University of Sindh Journal of Information and Communication Technology

A Novel Deep Learning Framework Approach For Identifying The Sugarcane Disease

Munazza Rasheed — 2026-07-30

Agriculture plays a significant role in ensuring the survival of the global economy and the growing need for food and resources. Sugarcane is a significant crop in the world and is primarily cultivated for sugar and biofuel production. But crop diseases constantly threaten their production, and crops are notoriously difficult to diagnose in early stages, and environmental conditions and slight visual differences between healthy and diseased leaves make diagnosis a challenge. In this work, a Deep Learning (DL) based method for automatic recognition of sugarcane diseases using sugarcane leaf samples is presented. Some state-of-the-art convolutional neural network (CNN) architectures were considered, as well as a novel, custom-made CNN?model proposed in this paper. The proposed custom CNN has achieved maximum accuracy of 99%, maximum precision of 99%, maximum recall of 99%, and maximum F1-score of 99% which is higher than DenseNet (93%), EfficientNet (96%), InceptionNet (91%), MobileNet (95%) and ResNet (94%). Our results clearly indicate that the proposed model can be effective in aiding farmers in early disease detection to prevent crop loss and increase?output.

An Enhanced Lightweight Trust Framework for Malicious Node Detection in Edge-Enabled IIoT Systems Networks

Atif hayat — 2026-07-30

The Industrial Internet of Things (IIoT) is a technology that allows efficient communication between interconnected devices, sensors, and edge nodes, making advanced industrial automation possible. But edge devices with limited resources are prone to security issues like malicious node behavior, fake recommendations, and attacks on the trust mechanisms. The existing lightweight trust management schemes are mainly based on direct interaction and static trust assessment, which is not suitable for dynamic environments. In this paper, a new lightweight trust management system for enhancing the trustworthiness of IIoT networks is proposed. The proposed approach is a hybrid trust evaluation approach combining direct trust evaluation based on observation and indirect trust evaluation based on recommendation feedback. In addition, for achieving continuous adaptation of trust values based on nodes' behaviors, a dynamic trust adaptation strategy is introduced. The results of the simulations under various network conditions and malicious node densities demonstrate that the proposed framework offers an effective improvement of the accuracy of the Trust Assessment, malicious node detection and defense against dishonest recommendations. The model is stable with as many as 50% malicious nodes in the network and has low computational and communication overhead. The framework offers a trustworthy edge computing trust management solution that is secure, scalable and efficient for IIoT environments.

Parts of Speech Tagging for Handwritten Sindhi Sentence using Deep Learning Models

Marya Soomro — 2026-07-30

Part-of-Speech (POS) tagging is an essential task in Natural Language Processing (NLP) that helps identify the role of each word in a sentence. For handwritten low-resource languages, this task becomes more difficult because of the lack of available datasets and differences in individual writing styles. In this study, a deep learning-based approach is developed for POS tagging of handwritten Sindhi sentences. For this purpose, a dataset of handwritten Sindhi sentences was collected and manually labeled with corresponding POS categories. The prepared dataset was then used for training and evaluating the proposed models. The study applies and compares two deep learning models, Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT), for automatic POS tagging. The models were evaluated using accuracy, precision, recall, and F1-score metrics. Experimental results demonstrate that BERT achieved better performance compared to LSTM due to its ability to capture contextual information from sentence structures. BERT achieved an accuracy of 87.3% while LSTM achieved 85.1%. The proposed work contributes to Sindhi language processing by providing a handwritten dataset and a baseline deep learning approach for POS tagging of a low-resource language.

Web Content Mining Concepts Techniques and a Comparative Methodology for Modern Web Data Extraction

Ali Hassan Sial — 2026-07-30

Rapid development of the World Wide Web (WWW) has led to vast amount of data having diverse structures, both structured, semi-structured and unstructured formats. The challenge has grown critical in the research field to be able to extract meaning and implications from such data. 1.5 Web Content Mining is an important process of applying data mining techniques, text mining and artificial intelligence techniques to convert the unstructured web documents to structured knowledge. This paper summarizes the concepts, techniques and tools used for web content mining, discusses the most recent achievements (with the aid of machine learning and natural language processing). This study is different from the survey metric-based approaches in that it presents a comparative analysis of classical and AI based mining methods, and evaluates them according to their efficiency and their benefits of being accurate and scalable. To implement the proposed lightweight framework, three web content mining tools are evaluated that are widely used in the Web, namely Web Info Extractor, Mozenda and Screen Scraper. In general, the results have broad implications regarding the capabilities and limitations of the existing tools and the better performance and versatility of the new tools based on AI for extraction accuracy. The study results can be of value to the researchers and practitioners involved in building the effective Web content mining systems for the real-world applications like, e-commerce and Cyber security.

A Novel Hybrid Approach with Siamese Networks and Domain-Adversarial Training

Naadiya Mirbahar — 2026-07-30

Balancing data utility and privacy is a persistent challenge in privacy-conscious machine learning, especially when handling personal or biometric data. We propose PES-DAT (Privacy-Enhanced Siamese Network with Domain-Adversarial Training), a novel framework that generates discriminative, privacy-preserving embeddings for clustering tasks. PES-DAT uniquely combines contrastive representation learning with a domain-adversarial classifier, enforced through a Gradient Reversal Layer, to produce task-relevant embeddings that are robust against sensitive attribute leakage. Furthermore, PES-DAT dynamically tunes the privacy-utility tradeoff during training via the Laplace mechanism, adapting privacy guarantees as learning progresses. Evaluations on benchmark clustering datasets, including human activity and campus behavior data, show that PES-DAT achieves up to 10% higher silhouette scores and reduces privacy leakage by 30% compared to state-of-the-art baselines such as AIB and PPGAN. Our results demonstrate that PES-DAT is a robust and flexible solution for privacy-preserving representation learning in both unsupervised and semi-supervised settings

Harmonic Mitigation in Grid-Connected Inverters Using Multi-Resonant Proportional Resonant (PR) Control

waseem javaid soomro — 2026-07-30

Modern microgrids and distributed generation facilities rely heavily on grid-connected voltage source inverters, making power quality compliance at the Point of Common Coupling (PCC) essential. Non-linear load demands introduce low-order harmonic currents that cause voltage distortion, often violating the limits established by the IEEE-519 standard. This paper presents a comprehensive implementation of a multi-resonant Proportional-Resonant (PR) control strategy operating within the stationary ?-? reference frame to eliminate the 5th, 7th, 11th, and 13th harmonic profiles. The complete digital discretization process based on Tustin's transformation is detailed alongside an operational software-based anti-windup clamping loop to protect against saturation during transient grid conditions. Developed within a MATLAB/Simulink environment, the proposed system is validated against composite inductive and capacitive non-linear load configurations. The verification results establish that the closed-loop controller successfully mitigates steady-state harmonic errors, reducing the PCC voltage THD from 10.028% to 3.755%, thereby satisfying the IEEE-519 harmonic limits for grid-connected inverter applications. The proposed structure provides structural simplicity and numerical stability, making it highly suitable for direct integration into real-time digital signal processors.