Thesis (Master of Science/Engineering in Computer Science and Engineering)

Browse

Search Results

Now showing 1 - 10 of 101
  • No Thumbnail Available
    Item
    Explainable dual-teacher knowledge distillation with confidence-aware knowledge filtering for lightweight wound segmentation
    (BRAC University, 2026-04) Dofadar, Dibyo Fabian; Galib, Syed Md.
    Medical image segmentation plays a crucial role in automated wound assessment; however, achieving high accuracy while maintaining computational efficiency and interpretability remains a significant challenge. Even though knowledge distillation can improve overall efficiency in most cases, relying entirely on the teacher(s) can lead to poor predictions. This thesis presents a novel framework that integrates explainable artificial intelligence with an efficient knowledge distillation pipeline for wound segmentation. A dual-teacher knowledge distillation strategy is employed to transfer complementary representations from heterogeneous high-capacity models to a lightweight student network. To enhance the reliability of knowledge transfer, a confidence-aware knowledge filtering mechanism is introduced, which selectively guides the student model using only high-confidence predictions from the teachers. In addition, a lightweight yet effective student architecture, termed SE-HybridConv- Tiny-UNet, is proposed. This model incorporates hybrid convolutional operations and channel attention mechanisms to improve feature representation while maintaining a low parameter count. Extensive experiments conducted on wound segmentation datasets demonstrate that the proposed framework achieves competitive performance compared to larger models, while significantly reducing computational complexity. Furthermore, explainability is incorporated using GradCAM++, enabling visual validation of model decisions and ensuring that predictions are based on clinically relevant regions. The consistency observed between validation and test-time explanations highlights the generalization capability and reliability of the proposed approach. Overall, this work establishes an effective balance between accuracy, efficiency, and interpretability, making it suitable for deployment in real-world, resource-constrained healthcare environments.
  • No Thumbnail Available
    Item
    Probabilistic security mapping of large language model integrations via stochastic Petri Nets
    (BRAC University, 2026) Mohammad, Zaber; Ferdous, Md Sadek; Sadeque, Farig Yousuf
    Large Language Models (LLMs) are becoming increasingly popular for use in modern software systems. However, with increasing popularity, newly introduced security risks have emerged while integrating LLMs in a software system. These security gaps do not align with the traditional cybersecurity framework. To address it, this study specifically focuses on modeling three distinct related threats: prompt injection, context extraction, and Denial of Service (DoS) by resource exhaustion. First, the research maps these three LLM security aspects with the traditional CIA triad (Confidentiality, Integrity, Availability) and maps the system assets with corresponding justifications to show exactly what component of a system is at risk during these specific attacks. After that, the research investigates three distinct and independent threat models across the LLM architecture. First, Prompt Injection is analyzed at the input processing layer to mathematically evaluate Defensive Depth theory. Second, Data Exfiltration is evaluated during output scanning to formalize the Temporal Defense theory. Finally, a Denial of Service (DoS) attack is modeled to validate the Saturation theory. To transition from theoretical risk to measurable impact, an independent threat model is developed using Petri Net diagram to simulate these distinct stages of the LLM pipeline. Mathematical analysis is then conducted using a Continuous-Time Markov Chain (CTMC) and finite queuing theories. Specifically for the DoS evaluation, the adversarial arrival rate (λ) and system processing bottleneck (ρ) are modeled to measure the queue wait times and resource depletion. Across all three threat vectors, Monte Carlo validation is used to ensure the theoretical mathematical calculations match the simulated reality. The result provides a formalized mathematical baseline for each independent vulnerability. The findings demonstrate the exact architectural trade-offs to implement input-layer defensive depth, the temporal cost for output sanitization, and the critical threshold where system queues saturate and drop legitimate requests during a DoS attack. These insights help developer to design more resilient, optimized, and mathematically verifiable security architecture for deployed LLM applications.
  • Thumbnail Image
    Item
    Unlearning to protect: a distilled reinforcement learning framework with privacy-preserving feature unlearning and XAI for IoT security
    (BRAC University, 2025-11) Hasan, MD. Nahid; Alam, Md. Golam Rabiul
    Botnets pose a significant cybersecurity threat, enabling attacks such as DDoS, data theft, and service disruptions on IoT devices. These devices often lack built-in botnet traffic filtering, leaving them highly exposed. Existing AI-based solutions improve detection capabilities but have limitations: (i) they are too heavy for IoT deployment, and (ii) they lack unlearning capabilities to forget sensitive or outdated features without retraining. To address these challenges, we propose DiRLU, a lightweight, reinforcement learning driven framework, while ensuring privacy by selectively unlearning sensitive or outdated features without requiring retraining. The framework leverages knowledge distillation to transfer knowledge from a teacher model into a lightweight student model, with both models trained using A2C. A post-hoc unlearning mechanism modifies weights to remove targeted features, while restored features show negligible performance loss, confirming reversibility. Unlike many benchmark models that used only 5% of the BoT-IoT dataset, this research leverages 25%, allowing us to develop a strong teacher model. Both the teacher and student models were trained using the A2C reinforcement learning algorithm, achieving impressive results, with the student model achieving 99.60% accuracy and a 99.80% F1 score. To enhance transparency, we integrated Explainable AI (XAI), particularly LIME, which helps interpret the model’s decisions and identify the key features influencing its predictions. Additionally, DiRLU requires only 2,370 FLOPS, approximately 3.87× more efficient than the state-of-the-art model, highlighting its efficiency for edge deployment. DiRLU combines efficiency with privacy, aligning with GDPR standards (right to be forgotten) to provide practical IoT security solution. By combining knowledge distillation, feature unlearning and XAI, this research not only strengthens botnet detection but also sets new standards for security, interpretability, and data privacy in cybersecurity.
  • Thumbnail Image
    Item
    Evidence-based workplace harassment guidance system: transforming #MeToo narratives into actionable knowledge through machine learning
    (BRAC University, 2026-01) Kabir, Mashphey Bintey; Alam, Md. Golam Rabiul
    Workplace harassment remains a pervasive global issue, yet victims often lack ac- cess to evidence about what actually happens when people in similar situations take action. Traditional resources offer generic procedural advice without outcome data, leaving individuals to make consequential decisions with incomplete informa- tion. This study presents an evidence-based workplace harassment guidance system that transforms 15,835 #MeToo narratives into personalized, actionable guidance grounded in documented outcomes from comparable cases. The system addresses a fundamental information asymmetry: while organizations accumulate knowledge about harassment cases, individual victims rarely know what outcomes others in similar situations experienced. By analyzing patterns across thousands of documented experiences, the system identifies a user’s specific vulner- ability profile, retrieves semantically similar historical cases, and presents evidence- based guidance including proven successful action sequences, outcome statistics, and high-impact actions that correlate with positive results. The guidance generation pipeline employs SimCSE-BERT embeddings for semantic similarity, multi-label vul- nerability detection leveraging these embeddings to identify seven co-occurring risk factors, and outcome pattern analysis across fifteen outcome categories—supported by BERT sentiment analysis (98.1% accuracy) and HDBSCAN clustering for eval- uation. The empirical analysis reveals sobering realities: negative outcomes predominate across all vulnerability types, institutional inaction occurs in 16.9% of cases, and harasser accountability remains rare at 3.6%. Rather than offering false reassur- ance, the system presents these evidence-based statistics to enable informed decision- making. Professional evaluation with eleven practitioners from HR, legal, mental health, and other sectors—91% with direct harassment case experience—validates that the guidance meets practical standards for appropriateness (M=4.09/5), use- fulness (M=4.09/5), actionability (M=3.91/5), and safety (M=3.73/5), with 91% endorsement and 73% rating the approach superior to typical harassment resources. This research demonstrates that machine learning can provide meaningful support for sensitive domains when developed with rigorous professional validation and com- mitment to user safety.
  • Thumbnail Image
    Item
    A lightweight time-series analysis model through multi-teacher knowledge distillation for food price forecasting
    (BRAC University, 2026-01) Zaman, Shifat; Alam, Md. Golam Rabiul
    Accurate food price forecasting is critical for food security planning, particularly in developing nations like Bangladesh where price volatility can significantly im-pact vulnerable populations; however, existing deep learning approaches for time series forecasting often require substantial computational resources, limiting their deployment in resource-constrained environments. This study presents a novel multi-teacher knowledge distillation framework that transfers knowledge from an ensem- ble of three distinct teacher architectures—DLinear, PatchTST, and N-BEATS— trained on World Food Programme (WFP) Bangladesh commodity price data from the Dhaka Division to compact student models (MLP, GRU, KAN) through a multi- component distillation loss comprising prediction-level matching, feature-level alignment, and price-difference learning, with an uncertainty-weighted mechanism that focuses training on confident teacher predictions while dynamically weighting teacher contributions based on validation performance. Experimental evaluation on four food commodities (Lentils, Oil, Rice, and Wheat flour) with a 6-month input window demonstrates that the proposed approach achieves a Mean Absolute Error (MAE) of 1.959 BDT/unit with a Mean Absolute Percentage Error (MAPE) of only 3.73%, representing a 37% improvement over the supervised learning baseline, 69% improvement over traditional ARIMA, and 81% improvement over LSTM baselines, with the three-teacher ensemble distilled to an MLP student achieving the best results and outperforming all single-teacher and two-teacher configurations. The resulting student model requires only 200K parameters (compared to over 1M in the teacher ensemble) and achieves inference in sub-millisecond time on standard CPU hardware without GPU acceleration, enabling deployment in humanitarian field offices with limited computational infrastructure. This study contributes a reproducible, configuration-driven framework for knowledge distillation in time series forecasting, demonstrating that sophisticated ensemble-level accuracy can be achieved with lightweight models suitable for resource-constrained field deployment in food security applications.
  • Thumbnail Image
    Item
    Integrated sensing and communication for next generation wireless systems
    (BRAC University, 2026-01) Zabin, Rifat; Alam, Md. Golam Rabium
    Wireless communication systems inherently interact with their surrounding physi- cal environment, causing transmitted signals to carry information beyond the data they are intended to convey. This observation forms the basis of Integrated Sensing and Communication (ISAC), where sensing and communication are jointly sup- ported using the same wireless infrastructure. Despite its promise, realizing ISAC in practical systems—particularly with commodity Wi-Fi—remains challenging due to high-dimensional channel representations, compute intensive channel sounding, hardware imperfection, and strong variability across environments. This thesis ad- dresses these challenges by developing practical, learning-driven ISAC frameworks that explicitly exploit the structural properties of wireless channels across time, fre- quency, and space. The thesis is organized around two complementary thrusts. First, communication-aided sensing is realized through PULSE , a lightweight framework that transforms raw Channel Frequency Response (CFR) measurements into compact, physics-aware temporal representations for sensing inference and keep- ing it generalized across domains. Second, sensing-aided communication is enabled through ELF , a scalable channel feedback framework that reformulates multi- antenna channel reporting as a structural inference problem, by feeding back only a small number of representative subchannel embeddings. Experiments show that PULSE achieves over 99% sensing accuracy across diverse activities while re- ducing the effective input dimensionality by approximately 85% and maintaining low inference latency suitable for real-time edge deployment. PULSE generalizes to unseen environments and devices using as little as 5 seconds worth of labeled data, outperforming state-of-the-art Wi-Fi sensing frameworks under domain shifts. On the other hand ELF reduces feedback overhead by up to 96% relative to dense sub- carrier reporting, while preserving near-identical communication reliability. Com- pared to standard IEEE 802.11ax explicit feedback, ELF achieves up to a 25× reduction in feedback size across wide bandwidths, and its feedback cost remains effectively independent of antenna count, enabling scalable operation in large Multiple-Input Multiple-Output (MIMO) systems. Extensive simulations and real- world Wi-Fi testbed evaluations demonstrate that the proposed approaches achieve strong sensing accuracy, robust cross-domain generalization, and substantial reduc- tions in communication overhead. Collectively, this thesis establishes a unified and deployable ISAC framework for commodity Wi-Fi systems, highlighting the role of channel structure in enabling efficient and reliable sensing and communication under real-world constraints.
  • Thumbnail Image
    Item
    Real-time scene description and interpretation using zero-shot learning and prompt-engineered vision-language models
    (BRAC University, 2024-08) Apurba, Md Shifatul Ahsan; Anwar, Md. Tawhid
    Real-time scene description and interpretation are essential for diverse applications such as surveillance, interactive media, and automated video analysis. However, most existing methods rely heavily on large-scale labeled datasets, thereby limiting their adaptability in dynamic or previously unseen scenarios. In this work, we propose a novel mixed-model framework that integrates Vision-Language Models (VLMs), Large Language Models (LLMs), and lightweight object detection networks (e.g., MobileNet-SSD) through advanced prompt engineering. By leveraging zeroshot learning, our approach generates contextually rich scene descriptions without requiring domain-specific or task-specific retraining. The prompt engineering component reduces sensitivity to subtle linguistic variations, enhancing robustness across diverse input formulations. Furthermore, the lightweight detector ensures real-time performance, making the framework suitable for resource-constrained environments. To address ethical and fairness considerations, we incorporate bias mitigation strategies that limit the propagation of harmful stereotypes from large-scale pretraining data. Experimental evaluations on multiple open-domain scenarios demonstrate that our system offers reliable and efficient scene interpretation, maintaining high accuracy in challenging conditions where traditional supervised techniques often fail. This research paves the way for more flexible, scalable, and responsible visionlanguage systems capable of operating effectively in real-world, zero-shot contexts.
  • Thumbnail Image
    Item
    WHISNER-BN: parameter-efficient end-to-end spoken named entity recognition for low-resource languages with morphology-aware alignment
    (BRAC University, 2025-12) Imran, Shah; Alam, Md. Golam Rabiul
    Named entity recognition from speech remains underdeveloped for low-resource languages such as Bengali, despite its importance for voice search, conversational AI, and accessibility. This thesis investigates Bengali spoken NER through three paradigmsdiscriminative structured prediction, generative multi-task learning, and multimodal instruction-tuned approaches-with the primary contribution whisNer-bn, a parameterefficient architecture employing Low-Rank Adaptation of Whisper encoders (7.1% trainable parameters), BiLSTM contextual encoding, and Conditional Random Field decoding with explicit BIO constraints. To enable end-to-end training, we introduce bnSpAligner, a morphology-aware forced alignment algorithm achieving 78% tokenlevel accuracy for Common Voice and 83% for SUBAKKO through adaptive thresholding and phonetic equivalence classes, and BSSC-Annotator, a multi-agent framework leveraging cross-lingual transfer to produce 50.57 hours of annotated Bengali speech comprising 38,504 entities across 36,237 utterances at less than 10% of manual annotation cost. Evaluated on 5,000 samples from combined test partitions using models trained on 20% of available data (4,989 samples), whisNer-bn achieves 0.681 F1, outperforming generative multi-task approaches by 15.6 percentage points and cascaded ASR-NER pipelines by 5.8 percentage points. The results demonstrate that discriminative structured prediction with joint acoustic-linguistic modeling provides superior inductive biases for entity recognition in morphologically complex languages, establishing the first systematic benchmark and transferable methodology for Bengali spoken NER with implications for thousands of underserved languages.
  • Thumbnail Image
    Item
    Causal inference in depression: understanding beyond correlation
    (BRAC University, 2025-10) Shoumo, Syed Zamil Hasan; Alam, Md. Golam Rabiul
    Depression remains one of the most pressing mental health concerns worldwide, intensified further by the socioeconomic and psychological impacts of the COVID-19 pandemic. Understanding the underlying mechanisms that contribute to depressive symptoms has therefore become a major research priority. While traditional statistical and machine learning models have been effective in identifying associations between risk factors and depression, they often fail to distinguish correlation from causation. Explainable Artificial Intelligence (XAI) methods, such as SHAP and LIME, have improved transparency by revealing which features influence model predictions; however, they remain fundamentally correlational and do not provide insight into the true causal pathways that drive depressive outcomes. To address this limitation, this research integrates machine learning, explainable AI, and causal inference to explore the causal factors behind depression among the Bangladeshi population during the COVID-19 pandemic. Using XGBoost for predictive modeling, the study first evaluates the relative importance of features through gain-based measures and SHAP value interpretation. Subsequently, a causal inference framework is constructed following Judea Pearl’s principles to identify and estimate direct causal effects using the backdoor adjustment method with a generalized linear model estimator. Finally, a combined feature-selection pipeline is developed that retains causally significant variables and iteratively removes weakly correlated ones to test their joint predictive strength. The results reveal that while several factors exhibit high correlation and feature importance in black-box models, only a subset demonstrates genuine causal influence on depressive outcomes. This distinction underscores the importance of causal reasoning in mental health analytics. Overall, the study establishes that integrating causal inference within predictive frameworks not only enhances interpretability and trustworthiness but also provides a clearer understanding of which factors can truly influence and potentially mitigate depression.
  • Thumbnail Image
    Item
    Chained semantic retrieval for rare disease and gene identification using clinical phenotype
    (BRAC University, 2025-10) Saiful, Md.; Alam, Md. Golam Rabiul
    Accurate identification of rare diseases and associated genes from patient phenotypic data represents a critical challenge in precision medicine and genomic research. Current diagnostic approaches face substantial limitations in scalability, interpretability, and real-time processing when integrating heterogeneous phenotypic and genotypic databases. This study presents a novel artificial intelligence-driven system employing chained semantic retrieval and knowledge graph integration to identify rare diseases, genes, and Human Phenotype Ontology (HPO) terms from patient phenotype descriptions. The methodology leverages sentence transformers (based on Bidirectional and Auto-Regressive Transformers) for generating vector embeddings, Facebook AI Similarity Search (FAISS) for efficient similarity computation, and a fine-tuned Llama 3.2 model integrated with a Retrieval-Augmented Generation (RAG) pipeline. The system implements a tiered scoring mechanism that chains retrieval across Human Phenotype Ontology, disease, and gene databases to progressively refine predictions through contextual enhancement. Evaluation on 50 patient phenotypes with confirmed Duchenne Muscular Dystrophy diagnosis, consisting of 30 true positive and 20 true negative cases, demonstrated strong performance with 28 true positives, 17 true negatives, 2 false negatives, and 3 false positives in top-ten retrieval results. The system achieved 90 % overall accuracy, 93.3 % recall, 85 % specificity, 90.3 % precision, and an F1-score of 91.8 %, with an average computational efficiency of 3.2 seconds per response. The proposed framework effectively addresses critical gaps in cross-database integration while maintaining interpretability through tiered confidence scoring for clinical decision support applications.