http://hdl.handle.net/123456789/10361 2026-04-04T12:14:12Z 2026-04-04T12:14:12Z Hassan Ayaz, 01-241222-010 http://hdl.handle.net/123456789/19220 2025-03-11T05:50:51Z 2025-01-01T00:00:00Z

Mobile Network Traffic Classification and Prediction Hassan Ayaz, 01-241222-010 Call Detail Records (CDRs) from mobile networks offer rich insights into network performance and user behavior. In this study, we analyze CDR data from Telecom Italia, encompassing spatiotemporal patterns across Milan, segmented into a 100x100 grid with each cell covering 0.3 kilometers. By analyzing the spatiotemporal dynamics of CDR data, we classify the network traffic into four categories: highest, high, moderate and low with high network traffic regions predominantly located in the city center. After network traffic classification, we predict future traffic patterns. We utilize automated machine learning (AutoML) tools and the state-of-the-art TimeGPT model for network traffic forecasting. Comparative analysis reveals that AUTOML performs better then TIMEGPT, delivering superior prediction and performing better on the various evaluation metrices resultantly capturing complex temporal and spatial relationships in the data. These predictive capabilities enable dynamic resource allocation, enhanced congestion management and improved network efficiency. Our findings underscore the potential of both AUTOML and TimeGPT, to some extent AUTOML appears to be more scalable and adaptable solution for mobile network traffic classification and forecasting, marking a significant advancement in network planning and optimization for urban environments Supervised by Dr. Kashif Sultan

2025-01-01T00:00:00Z Mahrukh Shakoor, 09-241241-001 http://hdl.handle.net/123456789/20458 2026-01-09T05:30:57Z 2025-01-01T00:00:00Z

Psychological Disord Ers Diagnosis Framework Using Reflective Listening and Generative AI Mahrukh Shakoor, 09-241241-001 Mental health disorders such as Generalized Anxiety Disorder (GAD), Major Depressive Disorder (MDD), and Borderline Personality Disorder (BPD) are rising globally, and this causes access to clinically trained professionals limited. To address this gap, our research proposes an innovative Psychological Disorder Diagnosis Framework Using Reflective Listening and Generative AI, integrating multimodal, natural language processing, DSM-5 diagnostic logic, knowledge-graph based reasoning, and validated psychometric assessments under one system. The system conducts a multi-stage evaluation process that starts with two rounds of conversational screening, followed by NLP-driven symptom extraction, keyword clustering, probability scoring, and DSM-5 validation checks. The reflective listening technique is used in fine-tuned LLM-based empathetic dialogue to enhance emotional understanding and user comfort. A secondary diagnostic stage executes standardized clinical tests, including the Penn State Worry Questionnaire (PSWQ) for GAD, Beck Depression Inventory (BDI) for MDD, and McLean Screening Instrument (MSI) for BPD, ensuring that AI predictions are verified against clinical standards. A comprehensive Gold dataset was created through system-generated session logs, enhanced with DSM-5 ground-truth labels, validated psychometric scores, corrected knowledge-graph structures, and reflective-listening quality ratings provided by licensed psychologists. This dataset supports the fine-tuning of a generative model capable of producing empathetic reflections, accurate disorder predictions, DSM-5 aligned reasoning, and appropriate treatment suggestions. The architecture combines a user-friendly multilingual interface, text and audio inputs, an NLP and reflective-listening engine, a clinical rule-based inference module, and a visualization layer that provides probability scores, knowledge graphs, and diagnostic summaries. Results determine that the proposed framework enhances diagnostic transparency, cultural adaptability, and clinical validity compared to traditional sentiment-based mental health models. The system achieves high stability between AI predictions and clinician evaluations, while reflective listening significantly improves user engagement and emotional coherence. This research contributes an explainable, empathetic, and clinically grounded AI diagnostic framework, establishing a foundation for next-generation intelligent mental health assessment systems suitable for real-world psychological support and early screening applications. Supervised by  Dr. Tamim Ahmad Khan

2025-01-01T00:00:00Z Faiza Asad, 01-241212-017 http://hdl.handle.net/123456789/16990 2024-02-26T08:03:08Z 2024-01-01T00:00:00Z

Value Cognizant Test Case Prioritization for Regression Testing Using Integration Based Test Data Faiza Asad, 01-241212-017 Test case prioritization (TCP) techniques are beneficial in regression testing tasks as they allow us to find test cases of high priority based on test data and metrics of previous executions. Existing TCP approaches can be categorized as value-neutral or value-neutral approaches. The value-neutral approaches are those where the underlying assumption is that all test cases, requirements, and faults have the same severity. Value-based approaches are approaches where a particular value of test cases, requirements, and faults is calculated and used in the TCP process. However, the majority of available methods and approaches are value-neutral where we propose a requirements’ business value quantification framework. This framework is proposed in such a manner that the impact of individual developers’ experience or the involved engineers’ professional experience is minimal. We also consider test case features and aspects such that we do test case prioritization considering the business value of a requirement as well as the test case execution data concerning faults found and their severity. We validate our framework using a previously established framework evaluation method (KANO), TOPSIS, and a case study. We use previously established evaluation criteria while validating through case studies, such as Average Percentage of Faults Detected (APFD), and Normalized Average Percentage of Faults Detected (NAPFD) for performance evaluation. We, finally, use white-box test data and provide a mechanism through extracting elements from Message-Method paths through the system under test (SUT) so that we may allow testers and developers to consider whitebox test cases for TCP as well. We make use of previously published criteria of Average Percentage of Element Coverage (APEC) for coverage analysis which is a novel contribution to our investigation. We compare our results with industrial benchmark results and we outperform them. Supervised by Dr. Tamim Ahmed Khan

2024-01-01T00:00:00Z Sadia Rashid, 01-241211-009 http://hdl.handle.net/123456789/16988 2024-02-26T07:47:40Z 2024-01-01T00:00:00Z

Android Malware Detection Using Dynamic Feature Analysis Sadia Rashid, 01-241211-009 Android is a widely used operating system but with its success, it also faces issues with malware threats. With advanced technology, malware is also becoming complex making it challenging to detect. Android malware detection techniques are used to detect malware on Android operating systems. There are mainly two types of detection techniques which are static and dynamic analysis. This thesis is focused on the dynamic analysis of Android malware. The use of machine learning or deep learning for malware detection requires datasets. These datasets Malware Researchers developed many machine learning models and deep learning models to detect Android malware considering static as well as dynamic features-based datasets. We consider features extracted from system executions and we perform multi-class classification of malware categories in this study, by considering all malware classes (Adware, Backdoor, File Infector, PUA, Ransomware, Riskware, Scareware, Trojan, TrojanBanker, Trojan-Dropper, Trojan-SMS, Trojan-Spy and zero-day) by using deep learning algorithm as it outperforms machine learning in high dimensional data. We use the CIC-AndMal-2020 dataset, the newly developed dataset for multi-class classification. It consists of 13 prominent malware classes and 141 features. We also perform statistical analysis on the dataset (p-value analysis and correlation) to identify the relationship between features and statistical analysis of the model (bias and variance) to arrive at the optimal model. We evaluate the proposed algorithm using performance metrics i.e. accuracy, precision, recall, ROC-AUC analysis, and F1-Score and, finally, compare our results with existing studies. Our results outperform previous dynamic analysis results. Supervised by Dr. Tamim Ahmed Khan

2024-01-01T00:00:00Z