In recent years, Content-Based Image Retrieval (CBIR) has advanced significantly with Deep Learning (DL), yet most of the existing models rely on global descriptors and attention-based local features, which struggle to capture explicit structural relationships within images. These models often operate as black boxes and provide limited interpretability; they exhibit reduced robustness under viewpoint and geometric variations. To address these limitations, an Interpretable Region-Attribute Graph Network (IR-AGN) is proposed in this research. Specifically, each image is decomposed into semantically meaningful regions, where each node encodes visual and semantic descriptors while edges represent geometric, appearance, and contextual relations. Consequently, a joint node-edge graph matching strategy computes similarity by aligning visual descriptors and relational structures between query and retrieval images. The results of the proposed IR-AGN model exhibit a higher mean Average Precision (mAP) of 95.19% on the Revisited Paris (RParis) dataset, when compared with the existing Multi-Layer Orientation Histogram (MLOH) model.
Malware classification is the process of categorizing malicious software into specific families based on its structural and behavioral features. However, malware classification remains challenging due to high code similarity across different families, leading to suboptimal performance. To address this issue, this research proposes a Dynamic Cluster Adaptive Aggregation Graph Neural Network with Prototypical Loss Function (DCAAGNN-PLF) for effective malware classification. DCAA is incorporated in the GNN to adaptively group similar nodes that effectively capture both global and local structural variations. PLF ensures robust intra-class clustering while enhancing inter-class separation in the feature space that minimizes misclassification. Min-max normalization scales all features to a common range that avoids larger numerical values and enhances model performance. Therefore, the proposed DCAAGNN-PLF achieves better accuracies of 99.90% and 99.25% on the Malimg dataset and the Microsoft BIG2015 dataset, respectively, compared to existing techniques such as Convolutional Neural Network (ConvNet).
This article presents the principle of adaptive position-trajectory control for moving objects, illustrated using a multi-link industrial robot manipulator. The principle is described through several stages derived from its main control objectives. A SCARA robot with adaptive position-trajectory control is chosen as the study object, and mathematical models are provided to compute its kinematic structure and dynamic characteristics. These models make it possible to determine the manipulator’s position in coordinates, the trajectory and position of the end-effector, and to analyze real-time trajectory generation. Based on the proposed principle, boundary conditions for the robot’s dynamic properties are formulated, enabling the development of control models, algorithms, and robust-adaptive position-trajectory control structures. The approach follows a classical PD controller framework using filtered and approximated tracking errors. Computer visualization of the robust-adaptive control structures is used to examine dynamic behavior. The proposed structures ensure high-accuracy and high-speed motion along specified trajectory positions.
This paper presents a new hybrid image encryption algorithm, which combines the SPECK block cipher with a Salsa20-based key generation process. Although these factors have been integrated in the past, what is unique about our work is that DNA encoding and transformation are applied only to the key stream, not to the image data itself. The method produces a non-linear, irreversible key obfuscation layer that enhances the system’s resistance to key-based attacks at a relatively small cost to the overall computation of the main image encryption operation. The system will have the key generated by Salsa20 to 128 bits, converted by DNA complement and reverse operations, and finally loaded into the SPECK cipher to be encrypted. The approach provides a high-security level but with memory-efficient computation, and, therefore, it is resource-constrained and can be applied to IoT devices. The system shows robust results in NPCR (>99.6%), UACI (~33%), and entropy (7.99).
The paper addresses the question of creating long-living software systems. One of the possible challenging comprehensive solutions for this might be to provide a viable framework which supports their complete life-cycle. The concept and architecture of a scalable framework for creating and maintaining software ontology-centric systems are proposed. The basis for the framework is the 2-level approach to the formation of system components of all types. It allows one to separate component models (ontologies) from components themselves. The 3-layer architecture of the framework is proposed. An implementation of the framework concept as a cloud platform was created. Several knowledge portals have been developed with its use. This evaluation showed that the proposed solution is a promising way for the organization of an ecosystem for collaborative development of viable information and software resources.
