. Syncordian avoids interleaving anomalies by detecting overlapping edit blocks and re-issuing the lower priority one, and it authenticates every operation with a keyed message authentication code rooted in a secret shared by the peers admitted when the document is created, so it needs no public key infrastructure. To evaluate Syncordian we use two collaborative documents taken from GitHub, with over 1000 edit operations contributed by over 100 network peers, together with a suite of controlled workloads that build the concurrent authoring a replayed history cannot. We validate that Syncordian keeps every peer’s document consistent, even in the presence of distrusted nodes, and we report what the guarantees cost in memory, in messages, and in bytes, and how each cost grows with the number of peers. Syncordian authenticates group membership rather than individual peer identity, so its guarantees are against nodes outside the admitted group, and a compromised member is out of scope.
Nowadays, Mobile Crowdsensing (MCS) is a crucial Internet of Things (IoT) paradigm leveraging mobile sensors for large-scale data collection. Dynamic task allocation remains a key challenge in real-world deployments. This study introduces Multi-Objective Fractal-based Artificial Rabbit Optimization (MOFARO), a multi-objective model within the Artificial Rabbit Optimization Algorithm for task allocation in edge-fog-cloud MCS environments. Tasks are modeled as Directed Acyclic Graphs (DAGs), incorporating Markov chain mobility predictions and experimental constraints to minimize execution time, energy consumption, cost, and Quality of Service (QoS) violations. Simulations on established workflows yield a Pareto front hypervolume of 0.85. MOFARO reduces execution time by 12–25
Underwater acoustic sensor networks (UASNs) are increasingly employed in both military and civilian applications. Owing to the challenging nature of underwater communication—characterized by unstable channels and dynamic environmental factors— UASNs are vulnerable to various security threats. Therefore, designing robust and adaptive trust management frameworks is vital to ensure their secure and stable operation. This paper introduces a novel, smart, and fast, fuzzy based-trust management (SFFTM) scheme in UASNs. In SFFTM, trust values are derived from three types of trust evidence: communication-based, energy-based and, data-based. To refine communication-based trust evidence, SFFTM distinguishes between communication failures caused by malicious nodes and those resulting from channel instability. To achieve this, a new control factor— combining link delay and link usage— is introduced. When failures are caused by malicious nodes, this factor exponentially reduces the trust evidence. Conversely, if the failures result from instability, the control factor enhances the trust evidence. Furthermore, the energy-based trust evidence is derived by analyzing the residual energy of each nod and evaluating the correlation between its actual and expected energy consumption rates using the Pearson correlation coefficient (PCC). SFFTM includes four trust elements: direct trust, recommendation trust, integrated trust, and indirect trust. Direct trust is obtained from a fuzzy inference system that processes the three trust evidences. Recommendation trust is determined by evaluating the honesty degree of recommenders and assigning higher weights to those with higher honesty level. Indirect trust is inferred through a multi-hop trust chain that links the evaluator to the target node, where intermediate recommenders are chosen according to both their spatial proximity and their assessed trustworthiness. The performance of SFFTM is evaluated through extensive simulations and compared with three existing approaches— CFFTM, LTrust, and TMC— under three attack scenarios: bad/good-mouthing, collision, and hybrid attacks. Simulation findings show that SFFTM outperforms the existing techniques by enhancing detection accuracy and significantly reducing the false alarm rate. Specifically, in the first scenario, SFFTM improves the indirect trust of legitimate nodes by 2.58
The use of data mining and artificial intelligence (AI) in educational contexts has attracted increasing attention from researchers over the past decade. However, extracting meaningful insights and actionable patterns from the large volume of student data and activities available in learning management systems (LMSs) remains challenging due to the diversity of course designs and engagement behaviors. Furthermore, differences in how students interact with learning management systems can significantly affect the predictive value of digital engagement metrics. This study investigates the relationship between Moodle engagement and students’ academic performance across three undergraduate courses with differing pedagogical orientations: Research Methods, Arabic Language, and Palestinian Studies. Using a multi-level quantitative approach that combined correlation analysis and predictive modeling, the study assessed how temporal patterns, activity types, and departmental contexts (STEM-oriented vs. humanities-based) shaped the explanatory power of digital engagement data. The results revealed that Moodle activity is a strong predictor of success in assessment-driven courses, such as Research Methods, where quiz- and assignment-related interactions were highly correlated with final grades, and machine learning models achieved predictive accuracy close to 97
Blockchain systems use distributed consensus to establish decentralized trust, yet the proliferation of protocol designs and adversarial capabilities makes it difficult to compare security guarantees and engineering trade-offs in a unified way. This survey reviews blockchain consensus from a security-oriented lens and links distributed-systems fundamentals to concrete attack vectors and defense mechanisms. We summarize the key assumptions that delimit what consensus can guarantee in permissionless settings, including network synchrony and partition models, crash and Byzantine fault models, honest-majority and quorum thresholds, and the safety–liveness tension under partial synchrony. We then introduce a threat model that classifies attacks across the networking layer, the consensus and finality layer, and the incentive layer, and apply consistent finality semantics to analyze representative protocols spanning crash-tolerant and Byzantine-fault-tolerant families, Nakamoto-style longest-chain consensus, proof-of-stake variants, DAG-based BFT, Proof-of-X constructions, and hybrid designs. We distill recurring adversarial patterns, mitigation strategies, and the resulting security–performance trade-offs. Finally, we outline open problems including the post-quantum cryptographic transition, cross-shard security and atomicity in sharded consensus, and adaptive defenses resilient to evolving network and economic adversaries.
The metaverse has recently gained traction as a persistent, realistic, and interconnected virtual-physical ecosystem with high levels of realism and immersivity. According to such a definition, it fundamentally relies on the generation of high-fidelity 3D scenes. Manual construction of these scenes remains a time-consuming process that requires advanced technical skills, posing a significant barrier to widespread adoption. Recent advancements in AI-assisted photogrammetry tools offer an alternative approach by enabling semi-automatic 3D reconstruction from real-world environments. Nevertheless, the usability, quality, and real-time applicability of the resulting assets remain underexplored. This study aims to assess the effectiveness of photogrammetry-based tools for creating functional and visually appealing 3D scenes for metaverse applications. To this end, a comparative experiment was conducted by reconstructing three environments using manual modeling and two photogrammetry tools based on distinct technologies: Polycam (MVS-based) and LumaAI (NeRF-based). The resulting models were evaluated using quantitative metrics, including modeling time, polygon count, visual similarity (SSIM), real-time performance (GPU usage and FPS), and Hausdorff distance. The findings revealed that photogrammetry significantly accelerates the modeling process but does not consistently outperform manual modeling in terms of polygonal optimization and real-time rendering efficiency. Furthermore, the choice between photogrammetry tools depends on object characteristics and application constraints. This study provides practical insights and empirical guidelines for developers and researchers, highlighting the trade-offs between automation, performance, and visual fidelity in 3D scene generation for the metaverse.
The package osESD in Python and R implements online sequential extreme studentized deviate tests, or osESD in short, a statistical testing method for anomaly detection in online streaming datasets with varying patterns. Its foundation lies in the general extreme studentized deviate test, or GESD test. The GESD test, a simple and effective anomaly detection test, lacks online learning and shows poor performance for pattern-varying data. Our proposed osESD algorithm uses sequential ESD, a variation of GESD, which implements online learning with change-rate trend vectors and residual trend vectors to deal with the problem of pattern-varying streaming data. We also automated parameter tuning in the package with multiple options the user can select for immediate deployment and use. We show that the model learning of our implementation is fast and the performance is excellent compared with numerous models with both various synthetic datasets and real datasets. Codes can be found in the PyPI [11].
As the scale of big data continues to expand, traditional single-machine processing methods have been replaced by distributed computing frameworks. With the limited potential for further CPU performance improvements, GPU-accelerated computing has emerged as a trend, particularly for single-instruction multiple-data (SIMD) and computationally intensive tasks. However, with the clustering of GPUs, traditional scheduling algorithms face limitations such as resource waste and inefficient allocation. To meet the diverse needs of big data applications, it is necessary to optimize the task scheduling and architecture of computing frameworks to enhance processing efficiency and adaptability. Addressing these challenges, this paper proposes a novel distributed computing framework, GDSF, based on a thorough examination of existing computing frameworks and task scheduling algorithms, combined with the computational characteristics of GPUs and the mechanism for multi-task GPU sharing. This framework incorporates various task scheduling strategies to ensure effective scheduling for both dependent and independent task models, achieving an optimal match between tasks and GPU resources. Experiments conducted on a real heterogeneous CPU-GPU cluster using diverse GPU workloads demonstrate that the performance of the GDSF computing framework and its scheduling algorithms meets the expected research objectives, thus validating the feasibility and effectiveness of the proposed design.
In solving constrained multi-objective optimization problems (CMOPs), not only is the optimization of the population in the objective space important, but the solutions in the decision space also play a crucial role. However, current constrained multi-objective evolutionary algorithms (CMOEAs) often neglect the impact of the solution generation process on algorithm performance. Therefore, this paper proposes an information-assisted solution generation based dual population constrained multi-objective evolutionary algorithm, called GSCMO. The information includes entropy information reflecting the population distribution and neighborhood information reflecting the proximity relationships between solutions. In GSCMO, we design a global solution generation strategy for the main population based on entropy information. This strategy dynamically optimizes the mutation strength, focusing on enhancing solution diversity and maintaining extensive exploration of the population in the decision space, thereby improving the algorithm’s ability to explore potential feasible solution regions. At the same time, we develop a local solution generation strategy for the auxiliary population based on neighborhood information. This strategy uses Manhattan distance to determine the neighborhood of each individual, selects individuals from the neighborhood as parents, and generates high-quality individuals, thus enhancing the algorithm’s local search capability. Extensive experimental results demonstrate that GSCMO has high competitiveness compared with some state-of-the-art methods.
In the research study, we investigated VoteChain, an energy efficient and secure framework for financial transactions. The proposed model integrates a lightweight CNN–LSTM hybrid anomaly detection model into the consensus layer. The design is an intelligent based validation during transactions, a preemptive identification of fraudulent behavior during block finality. we evaluated system under simulated permission blockchain network through benchmark dataset. The framework reduces energy and achieves 20.0 mWh in a transaction. The anomaly models’ energy metrics profile (CNN_LSTM) distributed as inference at 38
