Volume 4 Issue 2: September 2025

The advancement of the Internet of Things (IoT) provides a set of new possibilities and challenges within the Industrial and Manufacturing environments, as well as into the Private Sector and the user daily life in our houses. In this context it is important to design and provide an IoT integrated system with a scalable architecture while maintaining a set of competitive costs and performance. This paper presents the development of low cost IoT‑based Smart Home Temperature and Humidity Monitoring System. The proposed architecture aims to demonstrate core IoT principles such as real‑time data collection, remote device control, and scalable architecture using low cost technologies, such as Arduino Uno R4 Wi‑Fi and ESP32 microcontrollers. The system successfully simulated appliance control ‑ e.g., radiators, extractor fans ‑ via LEDs and basic actuators, combined with a mobile application providing real‑time environmental data ‑ e.g., temperature, humidity, Carbon Monoxide (CO) levels ‑ and remote‑control functionality. Thanks to the proposed design the architecture is also inherently scalable, customizable and expandable, combining a modular approach with a customized mobile app and an user friendly interaction. Challenges included hardware compatibility, power management, and software integration, with further work on security features (i.e., cryptography algorithms) and cloud integration.

• View pdf

Understanding public opinion at scale is both a scientific challenge and a practical necessity in the digital era, as the proliferation of online communication platforms has created unprecedented opportunities to monitor attitudes in near real time. Early work in subjectivity detection and semantic orientation laid the methodological foundations for automated sentiment extraction, focusing on distinguishing objective from subjective content and determining polarity. Contemporary applications, however, face far more complex requirements, demanding systems capable of processing massive, noisy, and dynamic data streams while integrating multimodal signals from text, images, audio, and video. This paper presents a historical review of sentiment analysis and opinion monitoring through the lens of artificial intelligence, tracing developments from the early 1990s to the present and classifying approaches from lexicon‑based heuristics to classical machine learning, deep neural architectures, transfer learning, and multimodal fusion, with an emphasis on both technical and conceptual advances. Extensive tables summarize algorithms, datasets, and case studies across various domains, including politics, finance, and entertainment, highlighting practical lessons and performance trends. The review also addresses pressing ethical concerns, including bias, fairness, and transparency, and considers the implications of rapidly evolving AI capabilities. We conclude by outlining future directions that emphasize adaptability, context awareness, and the seamless integration of emerging technologies into scalable and reliable opinion analysis systems.

• View pdf

Vehicular Ad hoc Networks (VANETs) face significant challenges in maintaining seamless connectivity due to frequent handovers caused by high vehicular mobility, fluctuating signal strength, and dynamic traffic conditions. Conventional static threshold‑based handover schemes are unable to adapt to these variations, often resulting in excessive latency, packet loss, handover failures, and poor reliability. Such drawbacks are particularly detrimental to critical applications like collision avoidance and real‑time traffic management. To address these limitations, this paper proposes a dynamic handover framework based on Q‑learning integrated with Li‑Fi communication technology. The framework intelligently evaluates real‑time parameters, such as vehicle speed, signal strength, and network occupancy, to make adaptive handover decisions rather than relying on preset thresholds. The proposed approach is modeled and tested using OMNeT++ and SUMO simulation platforms. Experimental results demonstrate that the proposed framework achieves a handover success rate of 90.0%, reduces failure rates to 9.8%, and limits ping‑pong handovers to 4.7%. Moreover, throughput improves by 20.3%, and handover delay decreases by 30.2% compared to other state‑of‑the‑art approaches. The packet delivery ratio is sustained at 95.6% even under high traffic density, indicating system robustness and scalability. These findings highlight that reinforcement learning‑based handover management provides a promising solution for next‑generation VANETs, offering adaptability, reliability, and efficiency for emerging intelligent transportation systems and critical vehicular applications, thereby ensuring safety.

• View pdf

Unmanned Aerial Vehicle (UAV) path planning in unknown environments continues to pose a significant challenge, as Deep Reinforcement Learning (DRL) solutions are often severely hampered by slow convergence rates as well as unstable training dynamics. To address this gap, we introduce a Genetic Programming–seeded Soft Actor–Critic (GP+SAC) approach in which Genetic Programming produces high-quality trajectories that are introduced into the replay buffer of SAC as a “warm-start” policy to prevent wasteful early exploration. Through experiments in three benchmark grid environments, we demonstrate that GP+SAC converges significantly more rapidly than the FA-DQN baseline, achieving superior returns in fewer episodes while capitalizing on the same reward design. We show that in large environments, GP+SAC achieved a mean path length of 30.55 units as compared to FA-DQN’s 28.38, thus validating that rapid convergence has no tradeoff in path efficiency. Observably, results also show that as much as GP+SAC obtains superior cumulative rewards, there is a visible fluctuation in the level of training that is indicative of instabilities under very constrained environments. Numerical evaluations show that the proposed GP+SAC agent converges significantly faster than the FA-DQN baseline, achieving higher episodic returns within only a few episodes. In terms of path efficiency, GP+SAC yields an average path length of 30.55 units, which is comparable to the FA-DQN’s 28.38 units, demonstrating that accelerated convergence is achieved without sacrificing path optimality.

• View pdf

The exponential growth of data traffic and user demand in modern communication systems has significantly increased the complexity of data streaming and management in Beyond Fifth Generation (B5G) networks. These networks face critical challenges such as network congestion, traffic load imbalance, latency, energy consumption, spectrum inefficiency, and limited storage capacity for real-time content delivery. Addressing these issues requires new architectural and conceptual approaches rather than incremental improvements to existing methods. This paper introduces a novel conceptual framework for cache-enabled uplink transmission within heterogeneous network environments comprising Macro Base Stations (MBSs), Small Cell Networks (SCNs), and mobile user devices. The proposed framework aims to optimize uplink content delivery by eliminating redundant cached data through curated content lists and employing content segmentation for distributed cache placement. The framework is organized into three interrelated components: Unified Distributed Cached Content Management at the MBS level, Content Deduplication and Segmentation at the SCN level, and Content Matchmaking at the mobile device level. Together, these components enable efficient data synchronization, enhance resource utilization, and minimize redundant data transmissions. Although this study is primarily conceptual, it establishes a strong theoretical foundation for future experimental validation. The proposed design is expected to improve traffic management efficiency, reduce energy consumption, and enhance Quality of Service (QoS) and user experience in future B5G and 6G communication environments.

• View pdf

Backdoor attacks pose a critical and increasingly realistic security threat to deep neural networks (DNNs), enabling adversaries to implant hidden behaviors that remain dormant under normal conditions while preserving high performance on benign data. Although numerous defenses have been proposed, most works treat the interaction between attackers and defenders in isolation, without a principled mechanism to analyze their strategic interplay under realistic resource constraints. This paper introduces BG_Cost, a zero‑sum game‑theoretic framework that formalizes backdoor attack–defense dynamics with explicit cost‑aware utility functions. The attacker seeks to maximize Attack Success Rate (ASR) while maintaining Clean Data Accuracy (CDA) above an acceptance threshold to remain stealthy, whereas the defender aims to limit ASR and preserve CDA while minimizing the computational and accuracy costs induced by mitigation. By embedding resource consumption directly into the utilities of both players, BG_Cost provides a structured benchmark to study equilibrium strategies across unconstrained, balanced, and high‑cost operational regimes. Through numerical simulations, we show that cost‑aware game modeling fundamentally alters equilibrium behavior: unconstrained settings drive extreme strategies, costly defenses weaken robustness, costly attacks suppress adversarial impact, and balanced configurations yield deployment‑friendly equilibria with low ASR and high CDA. Rather than proposing a new algorithmic defense, BG_Cost serves as a decision‑theoretic tool that complements existing mechanisms by revealing how cost constraints shape optimal attacker–defender behavior in practice, guiding the design of realistic and resource‑efficient protections against backdoor threats.

• View pdf

Information Retrieval (IR) has undergone a profound transformation in the field of Natural Language Processing (NLP), shifting from traditional keyword-based approaches to neural architectures and, more recently, to advanced generative Large Language Models (LLMs). Transformer-based models such as Bidirectional Encoder Representations from Transformers (BERT) have substantially improved semantic understanding and retrieval accuracy by enabling contextualized embeddings, deeper interaction modeling, and more effective ranking mechanisms. The rise of Retrieval-Augmented Generation (RAG) represents a significant development by integrating retrieval with generation to produce factually-grounded, context-aware, and explainable outputs, while reducing the likelihood of hallucinations commonly associated with LLMs. This survey provides a comprehensive review of modern IR techniques, focusing on BERT-based retrieval models, emerging generative retrieval frameworks, evaluation methodologies, and key application domains. We provide a structured taxonomy of IR methods and conduct a comparative analysis of state-of-the-art research to highlight performance trends and methodological distinctions. Ongoing challenges, including scalability, computational efficiency, interpretability, and limitations of current evaluation benchmarks, are critically discussed. Additionally, the survey explores emerging directions such as multimodal and cross-modal retrieval, hybrid dense-sparse architectures, knowledge-graph-enhanced retrieval, and the integration of LLMs as unified retriever-generator systems. These advancements illustrate the rapid evolution of IR. They also underscore the need for adaptive, reliable, and transparent retrieval solutions in increasingly complex information environments. This work aims to provide researchers and practitioners with a clear and organized overview of the evolution, current landscape, and future research opportunities in IR within the era of LLM-driven NLP.

• View PDF

This paper introduces a selective image encryption framework for color imagery, emphasizing computational efficiency without compromising practical security. The core idea is to encrypt only the most informative and perceptually critical components of each pixel, while bypassing nonessential data to reduce processing cost; a full-encryption variant is implemented to enable apples-to-apples comparison. Evaluated on a representative set of natural color images, the selective scheme achieves encrypted outputs with an average Peak Signal-to-Noise Ratio (PSNR) of 8.7 dB, Mean Structural Similarity (MSSIM) of 0.07, and Information Entropy (IE) of 7.8 bits. These values are closely aligned with those obtained under full encryption, indicating low residual similarity to the plaintext and near-uniform randomness in cipher histograms. In qualitative terms, the visual content is thoroughly obfuscated, while the selective strategy decreases the amount of data entering the cryptographic core, yielding measurable speedups. The design integrates permutation and diffusion stages suitable for block-based processing and common cipher modes, and supports region-of-interest operation when desired. Together, the empirical evidence and design choices suggest a practical trade-off: comparable security indicators at a fraction of the computational effort. The approach is particularly attractive for resource-constrained settings, batch protection of large image corpora, and latency-sensitive pipelines. Future work will extend the methodology to grayscale imagery, broaden the testbed, and incorporate a dedicated pseudo-random number generator to decouple randomness from platform dependencies.

• View PDF