The integration of artificial intelligence and the Internet of Things (AIoT) enables advanced edge computing but creates significant energy challenges for resource-constrained devices. To address these challenges, this paper introduces a novel comparative taxonomy and a systematic gap analysis matrix that directly maps hardware-aware TinyML paradigms to network-layer scheduling in sustainable AIoT systems. Synthesizing recent empirical studies (2023–2026) on power consumption models for IoT edge devices, machine learning techniques for energy optimization, and sustainable deployment strategies, we demonstrate that additive and regression-based models achieve low prediction error (MAPE 4–... More >

These days, many eco-systems related to federated learning, blockchain, self-driving cars, and scientific computing have many agents working together, each doing its own part. Using a single central system to check if all the agents are doing their work correctly is slow and gets more expensive as more agents are added. This paper introduces a new way called the Verification Game (VG). The agents don’t depend on a central system. The agents check each other’s work. If the agents are honest, they get rewards, so telling the truth is the best option. This type of method also saves a lot of computing power because it doesn’t check every single task. We also came up with a method called Ad... More >

This paper introduces a course-specific agentic retrieval-augmented generation (RAG) chatbot developed to support Information Technology students at Hai Phong University. The proposed system addresses the limitations of static FAQ bots, which are unable to manage open-ended academic tasks, and model-only large language model (LLM) assistants, which may generate fluent yet insufficiently grounded responses. The prototype combines local LLM deployment, semantic retrieval of course materials, and constrained, tool-oriented orchestration to perform four key tasks: question answering, document summarization, study planning, and quiz generation. The primary contributions include a six-layer privac... More >

The intrusion detection at the 5G network perimeter demands learning frameworks that are practically feasible and computationally efficient. This research proposes a lightweight, slice-sensitive intrusion detection approach designed for edge deployment, with a strong emphasis on minimizing information leakage while accounting for the resource constraints inherent in edge environments. A rigorous chronological and session-discontinuous experimental protocol ensures that training and test traffic remain temporally separated, faithfully replicating realistic deployment conditions. The proposed framework employs a classical Logistic Regression classifier using flow-based statistical features ext... More >

In the current academic landscape, students often face challenges in identifying suitable institutions for higher studies based on their academic and profile attributes. Existing advisory services and online tools are either expensive or lack predictive accuracy. This research proposes a machine learning-based admission prediction system that estimates the probability of university admission using historical applicant data. Linear Regression serves as a baseline model to capture linear relationships, Random Forest models non-linear feature interactions, and CatBoost is selected for its robustness on structured tabular data and native handling of categorical features. Comparative evaluation u... More >

The new progress in text-to-3D technology has greatly changed and improved the artificial intelligence (AI) applications in augmented and virtual reality (AR/VR) environments. Many different techniques in 2024-2025 like diffusion models, Gaussian splatting, and physics aware models have helped the text-to-3D much better by improving the visual fidelity, semantic coherence, and generation efficiency. Some models like Turbo3D, Dive3D and Instant3D are deigned to make the 3D generation faster by improving the working process of diffusion models. Other frameworks like LAYOUTDREAMER, PhiP-G and CompGS focus on creating scenes that are well organized and structured. Dream Reward and Coheren Dream... More >

This research presents a new framework for clinical summarization that combines the TinyLlama model with MIMIC-III and FHIR data using the Model Context Protocol (MCP). Unlike cloud-based models like Med-PaLM, our approach uses local processing to cut costs and protect patient data with AES-256 encryption and strict access controls, meeting HIPAA and GDPR standards. It retrieves FHIR-compliant data from public servers (e.g., \texttt{hapi.fhir.org}) for interoperability across hospital systems. Tested on discharge summaries, it achieves ROUGE-L F1 scores of 0.96 for MIMIC-III and 0.84 for FHIR, beating baselines like BioBERT (0.61, p < 0.001) due to efficient preprocessing and MCP’s accurat... More >

Owing to their dynamic user interactions, ever-changing structure, and real-time content changes, dynamic social networks pose significant privacy and security risks. The state of security and privacy-preserving techniques in these developing platforms is thoroughly examined in this study. We highlight the benefits and drawbacks of various approaches as we review recent studies on privacy-preserving tactics, security updates, and anonymisation methods. Important findings indicate that present approaches often fail in dynamic situations, even when they operate well in static network conditions. Beyond common problems, we also point out important security aspects influenced by hierarchical sys... More >