Journal of Soft Computing and Computational Intelligence (p-ISSN: 3107-4855, e-ISSN: 3048-6610)

ChaosFusion: An Adaptive 5D Hyper-chaotic Framework for Secure Image Encryption with 3D Pixel Permutation

2026-09-11T10:17:47+00:00

The growing exchange of digital images across public networks has made image encryption a critical area of research. Conventional cryptosystems such as DES and AES are ill-suited to image data because of its large volume, high redundancy, and strong inter-pixel correlation, motivating the shift towards chaos-based encryption. This paper surveys recent literature on chaotic and hyper-chaotic image encryption techniques, examining pixel reorganization strategies, nonlinear transformations, dynamic S-box design, and hybrid approaches that combine chaos with logic gates or deep-learning autoencoders. Each approach is analyzed for its contribution to confusion–diffusion strength, keyspace, and computational cost. The survey identifies a recurring trade-off between security strength and processing efficiency across existing methods and uses these gaps to motivate a proposed framework based on 3D pixel scrambling and a 5D hyper-chaotic keystream. The survey further considers how adaptive confusion–diffusion can strengthen a practical encryption pipeline by allowing the chaotic keystream and permutation behavior to respond to secret parameters rather than relying on a fixed transformation sequence. Particular attention is given to the relationship between spatial scrambling, hyper-chaotic sequence generation, key sensitivity, and measurable security indicators. The reviewed studies show that increasing algorithmic complexity can improve resistance to statistical and differential analysis, but may also increase execution time, memory consumption, and implementation difficulty. Therefore, the proposed direction emphasizes a balanced architecture in which 3D pixel scrambling reduces spatial redundancy and a 5D hyper-chaotic keystream performs adaptive diffusion.

Attention-Guided Ensemble Deep Learning Model for Skin Cancer Detection Using Lesion Images

2026-10-13T09:54:29+00:00

Due to its high prevalence and the rapid progression of its malignant forms, skin cancer is a critical public health issue where rapid and accurate detection significantly influences survival rates. Currently, diagnosing this disease primarily relies on a dermatologist's visual assessment, followed by histopathological confirmation. This traditional method is inherently time-consuming, requires significant specialist resources, and is susceptible to inter-observer variability. To address these challenges, the study introduces a novel attention-guided ensemble deep learning framework designed to automate lesion classification directly from dermoscopic imagery. This study system synergistically combines EfficientNet-B4, DenseNet121, and a Vision Transformer (ViT) into a unified architecture. This integration ensures that both localized textural details and macro-level contextual structures are comprehensively analyzed during the final classification phase, preventing the loss of crucial diagnostic information. The complete workflow encompasses image preprocessing, lesion segmentation, multi-backbone feature extraction, ensemble fusion, and rigorous performance evaluation. The model was trained and evaluated using the ISIC dermoscopic dataset, comprising 2,357 benign and malignant cases. The study's ensemble approach achieved an outstanding accuracy of 99.21%, outperforming standard architectures such as ResNet50, ResNet101, and ResNet152 across all key metrics, including precision, recall, and F1-score. These findings strongly indicate that employing ensemble deep learning strategies offers a substantial improvement over single-model architectures, presenting a highly viable tool for computer-aided diagnosis in dermatology.

Explainable Neuro-fuzzy System for Coral Reef Fish Classification Using Multi-Scale Visual Features

2026-09-18T06:54:00+00:00

Automated coral reef fish classification from underwater imagery is essential for scalable biodiversity monitoring, yet the deep convolutional networks achieving the highest accuracy remain largely opaque, limiting adoption by marine scientists who must justify conservation decisions. This article presents an explainable neuro-fuzzy system that classifies coral reef fish species from multi-scale visual features while producing human-readable justifications for every prediction. The proposed architecture extracts fine-, mid-, and coarse-scale convolutional features corresponding to local texture, body shape, and global coloration, fuses them into a unified embedding, and passes this embedding through a Takagi-Sugeno neuro-fuzzy inference layer whose linguistic rules and membership functions remain directly inspectable. A complementary explanation module combines fuzzy rule firing strengths with gradient-based saliency maps to generate joint symbolic-visual explanations. The architecture, its multi-scale feature extraction strategy, the fuzzy rule base, and representative underwater fish and coral datasets are described, and comparative tables summarize architectural components, datasets, and accuracy-interpretability trade-offs against convolutional and standard neuro-fuzzy baselines. Results indicate the proposed system retains competitive classification accuracy while substantially improving interpretability relative to conventional deep learning classifiers, supporting its use as a trustworthy tool for automated reef fish monitoring and conservation decision support.

A Zero-Day Vulnerability Detection and Automated Security Assessment Framework

2026-10-21T07:27:57+00:00

The increasing dependence on digital systems, software applications, cloud storage, and digital infrastructure has intensified cybersecurity risks, particularly those associated with previously unknown or undocumented vulnerabilities. This study aims to propose a Zero-Day Vulnerability Detection and Automated Security Assessment Framework for identifying potential unknown security weaknesses, evaluating their severity, and recommending appropriate mitigation measures. The proposed framework adopts a Vulnerability Assessment and Penetration Testing (VAPT) methodology that integrates asset discovery, vulnerability assessment, vulnerability classification, controlled validation, behavioural analysis, risk prioritization, remediation, and reassessment. Identified vulnerabilities are correlated with established security references, including Common Vulnerabilities and Exposures (CVE), Common Weakness Enumeration (CWE), Common Vulnerability Scoring System (CVSS), and the OWASP Top 10. Furthermore, the framework supports security-control mapping to relevant standards and compliance frameworks, including NIST, ISO/IEC 27001, and HIPAA. Findings that do not correspond to existing vulnerability records undergo additional behavioural analysis and validation to determine whether they represent potential zero-day vulnerabilities. The proposed framework establishes an integrated workflow for vulnerability identification, risk-based classification, and automated remediation guidance. Its expected outcomes include improved assessment consistency, more effective vulnerability prioritization, and reduced false-positive classification through controlled validation and reassessment. Experimental evaluation using measurable indicators, including detection accuracy, false-positive rate, and remediation effectiveness, is required to determine the framework’s actual performance.

Integration of Computer Vision and Digital Twin Technologies for Intelligent Industrial Monitoring: A Comprehensive Review

2026-10-05T11:42:21+00:00

Digital twin technology and computer vision have independently advanced as pillars of intelligent cyber-physical systems, and the convergence of the two fields is increasingly regarded as a pathway toward richer, more autonomous, and more accurate virtual representations of physical assets. This review examines the theoretical foundations, technical enablers, and application domains associated with vision-enabled digital twins. A layered conceptual framework is proposed, structured around sensing, visual perception, real-time synchronization, simulation, and decision-support components. The role of deep learning architectures, including convolutional neural networks, transformer-based detectors, and 3D reconstruction methods, is discussed in relation to their contribution toward automated state estimation, geometry capture, and anomaly detection within digital twin environments. Reported implementations across manufacturing, healthcare, smart infrastructure, and autonomous transportation are synthesized to illustrate performance trends, integration strategies, and reported benefits such as reduced downtime, improved predictive maintenance accuracy, and enhanced situational awareness. Persistent challenges are identified, including latency in bidirectional data synchronization, limited generalization of vision models across varying environmental conditions, high computational demand, and vulnerability to cybersecurity threats. Future research directions are outlined, emphasizing edge-based inference, multimodal sensor fusion, and standardized interoperability protocols. The review is intended to serve as a consolidated reference for researchers and practitioners pursuing vision-integrated digital twin solutions across industrial and urban contexts.