Journal of Intelligent Systems and Internet of Things

Student performance during the lecture needs to be closely watched to ensure effective learning takes place. This helps the lecturer monitor the performance of the students in real time. By observing the performance of the students, the lecturer can detect the ones who find performance difficult and assist them accordingly. Besides this, the lecturer can also modify the method of teaching whenever needed. By understanding that their performance can be checked through the system, the student remains motivated to perform even in class. The study will help to develop a system that can be used to monitor the performance of the student during the real time lecture using sound and image processing. The method of developing the system involves the use of two methods: image processing and sound processing. The image processing technique can be used to detect the image of the student, while the sound processing technique will be used to detect the sound of the student during the performance. In the proposed system, Gray Level Co-occurrence Matrix technique has been used along with the Viola-Jones method to detect images along with the weighted Euclidean distance method used in image processing. Additionally, the Mel Frequency Cepstral Coefficients method has been used to detect the relevant sound along with the classification method involving the K-Nearest neighborhood method. The experiment has shown the efficiency of the system developed because the accuracy of image and sound identification of the student was at an average of 89% and 90% respectively. All of this helped to ascertain the efficiency of the system in the development of the research study.

The proliferation of Internet of Things (IoT) technologies has transformed digital ecosystems, creating highly interconnected environments that demand robust and adaptive cybersecurity governance. Despite their widespread adoption, existing Information Technology Governance (ITG) frameworks—such as the NIST Cybersecurity Framework (CSF), ISO/IEC 27001, Center for Internet Security (CIS) Controls, and ISA/IEC 62443 vary considerably in scope, applicability, and alignment with the unique characteristics of IoT infrastructures. The absence of a unified approach to address IoT-specific challenges such as device heterogeneity, data provenance, and real-time monitoring underscores the need for a comprehensive comparative analysis. This study conducts a qualitative synthesis and thematic comparison of leading cybersecurity governance frameworks to evaluate their effectiveness in ensuring compliance and resilience within IoT-enabled environments. Each framework was examined across recurring governance domains, including risk management orientation, scalability, control comprehensiveness, interoperability, and contextual adaptability. The analysis integrated findings from scholarly literature, international standards documentation, and expert reports, allowing the identification of emergent patterns, convergences, and gaps in the frameworks’ conceptual foundations and implementation practices. The findings indicate that NIST CSF provides a highly flexible, sector-neutral architecture fostering adaptive governance, whereas ISO/IEC 27001 offers formalized, audit-oriented structures suitable for organizations emphasizing certification and policy compliance. The CIS Controls framework emerges as practical and accessible, favoring rapid implementation and community-driven updates, while ISA/IEC 62443 demonstrates unparalleled domain specificity and defense-in-depth design for industrial and cyber-physical systems. Nevertheless, all frameworks exhibit limitations when addressing IoT-centric issues such as dynamic risk contexts, interoperability among heterogeneous devices, and integration of operational and information technology governance layers. The study concludes that a composite, layered governance approach—anchored in the structural rigor of ISO/IEC 27001, the adaptability of NIST CSF, the practicality of CIS Controls, and the industrial depth of ISA/IEC 62443—can offer a more holistic foundation for IoT cybersecurity compliance.

With the rapid increase of digital content distribution, video watermarking ownership has become an essential tool for detecting certification and tampering. This paper proposes a novel 3D video Zero-Watermarking Framework that integrates machine learning, cryptographic protocol, and entropy-based keyframe selection to ensure strength, inconvenience, and safety. The method operates at two levels: client-side watermark generation and server-side certification. On the client side, the keyframe is extracted using entropy analysis, features are obtained with different 3D Convolutional Neural Network (S3D-CNN), and adaptive noise is generated through the generative adversarial network (GANS). These components are paired with XOR to create a binary watermark key, which undergoes NIST random tests before being safely sent with the original video. On the server, Feige-Fiat-Shamir (FFS) certifies the watermark without highlighting the sensitive information of the zero-knowledge protocol. The system is evaluated against general attacks such as Gaussian noise, JPEG compression, staining, salt-and-pepper, rotation, and scaling. Performance metrics (PSNR, SSIM, NCC, and BER) with FFS protocols, showing 98.7% accuracy in verifying watermark integrity, display strong strength and inevitability. Experimental results, supporting safe and decentralized certification, confirm the effectiveness of the framework proposed to maintain watermarks under various attacks. Future work will focus on integrating blockchain technology and increasing the GAN model for real-world deployment.

The rapid proliferation of edge-AI systems in IoT, autonomous robotics, and biomedical monitoring demands ultra-low-power, latency-aware intelligence that conventional deep neural networks struggle to provide due to heavy computation and memory overheads. Neuromorphic computing offers a promising biological-inspired alternative by processing information through sparse spiking events, enabling energy-efficient on-device learning and inference. This paper presents a neuromorphic VLSI accelerator based on a hybrid spiking neural architecture that combines Leaky-Integrate-and-Fire (LIF) neurons, adaptive threshold spiking units, and synaptic plasticity circuits to support both supervised and unsupervised learning modes at the edge. A hierarchical crossbar-memory topology integrated with non-volatile memristive synapses provides dense weight storage and real-time synaptic updates, reducing off-chip memory access by 78%. A pipelined event-driven computation engine and clock-gated spike scheduler minimize dynamic switching, achieving 61% reduction in power and 2.4× throughput improvement compared to conventional CMOS DNN accelerators. The proposed system performs dynamic visual-feature encoding, spike-based temporal fusion, and on-chip learning for anomaly and object detection tasks in low-power sensor nodes. Fabricated in 28-nm CMOS, the prototype achieves 0.29 mW power, 0.42 pJ/spike energy, and 94.3% inference accuracy, outperforming state-of-the-art neuromorphic platforms. Results demonstrate that hybrid spiking architectures integrated with VLSI-efficient plasticity circuits can deliver high-accuracy, self-adaptive AI within stringent edge constraints, enabling next-generation smart-sensing and autonomous micro-robotic intelligence.

Image enhancement remains a fundamental challenge in computer vision, particularly in scenarios involving low contrast, uneven illumination, and noise interference. While traditional spatial and frequency domain techniques efficiently address specific distortions, they often fail to generalize across diverse image conditions. To overcome these limitations, this paper proposes an Adaptive Hybrid Image Enhancement Framework that integrates deep learning-based enhancement networks with classical filtering algorithms for optimal visual restoration and detail preservation. The proposed method employs a Convolutional Neural Network (CNN) enhanced with an attention-guided residual block to learn fine-grained illumination patterns, followed by adaptive fusion with traditional filters such as Gaussian smoothing, histogram equalization, and bilateral filtering. This hybrid approach ensures a balance between structural clarity and natural color consistency. A dynamic weighting mechanism is applied to adjust enhancement intensity based on local luminance and texture statistics. Experimental validation on benchmark datasets such as MIT-Adobe FiveK, BSD500, and LIME demonstrates significant improvement over state-of-the-art methods. The proposed hybrid model achieves an average PSNR of 32.8 dB, SSIM of 0.95, and naturalness index improvement of 18%, outperforming standalone deep learning and filtering techniques. The adaptive framework effectively enhances visibility in underexposed, blurred, and noisy conditions, making it ideal for applications in medical imaging, autonomous vision, and surveillance systems.

Traditional recommendation systems primarily rely on user behavior, ratings, and content-based preferences to suggest products or services. However, they often overlook the nuanced emotional context that significantly influences consumer decision-making. This paper proposes a Sentiment-Enhanced Recommendation System (SERS) that integrates sentiment analysis with collaborative and content-based filtering to better capture the affective dimensions of user preferences. By analyzing user-generated content such as reviews, comments, and social media posts using deep learning-based sentiment classifiers, the proposed model quantifies emotional polarity and intensity. These sentiment signals are then incorporated into the recommendation pipeline using hybrid matrix factorization and attention mechanisms, enabling dynamic adaptation to users' emotional states. Experimental evaluations conducted on datasets from Amazon and Yelp demonstrate significant improvements in precision, recall, and user satisfaction scores compared to traditional models. The findings highlight the critical role of emotions in shaping consumer behavior and underscore the importance of affect-aware personalization in modern recommendation systems.

The current exponential growth in the demand for bandwidth is the most urgent challenge for next-generation wireless systems. One of the most appropriate techniques to overcome this situation is Free-space optical (FSO) communication due to the provision of an ample bandwidth. The main disadvantage of FSO communication systems is that the optical beam, propagating through atmospheric turbulence, can be distorted to an unacceptable level. In this work, a 64x64 MIMO-FSO system with Non-Orthogonal Multiple Access (NOMA) and QPSK modulation scheme is assessed. We compare the Bit Error Rate (BER) performance of the system under 4 theoretical turbulence channel models: Log-Normal, Gamma-Gamma, Fisher-Snedecor, Negative Exponential, as well as 4 real seasonal LogCn² datasets. Classical Maximum Likelihood (ML) detection was compared against the deep learning-based ML detection using a Deep Neural Network (DNN) as well as an Autoencoder model. We found that the autoencoder model has outperformed the classical ML detection in terms of BER performance, especially for the weaker user, when NOMA is considered. It was also found that using real datasets that represent real turbulence conditions the proposed system is highly effective and can serve as intelligent fronthaul/backhaul solutions for dense IoT networks such as smart cities, autonomous vehicles, and industrial automation.

The rapid expansion of Internet of Things (IoT) devices has significantly amplified cybersecurity risks, thereby necessitating advanced anomaly detection mechanisms. This research introduces a hybrid detection framework tailored for IoT networks, combining deep learning architectures with bio-inspired optimization techniques. At the core of the framework lies the IoT Autoencoder-Based Feature Extraction Network (IoTAE-FEN), designed to minimize data dimensionality while preserving key discriminative features. To further refine the selected attributes, a Binary Multi-Objective Enhanced Gray Wolf Optimization (BMOEGWO) strategy, modeled on the cooperative hunting behavior of gray wolves, is employed. For the classification phase, Random Forest (RF) is integrated, resulting in the proposed AE-BMOEGWO-RF hybrid model. The effectiveness of this approach was validated on benchmark datasets, including NSL-KDD and TON-IoT. Experimental findings highlight a feature selection accuracy of 96.85% on the TON-IoT dataset and an overall classification performance of 97.81% on NSL-KDD. Comparative evaluations against existing techniques underscore the framework’s superior detection capability, emphasizing its potential to strengthen IoT network security by addressing longstanding challenges in feature extraction and selection for anomaly detection.

When using mammography to diagnose breast cancer, segmenting medical scans is a crucial step.  Accurate segmentation facilitates early diagnosis, which in turn makes it possible to administer individualized treatment plans, ultimately improving patient outcomes. However, for these Deep Learning (DL) models to be trained efficiently and perform optimally, they require access to large datasets.   The lack of sufficient photographs in many publicly available datasets to adequately train deep learning models is a common flaw.   Therefore, this work aims to examine the effects of various affine data augmentations on the Dice Score of a U-NET model utilizing a recently released public dataset of Contrast-Enhanced Spectral Mammography (CESM) images.  The collection consists of 1003 CESM images and matching segmentation masks made by a certified radiologist.   Modifying certain model parameters on the CESM dataset and investigating the impact of single and combination data augmentations on the model's overall performance are the objectives of the study. Images that were moved in the x-direction and sheared vertically were used to train the best-performing model.  On the test set, the model's Dice Score was 56.6%, which was 9% better than the baseline result and showed how crucial data augmentation is when working with small datasets.

The Internet of Things real-time communications depend on a secure stream of data. For the secure communications, a stream cipher with the features of ease and speediness is appropriate. The development and testing of a novel cryptographic algorithm with the goal of enhancing encryption performance. This paper introduces novel A matrix-based nonlinear pseudorandom key stream generation method inspired by the principles of fundamental recursive relationship of Reinforcement Learning, aiming to enhance diffusion and randomness in stream ciphers. We also incorporate the encryption approach based on the Counter based transformation of keystream generation (CBTKSG) method to enhance the speed, which is particularly well-suited for efficiently handling large file sizes since it delivers fast throughput. The technique was thoroughly bench marked and compared to other well-known encryption schemes. Performance has significantly improved without sacrificing security, according to the data. The keystream output was placed through the NIST SP 800-22 statistical test suite to verify its cryptographic strength. It passed every test with high p-values, indicating high randomness quality. The cipher has a strong avalanche effect, meets standard security criteria like IND-CPA and IND-CCA, and resists common cryptanalysis methods including related-key, differential, and linear attacks.

The research-automated segmentation of brain tumors occurs due to the need to enhance diagnosis and/or treatment planning. The existing techniques suffer the effects of scale variation, redundant features, and the high dimensionality that causes ambiguous findings. We suggest the model named MFWX, which unites Multi-Scale CNN, Multi-Frequency Channel Attention (MFCA), Weighted Particle Swarm Optimization (WPSO) to identify features and XGBoost methods to classify them. The Multi-Scale CNN will capture the structure of the tumor at multiple resolutions, MFCA adjusts the features by zeroing in on significant frequency zones and WPSO eliminates redundancy to heavy-hit the strong forecasts of XGBoost. However, MFWX attained 94.2 accuracy and 92.5 Dice on the BraTS-2020 dataset surpassing ResNet50, EfficientNet-B7, and U-Net. It achieved an accuracy of 96.7%, and Dice of 95.1% on BraTS-2018, and performed well on classes of tumors. Ablation experiments proved the necessity of every part. In general, MFWX presents an efficient, clinically meaningful, scalable solution that outsmarts the current segmentation techniques.

The rapid advancement of telecommunication infrastructures and endpoint technologies has led to a significant incorporation of Internet of Things devices in modern lifestyles. IoT involves a wide range of applications, such as connected video surveillance systems for security, wearable body sensors for health monitoring, and temperature sensors for environmental control in agricultural fields. These devices are essential for gathering and transmitting data in real-time. However, data acquisition and transmission processes are often exposed to serious security threats, particularly concerning data integrity, user privacy, and communication reliability. Conventional security mechanisms are typically inappropriate to resource constrained IoT devices. Thus, to overcome these challenges, extensive research has been devoted to developing secure communication frameworks, with a particular focus on robust authentication and key agreement protocols. Authentication is essential to guarantee the legitimacy of the information source, and many proposed AKA schemes rely on asymmetric cryptographic techniques. In this paper, we introduce an Enhanced Lightweight Cryptography-based Authentication Protocol for IoT devices, conceived to meet the computational constraints of IoT devices by employing simple XOR and hashing operations. The protocol enables mutual authentication between IoT devices and routers without the need to share credentials directly. Prior to authentication, an offline registration phase is conducted through an Authentication Server (AS), which generates unique key parameters based on the identifiers of the devices and routers. These parameters are securely distributed to both parties. Authentication is then performed using these pre-shared parameters in a computationally efficient yet secure manner that safeguards against common security threats. Theoretical analysis demonstrates that the proposed protocol is resistant to several common attacks, including man-in-the-middle, impersonation, session key disclosure, replay, and eavesdropping attacks. Additionally, the protocol ensures device anonymity and data privacy while maintaining lightweight performance suitable for constrained IoT environments.

Indoor transportation systems are a key area of development, where Automated Guided Vehicles (AGVs) help to increase efficiency and reduce labor costs. However, high-precision positioning technologies such as LiDAR and GNSS are expensive, making them unsuitable for widespread use. This research has developed a low-cost positioning system for indoor AGVs using multiple sensors, including CCTV, UWB, inertial measurement units (IMUs), and encoders. The experiment was carried out under both static and dynamic conditions. In static tests, Trilateration distance measurements show a lower positioning error than the triangular method, with a maximum error of 1.4464 m (x-axis) and 1.0464 m (y-axis) in dynamic tests. The integrated Encoder and IMU sensor data yielded the lowest error (RMSE = 0.0732 m at 0.4 m/s, 0.0678 at 0.27 m/s), Next is CCTV, while UWB has the highest error rate. The application of a Parallel Sensor Fusion architecture optimized using a Generalized Reduced Gradient (GRG) nonlinear algorithm, significantly reduced localization errors. The RMSE values decreased to 0.0623 m (0.4 m/s) and 0.0411 m (0.27 m/s). The results, in a controlled environment laboratory, indicate that combining multiple sensors will improve the positioning accuracy. Combining the encoder and IMU effectively reduces accumulated errors and increases system stability. While Adjust the weight of the sensor offline, this proposed system offers a cost-effective positioning solution for indoor AGVs, which contributes to the development of affordable and accurate AGV navigation systems.

Solar energy systems play a crucial role in fulfilling global energy needs sustainably; however, their performance is often affected by dynamic environmental factors. This study investigates the use of Artificial Intelligence (AI) for real-time optimization and adaptive control to improve the operational efficiency of solar energy systems. The research specifically addresses output variability arising from fluctuations in solar irradiance, temperature, and panel soiling, limitations that conventional control approaches fail to manage effectively. The primary goal is to develop intelligent AI-based models capable of predicting and automatically adjusting critical system parameters in real time, thereby reducing manual intervention and enhancing operational reliability. Data from a solar photovoltaic (PV) and thermal hybrid testbed in Jodhpur, India were collected over a six-month period. The Indian Meteorological Department provided more than 10000 hourly data samples that included weather and seasonal variations. An NI DAQ system with high-precision sensors was used to measure important parameters such as solar irradiance panel, and ambient temperatures wind speed inclination angle and energy output. For predictive control, the suggested methodology uses a hybrid ensemble framework that combines Extreme Gradient Boosting (XGBoost), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Deep Neural Networks (DNN). In this framework, XGBoost carries out variable importance ranking to determine the dominant influencing factors ANFIS enables adaptive operational control and DNNs forecast energy output. In contrast to previous research that concentrated on distinct AI methods this work presents a cohesive hybrid approach that integrates feature significance analysis adaptive optimization and forecasting accuracy into a single system. The hybrid ensemble model outperforms individual approaches in achieving stable and effective energy generation according to evaluation using RMSE, R2, and MEF metrics. Furthermore, its compatibility with IoT-enabled edge devices underscores its potential for large-scale, real-time, and automated solar energy management within future smart grid infrastructures, advancing global efforts toward sustainable energy transitions.

The generation of cryptographic keys from biometric traits offers a secure alternative to password-based authentication, but is hindered by challenges related to entropy, reproducibility, and adversarial resistance. This work presents a dual-path framework in which a Continuous Thinking Machine Model (CTMM) extracts multimodal embeddings from iris and fingerprint data. Feature vectors undergo projection through principal component analysis and graph-based distance encoding, followed by chaotic sequence modeling with Lorenz-like dynamics and an error-correcting routine to stabilize bitstreams. A secure mixing function consolidates the outputs, while SHA3-512 ensures deterministic expansion. Final passkeys are generated using the Kyber512 post-quantum key encapsulation mechanism (KEM), with neuro-symbolic reasoning applied as a validation layer to enforce entropy, avalanche properties, and inter-user separation. Evaluation confirmed compliance with NIST statistical tests, including monobit, runs, and longest-run assessments, while the system maintained a near-zero false acceptance rate. The originality of this work lies in combining CTMM-driven multimodal feature extraction with a quantum-safe cryptographic pipeline, augmented by neuro-symbolic validation, to establish a reproducible and secure method for biometric passkey generation in high-assurance authentication contexts.