Muhammad Imran Khalid

E‐commerce platforms face an escalating challenge from sophisticated financial fraud and cyberattacks, yet current detection systems suffer from a fundamental trilemma: achieving high detection accuracy, providing expert‐level transparency, and ensuring user comprehensibility simultaneously. Graph neural networks (GNNs) excel at detecting complex fraud patterns through relational learning but operate as black boxes, while existing explainable AI (XAI) methods generate technical attributions that overwhelm non‐expert users. Standalone large language models (LLMs) can produce natural language explanations but lack access to relational graph structures and risk scoring mechanisms. We present MT‐GXL, a novel architectural framework that resolves this trilemma through triple‐granularity orchestration of multi‐modal GNN‐based detection, XAI attribution, and grounded LLM generation. Our framework introduces three key innovations: (1) a triple‐granularity XAI router that dynamically routes explanations based on threat taxonomy; (2) a fraud‐aware homophily index that quantifies closed‐loop fraud ring cohesion; and (3) a grounding mechanism with claim verification that prevents LLM hallucination. Extensive experiments across three benchmark datasets demonstrate that MT‐GXL achieves 97.8% accuracy (AUC = 0.962) on IEEE‐CIS and 98.9% accuracy (AUC = 0.991) on CIC‐IDS2017; on the extremely imbalanced PaySim benchmark (0.13% fraud), where raw accuracy is uninformative, it attains F1 = 0.950 and AUC‐PR = 0.976. Crucially, MT‐GXL is the first unified system to simultaneously satisfy all three trilemma objectives: high accuracy (Acc 0.95), expert transparency (Trans 0.85), and user comprehensibility (Comp 4.0/5.0). A human‐centric evaluation with 70 participants confirms that LLM‐driven explanations significantly outperform traditional XAI, achieving 47% improvement in clarity, 71% in actionability, and 119% in anxiety reduction (). Computational analysis confirms production readiness: 110 ms inference latency, linear scaling to 10M transactions, and a 95% reduction in LLM API cost through tiered deployment.

This study proposes a novel hybrid deep learning framework that integrates a Large Language Model (LLM) with a Transformer architecture for stock price forecasting. The research addresses a critical theoretical gap in existing approaches that empirically combine textual and numerical data without a formal understanding of their interaction mechanisms. We conceptualize a prompt-based LLM as a mathematically defined signal generator, capable of extracting directional market sentiment and an associated confidence score from financial news. These signals are then dynamically fused with structured historical price features through a noise-robust gating mechanism, enabling the Transformer to adaptively weigh semantic and quantitative information. Empirical evaluations demonstrate that the proposed Hybrid LLM–Transformer model demonstrates promising improvement over a Vanilla Transformer baseline, reducing the Root Mean Squared Error (RMSE) by 5.28% (p = 0.003). Moreover, ablation and robustness analyses confirm the model’s stability under noisy conditions and its capacity to maintain interpretability through confidence-weighted attention. The findings contribute to theoretical exploration of LLM–Transformer interactions, offering a step toward explainable, noise-resilient, and semantically enriched financial forecasting systems.

Gaze estimation plays a vital role in human-computer interaction, driver monitoring, and psychological analysis. While state-of-the-art appearance-based methods achieve high accuracy using deep learning, they often demand substantial computational resources, including GPU acceleration and extensive training, limiting their use in resource-constrained or real-time scenarios. This paper introduces GeoGaze, a novel, lightweight, training-free framework that infers categorical gaze direction (“Left”, “Center”, “Right”) solely from geometric analysis of facial landmarks. Leveraging the high-precision 478-point face mesh and iris landmarks provided by MediaPipe, GeoGaze computes a simple normalized iris-to-eye-corner ratio and applies intuitive thresholds, eliminating the need for model training or GPU support. Evaluated on a simulated 1,500-image dataset (SGDD-1500), GeoGaze delivers competitive directional classification accuracy while achieving real-time performance (~66 FPS on CPU), outperforming typical deep learning baselines by more than an order of magnitude in speed. These results position GeoGaze as an efficient, interpretable alternative for edge devices and applications where precise angular gaze is unnecessary and directional intent suffices.

The rapid global deployment of solar photovoltaic (PV) technology presents a significant and often overlooked challenge: the effective management of end-of-life (EoL) solar panels. This issue is particularly acute in developing and emerging economies, where established reverse logistics infrastructure is often lacking. A critical limitation in current academic literature is the oversimplified forecasting of EoL waste streams, which fails to account for the dynamic interplay of socio-economic, policy, and environmental variables. To bridge this gap, we propose a novel decision support system (DSS) for the design of a sustainable reverse logistics network. Our system uniquely integrates a hybrid, multi-factorial forecasting model combining a SARIMAX time series approach with a Gradient Boosting Regressor to provide a robust prediction of waste volume. The output of this predictive engine dynamically informs a multi objective, mixed integer linear programming (MILP) model, which optimizes the network design to simultaneously minimize economic costs and environmental impacts. Our findings demonstrate that this integrated framework offers a methodologically more realistic and adaptable approach for strategic planning than existing models, as validated through simulated scenarios. Our study provides an actionable blueprint for policymakers and industry leaders to proactively build a resilient and circular economy for a sustainable energy future.

This study proposes a novel hybrid deep learning framework that integrates a Large Language Model (LLM) with a Transformer architecture for stock price forecasting. The research addresses a critical theoretical gap in existing approaches that empirically combine textual and numerical data without a formal understanding of their interaction mechanisms. We conceptualise a prompt-based LLM as a mathematically defined signal generator, capable of extracting directional market sentiment and an associated confidence score from financial news. These signals are then dynamically fused with structured historical price features through a noise-robust gating mechanism, enabling the Transformer to adaptively weigh semantic and quantitative information. Empirical evaluations demonstrate that the proposed Hybrid LLM-Transformer model significantly outperforms a Vanilla Transformer baseline, reducing the Root Mean Squared Error (RMSE) by 5.28% (p = 0.003). Moreover, ablation and robustness analyses confirm the model's stability under noisy conditions and its capacity to maintain interpretability through confidence-weighted attention. The findings provide both theoretical and empirical support for a paradigm shift from empirical observation to formalised modelling of LLM-Transformer interactions, paving the way toward explainable, noise-resilient, and semantically enriched financial forecasting systems.

The escalating global prevalence of diabetes renders effective screening for Diabetic Retinopathy (DR) indispensable to prevent irreversible vision loss. Although deep learning models, particularly Convolutional Neural Networks (CNNs), attain diagnostic accuracy comparable to that of human experts, their black-box nature erodes clinical trust. To harmonize accuracy with interpretability, this paper proposes a novel CNN architecture that reformulates DR grading as a regression task. By substituting traditional dense layers with a Global Average Pooling (GAP) layer, our approach substantially reduces model complexity and training time while enabling the generation of Regression Activation Maps (RAMs). These RAMs deliver visual explanations by precisely highlighting the pathological regions that underpin the model's predictions. Evaluated on the Kaggle Diabetic Retinopathy Detection dataset, our model—through the replacement of dense layers with Global Average Pooling—markedly lowers model complexity while delivering diagnostic performance on par with baseline models employing fully-connected layers. The resulting system provides a simpler, more precise, and transparent alternative for automated medical screening, directly associating predictions with clinically relevant features.

Breast cancer is one of the most prevalent and life-threatening diseases among women worldwide. Accurate diagnosis from histopathological biopsy samples is essential, yet manual examination is time-consuming and subject to inter-observer variability, particularly given the shortage of trained pathologists alongside the increasing number of cases. Deep learning, especially Convolutional Neural Networks (CNNs), has emerged as a powerful tool for classifying medical images by automatically extracting discriminative features from raw data. In this study, we investigate the use of the publicly available Breast Cancer Histopathological (BreakHis) image database, which contains benign and malignant samples across multiple magnifications. Our proposed approach extracts isolated image patches, applies CNN-based feature learning and integrates multi-resolution information to improve classification performance. To enhance generalization under limited data, we adopt transfer learning with optimized fine-tuning. Experiments implemented in MATLAB demonstrate that our CNN-based framework achieves higher accuracy than traditional machine learning approaches relying on handcrafted texture descriptors. These findings highlight the potential of CNNs, combined with patch-based multi-resolution analysis, to support pathologists in reliable and efficient breast cancer diagnosis.

Studies have established that an ideal environment is critical for maximizing a student's learning potential. For educators, monitoring student behavior, engagement, and psychological state is essential for ensuring effective instruction. This paper introduces an automated learning analytics system designed to assist teachers by analyzing these parameters and providing actionable feedback. The system utilizes multiple cameras in conjunction with deep learning and computer vision to record and analyze classroom sessions, assessing student movement, gestures, and posture to generate summary reports. The framework integrates several high-performance models to achieve this. For facial recognition, the FaceNet technique is employed, demonstrating 99.63% accuracy on frontal profiles from labeled images and 95.12% on the YouTube Faces DB dataset. Concurrently, a Convolutional Neural Network (CNN) is used for expression monitoring, achieving competitive performance in identifying five distinct emotions. Furthermore, body posture is tracked and analyzed using OpenPifPaf, which shows a high-performance index greater than 0.8 on multiple benchmarks, including the CrowdPose and COCO datasets. By combining these technologies, the system provides a robust and accurate solution for automated, real-time student behavior monitoring.

Millimeter wave (mmWave) communication is a key enabler for future wireless networks, particularly in the context of unmanned aerial vehicle (UAV)-assisted systems. However, mmWave frequencies suffer from significant propagation loss and coverage limitations, especially in hybrid indoor-outdoor environments in addition to the rising secrecy issues. To address these challenges, we propose a hybrid reconfigurable intelligent surface (RIS)-assisted multi-hop secure communication scheme for UAV mmWave networks, where both simultaneously transmitting and reflecting (STAR)-RISs and traditional RISs are deployed to enhance signal coverage and improve system performance. First, we formulate secrecy energy efficiency (SEE) maximization problem. Next, we develop a novel generative artificial intelligence (GAI) and differential privacy (DP)-enhanced soft actor-critic (GAIDP-SAC) framework to jointly optimize precoding at the UAV, the phase shift profiles of hybrid RISs, and the UAV trajectory. The integration of differential privacy ensures secure communication and privacy preservation, while GAI accelerates convergence and improves policy learning efficiency. Simulation results demonstrate that the proposed GAIDP-SAC framework significantly enhances the coverage and security of UAV mmWave networks in hybrid indoor-outdoor environments. Furthermore, the results show that our method effectively handles the complex joint optimization problem, achieving state-of-the-art performance in multi-hop RIS-assisted UAV mmWave communication systems.