Early and sensitive identification of the lung cancer biomarkers is essential towards the process of enhancing the patient outcomes. Traditional plasmonic biosensors SP, LSPR, and SERS, provide sensitive detection of other biosensors, but spectral overlap, interference, and poor reproducibility in multi-biomarkers. Here, the plasmonic biosensing framework is augmented with AI and uses a convolutional neural network (CNN) with the sensors to facilitate real-time, automated deconvolution of overlapping resonance signals. CNN model was minimized to use convoluted layers of 4-6, ReLU activation, and 3x3 or 5x5 kernels, where the return accuracy in the prediction was up to 97.2 percent and a mean squared error was 0.0015 with single- and multi-biomarker spectra. It achieved 3-4x better signal-to-interference ratio, better resonance shift errors (0.012 - 0.015 nm) and a processing latency of less than 1s/sample compared to other integrations with SPR and LSPR sensors. Synthetic and clinical sample confirmation showed that the classification accuracy is over 95 percent, mean absolute error is less than 0.018 nm and reproducible variance smaller than 0.02 nm. Scalability test mains ensured that steady performance was achieved even with seven biomarkers and a variety of sensor types with generalization error of less than 0.002. These findings confirm that the proposed AI-assisted plasmonic system is a robust, sensitive, and fast platform of multi-biomarker lung cancer detection, which can be applied to clinical and pointof- care uses.
This study describes the design and creation of an intelligent, deep-learning-based evaluation of the sight-singing and ear-training assessment, which is an area of a longstanding gap in music education where evaluation has been based on human subjectivity and analysis of one skill at a time. The suggested system is a combination of four fundamental dimensions of musicianship, namely pitch accuracy, rhythmic precision, interval recognition, and solfege articulation, on a single multimodal fusion framework using convolutional neural networks, Bi LSTM temporal models, Transformer encoders, and spectral-temporal feature alignment. The system was trained and tested on a large, multi-level data consisting of beginner to expert vocal performances to be trained and tested on a wide scale of singing styles, skill patterns, and acoustic conditions. Experimental findings indicate high performance improvement against state-of-the-art models with a 14.2 cent pitch error, 28 ms rhythmic onset error, 92 percent interval accuracy and 90 percent solfege rating and a score correlation of 0.94 with expert human raters. Not only does the system provide high precision scoring, it also produces diagnostic as well as skill specific feedback that is pedagogically relevant in indicating the strengths and weaknesses of learners. Relative analysis has shown that the model is better than the current commercial and research systems and it is deeper, reliable, and has more granularity in analysis. The multimodal would have to be able to neutralize distracting noises as well as the varying levels of timbre and timing in an expressive rendition, and would thus have to be employed in the actual teaching environment and not necessarily in the highly controlled laboratory setting. The study presents an extensive framework that enables more sophisticated automated evaluation and assessment in music, sets the foundation for adaptive music learning training systems, supports real-time analytics of music performances, and allows for integrating more broadly the use of AI-based evaluations into music education.
Background: Bridging the gap between requirement analysis and system design remains a critical challenge in software engineering. Existing PF-based approaches lack automated mapping to Unified Modeling Language (UML), causing requirements-design disconnection, inefficient transformation, and poor traceability. Methods: We propose PF2UML, a model-driven transformation method that automatically converts PF models into comprehensive UML requirement models. The approach comprises three components: (1) an extended PF meta-model with semantic annotations addressing structural, behavioral, and service-level gaps; (2) a graphical modeling platform built on Eclipse Modeling Framework (EMF) and Sirius following the Meta-Object Facility (MOF) three-layer architecture; and (3) 24 declarative transformation rules implemented in Atlas Transformation Language (ATL), generating four UML diagram types: conceptual class diagrams, use case diagrams, service diagrams, and system sequence diagrams. Results: Validation on five representative open-source systems with 12 participants demonstrates a transformation accuracy of 93.6% (130% improvement over baseline), an average reduction in end-to-end time from Requirements Modeling and Model Transformation (RM+MT) of 26.8%, and a practical average ROI of 236.6%. Conclusion: PF2UML establishes seamless and traceable integration from requirements modeling to system design, significantly reducing manual effort and improving consistency for complex interdisciplinary system development.
Although large language model (LLM)-based autonomous agents are increasingly being utilized in complex software systems, ensuring their reliability remains a critical challenge due to unpredictable defects such as hallucinations, execution errors, and inconsistent reasoning. This study proposes a reliability-aware self-healing framework for LLM-based software agents. This framework integrates failure detection, reliability assessment, and automated recovery mechanisms. First, we define a failure type classification system and introduce a quantitative reliability assessment model. Next, we present a failure detection technique that identifies abnormal agent behavior based on execution patterns and output consistency. Finally, we design a self-healing mechanism that dynamically recovers from failures through adaptive replanning and prompt correction strategies. The proposed framework was tested in a multi-agent workflow environment simulating real-world operational conditions, and its performance was evaluated through task scenarios. The experimental results show that the approach in this study significantly increases the task success rate, reduces fault propagation, and improves the overall robustness of the system compared to existing methods. In particular, this study is distinctive in that it establishes an integrated monitoring system that combines the agents' internal reasoning processes with external execution results. The proposed framework provides a practical foundation for developing trustworthy intelligent agents capable of autonomous failure recovery and adaptive decision-making in dynamic environments. The proposed approach can serve as a core reliability mechanism for next-generation expert systems and autonomous AI applications.
Binary Code Similarity Detection (BCSD) plays a crucial role in vulnerability discovery and malware analysis. However, it faces two major challenges in cross-architecture scenarios: syntactic variations arising from architectural differences and semantic distortions induced by compiler optimizations. To address these issues, we propose SemASTer, a robust framework for cross-architecture binary code similarity detection that leverages Abstract Syntax Tree (AST) as its semantic backbone. A degradation-aware channel attention mechanism dynamically incorporates compensatory information only when the AST backbone is deemed unreliable due to aggressive optimizations. Furthermore, we introduce two complementary compensation pathways: behavioral semantic compensation to recover lost runtime dynamics, and control-flow compensation to reconstruct disrupted logical dependencies. Experimental results demonstrate that SemASTer achieves an 8.3% improvement in Recall@1 over the current state-of-the-art method in cross-architecture tasks, and an 11.4% gain under fully heterogeneous compilation settings (varying architectures, compilers, and optimization levels). Moreover, its effectiveness and robustness are validated through real-world vulnerability detection scenarios.
This study develops an artificial neural network-based knowledge engineering model for predicting urban–rural integration development. Urban–rural integration is treated as a complex socio-technical process involving nonlinear interactions among information transfer, knowledge spillover, social capital, and digital infrastructure maturity. Survey data were collected from 759 respondents in four cities of Jiangsu Province, China. A hybrid PLS-SEM–ANN approach was applied, where PLS-SEM validated the theoretical relationships and ANN captured nonlinear predictive patterns. The PLS-SEM results confirm that information transfer enhances knowledge spillover and social capital, which subsequently improve urban–rural integration development. Digital infrastructure maturity strengthens the effect of knowledge spillover but weakens the social-capital pathway. The ANN results demonstrate stable predictive performance across training and testing datasets, with sensitivity analysis identifying social capital and information transfer as key predictors. The study contributes to knowledge engineering by transforming social, informational, and digital indicators into an intelligent predictive framework for decision support.
In the modern era, cloud computing (CC) is more important as corporations increasingly adopt hybrid cloud infrastructure to enhance scalability and flexibility. Here, the hybrid cloud setup integrates one or more external public cloud systems to execute the workloads securely within the internal system. In some cases, cloud providers have missed the different security policies that can manage several cloud environments. This may be a reason for decreased performance under latency and data transfer ratios. These cloud environments fail to ensure safe and secure trustworthy communication between the decentralized distributed services. However, the traditional strategies are slowly securing the data seekers from the cloud networks. Therefore, this article introduces a novel framework to address the challenges faced in cloud infrastructure. Consequently, the proposed framework is named Reinforcement Learning-based Zero Trust with Trust Score (RLZT-TS) architecture. This integrates an RL algorithm, a dynamic trust scoring (DTS) model, and a Zero Trust architecture (ZTA) in a hybrid cloud mesh infrastructure. This model can effectively detect cyber threats. Also, for security purposes, an Advanced Multi-Factor Authentication (AMFA) protocol has been implemented in the security layer to secure the data during transmission over the cloud networks. The novel feature of RLZT-TS is a closed-loop adaptive trust model, where the trust score is continuously updated using RL algorithms, the access policy is dynamically adjusted based on the trust score, and the authentication requirements are dynamically changed based on the predicted risk level. The proposed RLZT-TS paradigm can support automated security reconfiguration and trust assessment towards the zero-trust values. Consequently, the model ensures intelligent security replica against evolving cybersecurity threats. Consequently, the developed framework is analysed through the MATLAB simulation, and outcomes illustrate that the significant enhancement in trust score accuracy (98%), attack detection rate (99%), resource utilization (8.82%), policy adaptability (98%), and latency (8.2ms).
End-to-end multimodal sentiment analysis fuses raw visual, acoustic, and textual streams whose tokenized lengths differ by up to 20:1 (1568 visual versus 77 text tokens on CMU-MOSEI). Dense attention biases fusion toward the longest modality (68% visual attention in bimodal TVLT), leaving shorter modalities—particularly the information-sparse acoustic stream—underrepresented. Current trimodal baselines consistently exhibit severe cross-category emotion imbalance (Fear F1: 25–29), a consequence of the same length-asymmetry bias that our architecture is designed to mitigate. We propose HAFT (Hierarchical Audio-Enhanced Fusion Transformer) to address three challenges arising from this asymmetry: computational bottleneck, modality imbalance, and audio under-utilization. HAFT routes audio–visual interaction through bottleneck tokens with adaptive depth-wise gating before integrating text, jointly reducing attention cost and counteracting modality bias; the Cascaded Audio Feature Enhancement (CAFE) framework strengthens prosodic representations via multi-scale time–frequency extraction; and grouped projections, decoupled positional attention, and layer-wise parameter sharing compress the remaining overhead. On CMU-MOSEI under end-to-end raw-input settings, HAFT attains [Formula: see text] Acc-2 and [Formula: see text] WF1 with 85 M parameters and 145 ms latency, improving Acc-2 and WF1 over bimodal TVLT-VA by 6.4 and 4.3 percentage points, respectively. Cross-category emotion F1 standard deviation falls from 15–17 (trimodal baselines) to 6.7, with per-category F1 range narrowing from over 43 to 16.0 points. The resulting trimodal attention distribution (33.4% visual, 32.6% audio, 34.0% text) indicates substantially reduced token-length-induced modality bias.
Software aging remains a persistent challenge in long-running software systems because performance degrades over time through resource exhaustion, memory leaks and related faults. To examine how supervised learning techniques are used to address this problem, this paper presents a search-based literature survey of 93 studies selected through predefined search strings and inclusion and exclusion criteria. The survey focuses on Linear Regression (LR), Support Vector Machines (SVM), K-Nearest Neighbours (KNN) and Backpropagation (BP) neural networks, both as standalone models and in hybrid configurations for software aging, software defect prediction and related reliability tasks. It analyses the modelling roles of these techniques, the reported evaluation settings and the target system types, and reviews BP-based hybrids that integrate LR, SVM, KNN, Random Forest, Decision Tree and Naive Bayes with feedforward, recurrent and convolutional neural networks. The findings identify recurring hybrid patterns, summarise the strengths and limitations of the model families and highlight open issues, including limited comparative evaluation and weak links between prediction models and concrete maintenance or rejuvenation decisions.
Accurate software project effort estimation remains a persistent challenge due to the limitations of traditional models in capturing rich contextual information from project artifacts. Conventional approaches such as COCOMO II and machine learning models primarily rely on structured numerical features, often ignoring the semantic depth present in Software Requirements Specification (SRS) documents. This gap leads to suboptimal estimation performance, particularly in early project stages. To address this limitation, this paper proposes a novel framework that integrates Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) for context-aware software effort estimation. The proposed system leverages SRS documents to extract structured attributes, converts historical project data into semantic embeddings, and retrieves similar past projects using vector similarity search. These retrieved analogues are then utilized by a reasoning-based LLM to generate effort predictions grounded in both contextual and historical evidence. The framework is evaluated on the ISBSG dataset using both traditional error metrics and RAG-specific evaluation measures. Experimental results demonstrate that the proposed approach achieves an MMRE of 10.74%, outperforming baseline models including COCOMO II [4], Random Forest, and Support Vector Regression under identical preprocessing settings. Additional analysis using RAGAS metrics shows high context precision (0.9760) and competitive answer correctness (0.7020), validating the effectiveness of the retrieval-generation pipeline. The results highlight the potential of combining LLMs with retrieval mechanisms to bridge the gap between unstructured requirements and quantitative estimation, offering a scalable and interpretable solution for modern software project planning.
Due primarily to delayed diagnosis, lung cancer remains one of the leading causes of cancer-related deaths worldwide. Automatic detection of lung cancer has been investigated using deep learning (DL) algorithms with chest computed tomography (CT) images, with promising results. Most current models, however, only work on small local image features, making them sensitive to noise and variations in image quality. Also, single-model methods might not ensure sufficient reliability for real clinical use. To address these limitations, a hybrid network, termed Swin-CNN-EL, is developed in this work. The proposed framework employs a Swin Transformer to extract meaningful characteristics from CT scan images, while an ensemble of three Convolutional Neural Network models—VGG19, EfficientNetB7, and Inception-v3—is utilised for classification. The outputs from these classifiers are combined using majority voting, where the final class label is determined by the prediction of the majority of models. This ensemble mechanism improves both classification reliability and overall predictive performance. Before model training, several preprocessing steps are applied to the CT images, including contrast-limited adaptive histogram equalisation (CLAHE), median filtering, and morphological enhancement. In addition, data augmentation methods are incorporated to improve the network's adaptability to diverse imaging variations and reduce overfitting. The effectiveness of the developed model is evaluated using a publicly accessible Kaggle lung cancer CT scan dataset containing 2,274 images categorised into four groups: adenocarcinoma, large cell carcinoma, squamous cell carcinoma, and normal lung tissue. For experimentation, the dataset is split into 70% training, 10% validation, and 20% test data. Furthermore, 5-fold stratified cross-validation is implemented to ensure dependable and unbiased performance assessment. Experimental findings reveal that the proposed Swin-CNN-EL architecture achieves strong classification capability, obtaining 99.6% accuracy, 99.1% precision, 98.7% recall, 98.9% F1-score, and 99.2% AUC. The narrow 95% confidence intervals additionally indicate the consistency and stability of the framework. Overall, the results confirm that the proposed approach can accurately and consistently classify lung cancer from chest CT images.
Adversarial attacks on machine learning-based malware detection models could enable attackers to craft evasive malware variants that remain functional while bypassing detection. However, existing methods primarily focus on feature-space adversarial attacks, lacking effective mapping from feature-space to code-space (i.e. the modifications of source code of malware to implement the adversarial attacks), which limits its practical application. To address this issue, this paper proposes an large language model (LLM) enhanced code-space adversarial attack method against the dynamic malware detection models, achieving end-to-end generation of executable and source code-level adversarial samples. First, we apply a feature-space adversarial attack algorithm to perturb the feature vectors and pinpoint the changes of dynamic behaviors (e.g. API calls). Second, we propose a prompt mechanism guided by the dynamic behavior changes and code generation hints to automatically modify the source code of malware samples by leveraging an LLM. Experiments on five types of typical malware demonstrate that the proposed method can effectively modify the source code of malware based on the adversarial perturbations, and consistently outperforms existing approaches in terms of evasion capability and functionality retention.
Software projects frequently report requirements-related failures even when systems conform to documented requirements and pass acceptance criteria. This paper argues that such failures stem from a conceptual conflation of requirements with technical specifications. We propose a socio-technical, contract-based conceptualization of software requirements, framing them as negotiated commitments defined by assumptions, constraints, and accountability structures. Drawing on requirements engineering and organizational theory, we analyze how requirements function as evolving agreements rather than static technical prescriptions. We identify core elements of requirement contracts and show how misalignment arises when underlying assumptions shift while technical conformance is maintained. The proposed perspective clarifies recurring failure scenarios and provides a structured lens for requirements elicitation, documentation, change management, and traceability. Rather than introducing a new method, this work offers a conceptual model that integrates socio-technical considerations into requirements artifacts and supports improved alignment between stakeholder commitments and technical realization.
Falls in the older population constitute a significant public health issue, motivating the development of automated fall risk assessment methods. However, existing approaches based on wearable sensors or depth cameras are often limited by hardware cost and deployment complexity. To this end, we present DualFRA, a vision-based deep learning approach that leverages standard cameras and integrates complementary mobility information from Timed Up and Go (TUG) and Five-Times Sit-to-Stand (FTSTS) test videos for comprehensive fall risk assessment. DualFRA first applies 3D pose estimation to extract skeletal sequences from the two clinical tests. For each sequence, multiple complementary re``presentations are constructed, including a spatio-temporal skeletal graph and auxiliary views that emphasize joint configurations and motion dynamics. These representations are then processed by a spatio-temporal feature encoder combining self-attention-enhanced graph convolution and multi-scale temporal modeling to capture discriminative movement patterns. The resulting dual-stream features are integrated through cross-attention to produce a coherent semantic embedding for fall risk prediction. Experimental results averaged across multiple independent runs demonstrate that DualFRA attains accuracy and weighted-[Formula: see text]-scores of 87.90% and 87.83%, surpassing baselines with average performance gains of 2.92% and 2.97%, respectively. Ablation analyses further confirm the benefits of the dual-stream design.
In recent years, billions of smart gadgets have been connected by the Internet of Things (IoT) to enhance people's quality of life. However, abnormalities or malevolent attacks create security weaknesses, resulting in a risk to data safety and an impairment in performance for IoT operations. Therefore, IoT security solutions must monitor and prevent undesired activities within the IoT network. The security requirements of IoT applications cannot be satisfied by traditional intrusion detection systems (IDS) since these methods require a lot of processing power, storage space, and training time. As a result, IDS with low weights, quick training times, and excellent detection accuracy must be created for IoT. Therefore, this article developed a deep-learning IoT security framework based on a hybrid TResNet-MiXNet architecture that can identify and anticipate vulnerabilities. This model consists of multiple phases. The first step uses a Conditional Tabular Generative Adversarial Networks (CTGAN) based method to solve the imbalanced data issue in the dataset. After that, the TResNet-based deep learning approach is used to extract the key features of each class from the dataset. Lastly, a classifier based on MiXNet is implemented to identify each attack separately. This study examines three distinct datasets, namely N-BaIoT, ToN-IoT and IoTID20. The empirical inquiry reveals that the presented methodology attains correspondingly remarkable accuracy of 99.38% and 99.45%, 99.41% on IoTID20, N-BaIoT and ToN-IoT datasets. Additionally, it demonstrates notable benefits over existing state-of-the-art models, as demonstrated by multiple evaluation measures.
This study examines how small language models (SLMs) behave under domain-specific fine-tuning in Turkish, a non-English and comparatively lower-resource language. Seven different models with varying sizes and architectures are evaluated across legal reasoning, reading comprehension, analytical abilities, and conceptual knowledge. Fine-tuning sets range from 100 to 10,000 supervised legal examples, enabling systematic analysis of data-scaling effects in the fine-tuning regime. The results show that performance does not improve monotonically with increased training data. Instead, models consistently exhibit a narrow efficiency zone, typically at 100-400 examples, where legal accuracy peaks before degrading due to over-specialization. Larger datasets amplify LoRA update magnitude and density without corresponding gains in accuracy, revealing a divergence between optimization progress and generalization. These effects are attributed to the fragmented nature of legal knowledge, which limits cross-topic transfer and makes models highly sensitive to overshooting the optimal fine-tuning range. Smaller end of considered models seem to lack the representational capacity for stable legal reasoning, whereas larger ones overfit rapidly without careful control. Overall, the study demonstrates that no general-purpose fine-tuning recipe can be assumed effective without comprehensive evaluation across multiple cognitive dimensions for SLMs.
This study proposes an improved N-AKAZE algorithm, which systematically optimizes the efficiency bottleneck of the traditional AKAZE algorithm in the feature matching of bronze drum images. By introducing seven key technical improvements, including adaptive scale selection (ASS), edge-preserving diffusion (EPD), and multi-threshold detection (MTD), a complete algorithm framework with bidirectional matching verification (BMV) is constructed. Experiments show that the average feature point detection time on the bronze drum image dataset is reduced to 85.96ms (p<0.01), and the matching accuracy is improved by 37.2%. This has significant advantages over the traditional AKAZE algorithm (t=5.32, df=12, p=0.0002). The research results provide a new technical paradigm for the digital preservation of cultural heritage.
Cross-Site Request Forgery (CSRF) vulnerabilities exploit legitimate user actions through unauthorized mechanisms, leading to data tampering, sensitive information leakage, and even system outages, thereby posing a serious threat to the widespread deployment of the industrial Internet. Traditional vulnerability detection methods struggle to effectively address the inherent complexities of CSRF vulnerabilities, resulting in low detection efficiency and limited practical effectiveness. To improve detection efficiency and accuracy, we propose CSRFinder, an enhanced CSRF vulnerability detection method based on dangerous requests and Gradient Boosting Decision Tree (GBDT) algorithm. It first addresses the challenge of collecting dangerous request data by defining explicit rules for identifying dangerous requests and constructing feature datasets that incorporate both attribute-based and textual features. Building on these datasets, CSRFinder employs an ensemble learning approach with multiple classifiers to train a request classification model. Finally, for candidate dangerous requests identified, request replay techniques are employed to construct request forgery scenarios and corresponding vulnerability determination strategies. CSRF vulnerabilities are identified by comparing the responses and corresponding database operations triggered by the original and forged requests. CSRFinder was validated on 15 real-world open-source web applications. Compared with single classifiers and standalone deep learning methods, the GBDT classifier demonstrated superior performance. Specifically, CSRFinder identified 3 CSRF vulnerabilities in PrestaShop websites that were missed by the SOTA method Deemon, while producing only one false positive. Additionally, it detected 9 more CSRF vulnerabilities than the classic method Mitch and exhibited superior overall performance across multiple evaluation metrics.
Self-Admitted Technical Debt (SATD) refers to situations where developers explicitly acknowledge design or implementation compromises in source code comments. Although SATD detection has been extensively studied in recent years, most datasets are severely imbalanced, leading models to bias toward the majority class. This paper empirically investigates how class imbalance affects the performance and stability of SATD prediction models. We design a dataset generation algorithm to construct datasets with varying imbalance ratios and benchmark four representative approaches: unsupervised pattern matching (MAT), traditional text mining (TM), maximum-entropy-based natural language processing (NLP), and convolutional neural networks (CNNs). We evaluate these approaches under both many-to-one (MTO) and one-to-one (OTO) prediction scenarios. Experimental results show that CNN achieves competitive performance and stability in cross-project (MTO) prediction, while MAT remains stable in single-project (OTO) settings and under highly imbalanced conditions. TM and NLP methods can reach local peaks under moderate-to-high imbalance ratios. These findings delineate the trade-offs among methods, offering actionable guidance for selecting SATD detection models under varying imbalance conditions.
Software architecture quality evaluation is an important means of assessing architectural soundness and informing the design, maintenance, and evolution of complex software systems. However, existing architecture quality evaluation approaches are largely source-agnostic, and mixed-source-specific quality concerns remain underexplored in architecture quality evaluation. They do not explicitly capture the source-related quality concerns introduced by the coexistence of autonomous and reused code. In mixed-source architectures, autonomous and reused entities often differ in how they are organized, evolved, and constrained, giving rise to concerns related to source interweaving, reduced dependency autonomy, and license-induced architectural constraints. As a result, architectures with substantially different mixed-source configurations may still receive similar quality evaluation results under source-agnostic methods. To address this limitation, we propose MSAQM, a mixed-source architecture quality model that introduces explicit source awareness into architecture evaluation. MSAQM formalizes mixed-source-specific quality concerns through three complementary attributes: Source Interweaving Degree (SID), Path Source Interference (PSI), and License Constraint Severity (LCS). Rather than replacing existing evaluators, MSAQM is designed as a plug-in layer that enhances them with mixed-source awareness while preserving their original general architecture quality evaluation capabilities. We evaluate MSAQM on representative mixed-source architecture changes. The results show that MSAQM improves the ability of existing evaluators to capture mixed-source-specific architecture quality changes while maintaining high consistency with their original assessment behavior.
