Array最新文献

{"title":"Enhancing the robustness of CNN-based lung cancer detection models against label-flipping poison attacks using defensive distillation","authors":"Ahmeed Yinusa ,&nbsp;Misa Faezipour","doi":"10.1016/j.array.2025.100637","DOIUrl":"10.1016/j.array.2025.100637","url":null,"abstract":"<div><div>Deep learning has significantly advanced automated medical imaging, particularly in lung cancer detection, yet vulnerability to adversarial manipulation continues to limit clinical reliability. This study investigates the impact of a 30% random-uniform label-flipping poisoning attack on Convolutional Neural Network (CNN) models trained using the IQ-OTH/NCCD dataset comprising 1190 CT images. A multi-strategy defense pipeline is proposed, integrating defensive distillation, Isolation Forest data sanitization, and a noise-tolerant loss function to enhance robustness against training data corruption. To ensure a valid evaluation framework, Synthetic Minority Oversampling Technique (SMOTE) was applied only to the clean training subset before any poisoning was introduced, preventing the propagation of corrupted labels and establishing a balanced and uncontaminated foundation for teacher training. A high-accuracy teacher model trained on this clean SMOTE-balanced dataset produces temperature-scaled soft labels that guide the Student model. The Student is then trained on a sanitized dataset filtered to remove anomalous or inconsistent samples, using Symmetric Cross-Entropy loss to reduce sensitivity to mislabeled data. Experimental results show that the pipeline maintains strong performance, achieving 99% accuracy on clean data and 95 to 96% accuracy under poisoning, while preserving stable precision, recall, and confidence calibration across all classes. These findings demonstrate that the proposed strategy effectively mitigates label-flipping poisoning, offering a reproducible path toward secure and trustworthy AI systems for medical imaging applications.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100637"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145735330","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Design and implementation of a scalable microservice-based cloud computing service for stochastic fluid flow modeling","authors":"Zhanars Abdiramanov","doi":"10.1016/j.array.2026.100679","DOIUrl":"10.1016/j.array.2026.100679","url":null,"abstract":"<div><div>This paper presents the design and performance evaluation of a scalable microservice-based cloud computing framework for distributed stochastic fluid flow simulations. The system integrates Docker containerization, NGINX load balancing, and Apache Kafka for asynchronous task coordination and efficient workload distribution across computational nodes. The backend combines PHP and Julia for orchestration and computation, with PostgreSQL managing task and result data. Benchmark experiments demonstrate near-linear scalability and stable performance under varying loads, confirming the system’s suitability for high-performance scientific computing. The proposed framework advances parallel and distributed computing by introducing a modular, reproducible architecture for executing complex stochastic models in cloud environments.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100679"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145973365","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A novel nature inspired feature optimization algorithm to detect breast cancer using machine learning","authors":"Md Anisur Rahman,&nbsp;M. Rubaiyat Hossain Mondal","doi":"10.1016/j.array.2026.100690","DOIUrl":"10.1016/j.array.2026.100690","url":null,"abstract":"<div><div>Feature optimization (FO) is a machine learning (ML) preprocessing step that improves performance by reducing the number of features. However, FO struggles with high-dimensional medical data, leading to overfitting. In this paper, we propose Fire-Gray Optimization (FGO), a nature-inspired FO algorithm designed to obtain an optimal feature subset and overcome single FO limitations. The proposed method capitalizes on the rapid flashing behavior of fireflies and the predatory habits of gray wolves. By simulating the firefly's flashing behavior and the gray wolf's social hierarchy sequentially, this FO approach iteratively optimizes features using a dataset, providing improved and balanced exploration and exploitation capabilities for resolving challenging optimization problems. Each algorithm refines the shortcomings of others and adds a hierarchical role with enhanced global and local search capabilities. The proposed technique has been evaluated on a well-known medical dataset, the Wisconsin Breast Cancer Dataset (WBCD), which has 32 features from 569 samples. Data in WBCD is classified using six learning methods: voting classifier (soft and hard voting), decision tree, logistic regression, random forest, extreme gradient boosting, and multilayer perceptron. The experimental outcome illustrates that the proposed approach outperformed various prior works on the WBCD by achieving a detection accuracy of 98.24 % for the voting classifier with 16 features. The same optimization process was employed on another breast cancer dataset (METABRIC) to predict 5-year survival status with six similar ML algorithms. FGO-optimized data achieved 95.96 % accuracy for the voting classifier with 16 features. In summary, FGO shows promise for high-dimensional biomedical classification.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100690"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146184698","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Hierarchical conditional forests for meta-frontier analysis under input uncertainty","authors":"Yu Zhao","doi":"10.1016/j.array.2026.100708","DOIUrl":"10.1016/j.array.2026.100708","url":null,"abstract":"<div><div>Conventional meta-frontier benchmarking with Data Envelopment Analysis (DEA) treats observed inputs as exact, even though measurement is often noisy. We propose a decision-centric alternative that learns the distribution of inputs given outputs and group labels via a hierarchical conditional forest, then propagates input uncertainty by drawing nonnegative pseudo-inputs and recomputing group and meta frontiers across replications. The forest shares splits across groups and partially pools leafwise moments, stabilizing small or heterogeneous groups without imposing a common inefficiency law. This yields level uncertainty: interpretable intervals and probabilities for efficiency distances and technology-gap ratios that reflect realistic re-measurement of inputs. We emphasize that these intervals represent conditional predictive distributions under the learned input law, conceptually distinct from Bayesian posterior intervals for latent parameters or frequentist confidence intervals for population quantities. We establish measurability and Monte Carlo consistency, validate on synthetic data, and apply the method to the 2019 Penn World Table. Compared with single-shot benchmarking, our approach delivers more conservative gap assessments, particularly for the pooled middle-income group, and expresses uncertainty in a format that is directly usable for policy thresholds and cross-group comparisons.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100708"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146184900","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Vehicle-Scale LLMs: Integrating low-rank residuals and 4-bit quantization for in-vehicle AI","authors":"Anubha Parashar ,&nbsp;Apoorva Parashar ,&nbsp;Bhavya Joshi ,&nbsp;Kavita Jhajharia ,&nbsp;Aditya Sinha","doi":"10.1016/j.array.2026.100709","DOIUrl":"10.1016/j.array.2026.100709","url":null,"abstract":"<div><div>The proliferation of large language models (LLMs) has enabled transformative advances in in-vehicle conversational artificial intelligence (AI), traffic forecasting, and natural language navigation. However, deploying LLMs on automotive edge devices faces stringent memory and compute constraints, as modern vehicle processors typically provide less than 16 GB of dedicated memory for AI workloads. We present Dynamic Adapter Compensation + 4-Bit Quantization for Intelligent Transportation Systems (DAC＋Q4-ITS), a fully training-free compression framework that fuses zero-initialized low-rank adapter compensation with on-the-fly 4-bit dynamic quantization. Our approach begins by collecting a compact in-vehicle corpus comprised of diverse driving-related utterances, traffic reports, and route-planning dialogues. We compute layerwise activation eigenspaces via singular value decomposition (SVD) from this calibration data and project the quantization-induced weight perturbations onto the top eight eigenvectors per transformer layer. This process yields rank-8 adapter compensation modules that are injected directly into each layer without any gradient updates, thereby recovering accuracy lost to aggressive 4-bit quantization. During inference, all linear projections are executed in 4-bit integer format – achieving a fourfold reduction in weight storage – while the corrective adapter paths remain in 32-bit floating-point (float-32). We demonstrate that a 1 billion-parameter (1 B-parameter) LLaMA-derived base model can operate on commodity automotive hardware (e.g., NVIDIA Graphics Processing Unit (GPU) Jetson Xavier NX with 16 GB Low-Power Double Data Rate 4X (LPDDR4x) memory), incurring only a 1%–3% latency overhead relative to an 8-bit baseline. On downstream tasks – natural language navigation (NLN), conversational route planning (CRP), and traffic forecasting (TF) – it achieves over 98% of full-precision performance. By enabling low-latency, privacy-preserving, on-device LLM inference under strict memory budgets, DAC＋Q4-ITS establishes a practical pathway for embedding advanced conversational and predictive AI directly in intelligent transportation systems, eliminating reliance on cloud connectivity and enhancing robustness, reliability, and user privacy.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100709"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147395519","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automated quality assessment of online medical science popularization Videos: A framework based on large language models","authors":"Shiqi Zhou ,&nbsp;Hua Wu ,&nbsp;Yongjian Zhang ,&nbsp;Qianqian Zhong ,&nbsp;Yueqin Diao ,&nbsp;Huihui Fang ,&nbsp;Yanwu Xu ,&nbsp;Hanyi Yu","doi":"10.1016/j.array.2026.100717","DOIUrl":"10.1016/j.array.2026.100717","url":null,"abstract":"<div><h3>Background</h3><div>Unverified information in online medical science popularization videos poses significant public health risks. In the era of user-generated content, the sheer volume of daily video uploads overwhelms conventional human moderation, necessitating scalable automated assessment technologies. This study aims to develop and validate a deterministic large language model (LLM) framework for automated assessment of online medical science popularization videos.</div></div><div><h3>Methods</h3><div>A dataset of 121 Chinese dietary weight loss videos was curated from TikTok, Bilibili, and Kwai. Audio tracks were transcribed using OpenAI's Whisper-1 toolkit, excluding videos with excessive background noise or no spoken content. Videos were annotated by physicians using mDISCERN, a validated tool assessing clarity, reliability, fairness, references, and rigor. Thirteen LLMs—spanning general-purpose, medically tuned, and Chinese-optimized architectures—were evaluated via a novel likelihood selection strategy. This approach converted mDISCERN scoring into binary classification tasks, calculating answer likelihoods in a single forward pass to ensure deterministic outputs. Performance was measured using accuracy, F1 score, and Matthews correlation coefficient, with statistical significance assessed via paired Chi-square tests.</div></div><div><h3>Results</h3><div>Llama3-based models (Meta-Llama-3-8B-Instruct and Llama3-8B-Chinese-Chat) achieved the highest macro-average accuracy (79.50%), closely matching expert ratings in clarity (95.87%), references (97.52%), and rigor (99.17%). However, fairness assessment revealed systemic limitations, with 12/13 models scoring 19.83% accuracy, except Deepseek-R1-Distill-Llama-70B (70.25%).</div></div><div><h3>Conclusion</h3><div>Our study validates the potential of LLM-driven quality assessment for weight loss videos. Performance comparison emphasizes architecture over scale or medical specialization for reliable assessments. By open-sourcing the pipeline, we provide a foundational framework that holds promise for broader application in medical science popularization.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100717"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396116","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross domain patent retrieval","authors":"Iliass Ayaou,&nbsp;Denis Cavallucci,&nbsp;Hicham Chibane","doi":"10.1016/j.array.2026.100720","DOIUrl":"10.1016/j.array.2026.100720","url":null,"abstract":"<div><div>Patent prior-art retrieval becomes especially challenging when relevant disclosures cross technological boundaries. Existing benchmarks lack explicit domain partitions, making it difficult to assess how retrieval systems cope with such shifts. We introduce <span>DAPFAM</span>, a family-level benchmark with explicit <strong>IN</strong>-domain and <strong>OUT</strong>-domain partitions defined by a new IPC3 overlap scheme. The dataset contains 1247 query families and 45,336 target families aggregated at the family level to reduce international redundancy, with citation-based relevance judgments. We conduct 249 controlled experiments spanning lexical (BM25) and dense (transformer) backends, document- and passage-level retrieval, multiple query and document representations, aggregation strategies, and hybrid fusion via Reciprocal Rank Fusion (RRF). Results reveal a pronounced domain gap: <strong>OUT-domain performance remains roughly five times lower</strong> than IN-domain across all configurations. Passage-level retrieval consistently outperforms document-level, and dense methods provide modest gains over BM25, but none close the OUT-domain gap. Document-level RRF yields strong effectiveness–efficiency trade-offs with minimal overhead. By exposing the persistent challenge of cross-domain retrieval, <span>DAPFAM</span> provides a reproducible, compute-aware testbed for developing more robust patent IR systems. The dataset is publicly available on huggingface at <span><span>this repository</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100720"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396119","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Design of a GenAI UX layer with small language models for edge computing in smart agriculture","authors":"Juan M. Núñez V. ,&nbsp;Carlos Alberto Peláez ,&nbsp;Andrés Solano ,&nbsp;Juan M. Corchado ,&nbsp;Fernando De la Prieta","doi":"10.1016/j.array.2025.100632","DOIUrl":"10.1016/j.array.2025.100632","url":null,"abstract":"<div><div>This study proposes a framework for designing a GenAI-UX layer in edge computing systems, integrating user experience (UX) principles and small language models (SLM) to optimize human–machine interaction in agricultural environments. Applied to indoor coriander cultivation, the enhanced UX design reduced the harvest cycle from 45 to 32 days by optimizing irrigation and environmental conditions through a voice-controlled interface with audio generation. The implementation of a structured UX approach enabled more efficient interaction with the system, facilitating data-driven decision-making processed at the edge. The study highlights the importance of UX design guidelines aligned with system requirements to ensure accessibility, intuitiveness, and efficiency in embedded environments. The results demonstrate that using SLM in edge systems improves personalization and response speed, enhancing agricultural productivity. This work underscores the need for a reference framework for UX design in embedded systems, ensuring effective interactions and a greater impact on the efficiency of controlled crop environments.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100632"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145684010","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Data-driven risk mitigation and flexibility enhancement in perishable supply chain networks","authors":"Mohammad Mehdi Tahmouresi,&nbsp;Javad Behnamian","doi":"10.1016/j.array.2025.100641","DOIUrl":"10.1016/j.array.2025.100641","url":null,"abstract":"<div><div>Supply chains operating in dynamic, high-risk environments need mechanisms that preserve economic efficiency while strengthening resilience to disruptions. This study develops an integrated, data-driven framework for four-echelon supply chains (suppliers, manufacturers, distributors, retailers) that explicitly models route disruptions, transportation capacity limits, product perishability, and production and storage constraints for finished goods and raw materials, together with uncertain customer demand. By considering these factors simultaneously, the model shows that production and storage limits critically influence system performance and that demand uncertainty increases operational complexity and cost. To this end, the study formulates a bi-objective, linearized Mixed-Integer Programming (MIP) model that minimizes (<em>i</em>) overall operational cost and (<em>ii</em>) network inflexibility, measured by the number of critical points and allocation counts, thereby capturing trade-offs among efficiency, risk mitigation, and flexibility. To address practical Just-In-Time (JIT) shortcomings under uncertainty, the model allows multi-sourcing (distributors can source from multiple manufacturers; manufacturers can procure from multiple suppliers), enhancing robustness relative to conventional configurations. Uncertainty is treated with a data-driven Distributionally Robust Optimization (DRO) approach. The model is solved with exact CPLEX routines via the augmented ε-constraint method for moderate-sized instances, and with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) for large instances. Performance is benchmarked against a Multi-Objective Particle Swarm Optimization (MOPSO) comparator. Computational experiments—conducted on datasets supplied by Khoshgovar Company (soft-drink production and distribution)—demonstrate that NSGA-II yields superior Pareto fronts (proximity, dispersion, objective attainment) while enabling tractable solutions at practical scales. Results further indicate that risk-averse strategies, although costlier in the short term, materially improve long-term resilience by lowering disruption impact and systemic exposure. The integrated framework advances theory by bridging resilience, JIT, and robust data-driven planning, and offers actionable managerial guidance for industries handling perishable goods (e.g., food and cold-chain pharmaceuticals), including strategies to balance cost and flexibility under uncertainty.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100641"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145788383","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"JPEG-compression agnostic identification of generative art using explainable spatial domain features","authors":"Vrinda Kohli ,&nbsp;Janmey Shukla ,&nbsp;Harish Sharma ,&nbsp;Narendra Khatri","doi":"10.1016/j.array.2025.100635","DOIUrl":"10.1016/j.array.2025.100635","url":null,"abstract":"<div><div>In recent years, the art landscape has undergone a considerable transformation with the emergence of AI-powered generative art tools, challenging traditional notions of artistic authenticity and ownership. The exponential growth of generative artwork sharing on social media platforms has created an urgent need to protect artists' intellectual properties from impersonation, forgery, and style appropriation. This study introduces an innovative, lightweight detection framework that efficiently distinguishes AI-generated art from human-created artwork by analyzing spatial domain features using tree-based ensembles. The study focuses on two prominent generative image architectures, StyleGAN2-ADA and Stable Diffusion, to explore the method's effectiveness across various classes of probabilistic deep generative models while incorporating JPEG compression considerations to reflect real-world social media conditions. The framework was evaluated across a diverse dataset of 10,000 images, achieving a detection accuracy of 94.43 % for StyleGAN2-ADA and 97.97 % for Stable Diffusion outputs on average across varying quality factors (QF). A key limitation observed is the lack of cross-architecture generalization-classifiers trained on one generative model do not reliably detect outputs from others, highlighting the need for architecture-agnostic detection strategies for real-world deployment. These results demonstrate comparable or better performance to existing deep learning solutions, requiring significantly less computational resources and training data. The proposed approach represents a significant step towards digital art authentication, offering a practical solution for real-time detection of AI-generated artwork in social media environments. Future work will focus on expanding the framework's capabilities to address emerging generative models and developing and integrating tools for automatic art authentication across various social media platforms.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"29 ","pages":"Article 100635"},"PeriodicalIF":4.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145788384","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}