Computer Engineering Commons^™

Retrieval-Augmented Generation (Rag) Chatbots: A Comparative Study Of Claude, Gpt-4o, Deepseek, And Llama, Kalindi Vijesh Parekh

Master's Projects

The use of Retrieval-augmented generation (RAG) in chatbot platforms has transformed academic spaces by significantly improving information accessibility. RAG has become a viable approach to upgrading Large Language Models (LLMs) with external knowledge access in real time. With the growing availability of advanced LLMs such as GPT, DeepSeek, Claude, Gemini, and Llama, there is a growing need to compare RAG systems based on different LLMs. This study compares the responses of four different RAG chatbots using popular LLMs against a uniquely designed evaluation dataset. Specifically, the study compares the responses and performance of closed-source (GPT-4o and Claude) and open-source models …

Smart: Semantic Mapping And Analysis For Regional Terrain Using Multi-Scale U-Net And Topsis, Rashmi Sonth

Master's Projects

Accurate land use classification is the backbone for urban planning. But with poor quality satellite images, varied landscapes and structures which are changing faster than ever, it becomes a challenge to define clear boundaries and hence to urban planning. This research explores the application of deep-learning model for land use classification and asses the suitability of the land. The proposed model combines a multi-scale U-Net architecture with Transformer blocks applied on a multi-spectral satellite images that improves the semantic segmentation greatly across the urban and rural regions. Additionally, a patch-wise segmentation is applied to overcome the common problem of feature …

Visionmate: Ai-Powered Image Captioning Web Application, Sai Anoushka Kokku

Master's Projects

VisionMate is a web application that generates captions for camera-captured images. It is designed
to assist users with visual impairments by converting visual input into spoken and written text. The
application uses the GIT-base model from Hugging Face, which processes the image and returns a
descriptive caption. Users can take a picture using the device camera—either via webcam on
desktop or the native camera interface on mobile. The app provides audio output using the
SpeechSynthesis API and uses full-screen tap interaction to simplify accessibility.

The frontend is implemented in React.js, and the backend is built with FastAPI. The backend calls …

Adaptive Cobot Interaction Via Smartwatch Data Fusion For Car Assembly Automation, Riddhik Tilawat

Master's Projects

In modern car manufacturing, collaborative robots (cobots) work with human operators during shared workcell interactions to maximize production speed and flexibility. Collaboration between humans and robots is safe and effective only when operator intent recognition via a single wrist-worn inertial measurement unit (IMU) is accurate and low-latency. This thesis develops an IMU-only intent recognition pipeline, and is evaluated on three datasets: the public OPPORTUNITY dataset, the Sony Smartwatch Gesture dataset and a custom Samsung Galaxy watch 6 dataset. The proposed framework leverages five step sequence-to-label problems which are stepwise posed as data streams transforming raw IMU data into trainable tensors. …

Real-Time Adaptive Framework For Topic Modeling In Social Engineering Attacks, Manav Bhasin

Master's Projects

Detecting social engineering attempts is crucial for security, as these threats are becoming more frequent and increasingly exploit human vulnerabilities. This research focuses on topic modeling using conversational data from Kevin Mitnick’s ”The Art of Deception” with dialogues that illustrate various social engineering strategies. The dataset comprises manually extracted and synthetically augmented conversations to ensure natural dialogue flow. Two methodologies are presented for utterance-level and global topic extraction: prompt engineering leveraging OpenAI’s GPT-4o-mini, characterized by few-shot learning and chain-of-thought prompting, and Quantized Low Rank Adaptation (QLoRA) utilizing Mistral’s 7B instruct model for efficient fine-tuning. Through experimentation and evaluation, this study …

Augmenting Missing Sensor Data For Robust Human Activity Recognition, Suryakangeyan Kandasamy Gowdaman

Master's Projects

Applications of ubiquitous computing, including health monitoring, sports analytics, and ambient-assisted living, rely on Human Activity Recognition (HAR) using wearable sensors. However, model robustness is challenged by missing sensor values, class imbalance, inter-subject variability, and temporal noise. This work proposes a complete HAR pipeline that addresses these challenges through sampling, time-series augmentation, dynamic feature handling, and GAN-PCA-based imputation. Built on the DeepSense architecture, the model integrates convolutional feature extraction with bi-GRUs for temporal modeling. The system is evaluated using 5-fold cross-validation, subject-aware holdout, and LOSEO strategies on the Opportunity dataset. Results demonstrate consistent accuracy across folds and strong generalization to …

Effects Of Data Augmentation On Sponge Identification Using Computer Vision Models, George Ku

Master's Projects

Coral reefs can be primarily found in tropical and sub-tropical regions of our oceans, providing a thriving habitat for millions of species. Marine sponges, which can be frequently found in coral reefs, play a critical role that contributes to the maintenance of these ecosystems, including the recycling of nutrients through water filtration. However, rising ocean temperatures and acidification due to climate change have resulted in the bleaching and death of coral reefs worldwide. In order to preserve these reefs and the sponges that depend on them, scientists have been performing studies on their biodiversity. This includes collecting numerous images of …

Detecting Ai-Generated News Articles Using Unsupervised Machine Learning Algorithms, Lilou Sicard-Noel

Master's Projects

The widespread adoption of Large Language Models (LLMs) has revolutionized text generation and heightened concerns over misinformation and the erosion of journalistic integrity. Detecting AI-generated text is critical to addressing these challenges, yet current detection methods face adaptability, scalability, and accuracy limitations. This research paper uses machine-learning techniques to explore the classification of human and AI-generated articles, including a mix of human and AI-written content. The primary focus is on evaluating the effectiveness of clustering algorithms (K-Means and Agglomerative Clustering), auto-encoders, and Part-Of- Speech Tag Transition Matrix Log-Likelihood for distinguishing between AI-generated and human-written texts. Our findings reveal that while …

Mycelia: Cross-Chain Data Oracle Using Frost Signatures, Bala Komatireddy

Master's Projects

The interoperability of heterogeneous blockchain networks is the basis for the widespread application of blockchains in various fields. Cross-chain data oracles play a significant role in enabling distributed applications to exchange data and assets across different blockchains, thereby greatly enriching and expanding the application scenarios and use of blockchains. With the continuous advancement of blockchain technology, more and more researchers and industry participants have begun to focus on developing cross-chain data oracles. Current cross-chain data oracles face issues with trust, as they rely on centralized intermediaries or limited validator networks, increasing the risk of manipulation or single points of failure. …

An Evidence-Based Approach To Predicting Pancreatic Ductal Adenocarcinoma, Surya Teja Nalluri

Master's Projects

Pancreatic ductal adenocarcinoma (PDAC) is a complex disease with hidden clinical indicators, so a reliable diagnosis of PDAC requires high precision and sophisticated analysis. Traditional probabilistic methods often rely on making unwarranted assumptions or undesirable approximations about probabilistic estimates, limiting their ability to provide the precision needed for correct diagnosis and treatment planning. In contrast, Dempster–Shafer Theory offers a formal framework for integrating uncertain and potentially conflicting evidence. This makes it well-suited for analyzing incomplete and ambiguous data typically associated with PDAC. By employing an evidential reasoning (ER) model based on Dempster-Shafer Theory, this approach systematically combines and evaluates imperfect …

Transformers In Time-Series Forecasting: Enhancing Robustness Via Dynamic Attention Mechanisms, Kush Patel

Master's Projects

Transformer architectures have emerged as powerful tools for time series forecasting, excelling at capturing complex temporal dependencies across multivariate inputs. However, these models are highly susceptible to adversarial attacks such as the Fast Gradient Sign Method (FGSM) and Basic Iterative Method (BIM), which can significantly degrade predictive performance through small, targeted perturbations. This work integrates dynamic attention mechanisms, adaptive masking modules that introduce controlled variability into attention pathways, into a transformer forecasting model to enhance robustness against such attacks. Using two distinct datasets, we compare the performance of a standard transformer and a dynamic attention-enhanced transformer under both clean and …

On The Adversarial Robustness Of Quantized Neural Networks Against Common Adversaries In Time-Series Forecasting, Maanak Arora

Master's Projects

Real-world edge applications now use modern machine learning models which require both resource efficiency and robustness against adversarial threats. Deep neural networks which include time series forecasting models still face risks from adversarial perturbations while quantization techniques used for memory and compute efficiency create unpredictable robustness challenges. This project investigates the adversarial resistance of Long Short-Term Memory (LSTM) models after applying post-training quantization at three different precision levels: 16-bit floating point (FP16), 8-bit integer (INT8) and custom 4-bit quantization. The Jena Climate dataset serves as our main benchmark for training a fullprecision LSTM model followed by multiple quantization strategies which …

Retrieval-Augmented Generation For Survival Analysis In Cancers: Methods And Evaluation On The Surveillance, Epidemiology, And End Results Database, Jyothi Vaidyanathan

Master's Projects

Healthcare is one of the most important fields that benefits from advancements in Artificial Intelligence (AI). From classic models like linear regression to cuttingedge transformers, AI is applied across various healthcare subdomains, such as drug discovery, predictive analytics, and personalized medicine, to name a few. These techniques enable medical practitioners to make more informed decisions, significantly improving both the speed and accuracy of diagnoses and treatments. Machine learning has played a transformative role in oncology, especially in areas like early detection, diagnosis, treatment planning, and patient monitoring, by analyzing medical images, clinical information, genomic data, sensor information. Our research aims …

Medilightrag: A System For Medical Query Response Using Fine-Tuned Llms And Graph Based Retrieval, Rajiv Karthik Reddy Kodimala

Master's Projects

The exponential increase in medical data has created a greater demand for precise and efficient information retrieval systems. Existing Large Language Models (LLMs) face domain-specific difficulties such as sophisticated medical jargon, situational comprehension, and the continual advancement of healthcare knowledge. To tackle these challenges, we present MediLightRAG, an innovative two-stage system which integrates parameter-efficient fine-tuning of Large Language models with LightRAG’s graph-based retrieval. The first stage focuses on enabling accurate resource-efficient model adaptation for the medical domain through QLoRA fine-tuning. In the second stage, LightRAG’s two-tiered retrieval architecture that combines graph-based indexing with dynamic knowledge retrieval is employed to enhance …

Moving Target Defense With Quantized Morphence: Defense Quantification Against Common Adversaries In Image And Time Series Problems, Rithika Dhamala

Master's Projects

In recent years, the vulnerability of deep learning models to adversarial attacks has emerged as a serious threat, particularly in domains where reliability and robustness are critical. This project builds upon the Morphence framework, a Moving Target Defense (MTD) strategy designed to counter adversarial threats by maintaining a dynamic pool of models and introducing randomness at inference time. While Morphence was originally developed for image classification tasks, this work not only reproduces the original architecture using MNIST and CIFAR-10 datasets but also extends the core principles to an entirely new domain: time series forecasting. The project proposes a unified defense …

Energy Considerations For Large Pre-Trained Neural Networks, Leo Mei

Master's Projects

In recent years, neural network models have achieved phenomenal performance due to the increasing parameters and complexity of model architectures. However, these advancements come with high environmental costs as they require massive computational resources and consume substantial amounts of electricity, leading to high carbon emissions. Previous studies have demonstrated that substantial redundancies exist in large pre-trained models, and reducing these redundancies through compression would not compromise model performance. While these studies focused on retaining comparable model performance, the direct impact of compression on energy consumption when training models appears to have received little attention. By quantifying the energy usage associated …

St-Hybrid: Dynamic Graph Learning With Multi-Scale Spatio-Temporal Attention For Traffic Forecasting, Dhe Yeong Ewaza Tchalla

Electronic Theses and Dissertations

Accurate short-term traffic forecasting is central to modern Intelligent Transportation Systems, supporting route guidance, adaptive signal control, and incident response. Yet producing reliable predictions remains difficult because traffic is highly non-stationary. The relationships among roadway sensors shift during congestion, incidents, weather changes, or fluctuations in demand, and the temporal structure of traffic spans several scales from abrupt minute-level variations to broader daily and weekly rhythms. Models that rely on fixed spatial graphs or a single temporal scale tend to miss these evolving and layered dependencies. This thesis addresses these challenges by developing a graph-learning framework that adapts to changing traffic …

Power Saving In Open Ran By Using Advanced Cpu Scheduling Algorithm, Saish Urumkar, Sachin Sharma

Conference papers

Open RAN is an emerging wireless technology that is gaining significant attention for its potential to enable flexi- ble, cost-efficient, and interoperable networks. Reducing power utilization in Open RAN, particularly for 5G base stations (gNodeBs) deployed in remote areas, remains a critical challenge due to limited power availability. In our previous work, we developed a CPU scheduling algorithm that optimized core allocation based on load conditions, reducing power utilization for gNodeB in a virtualized Open RAN environment. Extending our previous work, this paper introduces an advanced CPU scheduling for Open RAN designed to reduce power utilization in real hardware Open …

Smartphones On Wheels In Southeast Asia: A Crossroads For Data Governance, Attamongkol Tantratian, Gunn Jiravuttipong

Journal of Law and Mobility

While the transformation of automobiles into data-generating “smartphones on wheels” has revolutionized mobility, it has also raised critical concerns over data privacy and sovereignty. Equipped with sensors and connected technologies, smart vehicles collect vast amounts of data, including personal information, driving patterns, and biometric identifiers. While auto-exporting jurisdictions such as the United States, the European Union, and China have introduced regulatory measures to address these challenges, countries importing smart vehicles remain vulnerable due to their limited influence over the auto companies’ integrated technology and data policies.

This Article examines the regulatory and economic challenges faced by developing nations integrating foreign-designed …

A Comprehensive Review And Bibliometric Analysis On Collaborative Robotics For Industry: Safety Emerging As A Core Focus, Aida Haghighi, Morteza Cheraghi, Jérôme Pocachard, Valérie Botta-Genoulaz, Sabrina Jocelyn, Hamidreza Pourzarei

Revues de littérature, synthèses de connaissances

Research organizations and academics often seek to map the development of scientific fields, identify research gaps, and guide the direction of future research. In cobot-related research, the scientific literature consulted does not propose any comprehensive research agenda. Moreover, cobots, industrial robots inherently designed to collaborate with humans, bring with them emerging issues. To solve them, interdisciplinary research is often essential (e.g., combination of engineering, ergonomics and biomechanics expertise to handle safety challenges). This paper proposes an exhaustive study that employs a scoping review and bibliometric analysis to provide a structured macro perspective on the developments, key topics, and trends in …

Demonstrating The Impact Of Cpu Scheduling On Power Consumption In Virtualized Open Ran, Saish Urumkar, Sachin Sharma

Conference papers

Open RAN (Open Radio Access Network) is a next- generation wireless network gaining significant research interest globally due to its potential to provide a cost-efficient and scalable solution for growing network demands. Energy efficiency is an important area of focus in Open RAN deployments, as reducing power consumption while maintaining network performance is essential for sustainable wireless communication. This paper demonstrates the impact of CPU (Central Processing Unit) scheduling process priorities on power consumption and network performance in an Open RAN NodeB deployed on a testbed in the USA. The experimental results are demonstrated using two scenarios: (1) CPU Priority-Based …

Multi-Objective Deep Reinforcement Learning For Dynamic Algorithm Selection In Open Ran, Saish Urumkar, Byrav Ramamurthy, Sachin Sharma

Conference papers

Open Radio Access Networks (Open RAN) provide flexible, modular multi-vendor interoperability. Growing mobile data demand requires balancing network performance with power efficiency. Mobile operators need intelligent resource management to achieve Key Performance Indicator (KPI) targets while maintaining operational efficiency. This paper proposes a solution using a multi-objective deep reinforcement learning (MODRL) model deployed on the Open RAN Intelligent Controller (RIC). Three customizable operator profiles (Power Saving, Balanced, and Performance) are used which define specific priority ratios between performance and power saving objectives.

To evaluate, individual algorithms (CPU scheduling and UE connection state switching) are implemented in Open RAN, achieving 5–20%CPU …

Performance Evaluation Of Managed Switch Configurations For Secure And Efficient Plc-Based Industrial Automation Networks, Ahmed Salama

All Graduate Theses, Dissertations, and Other Capstone Projects

This thesis explores how managed switches can improve network performance and security in PLC-based industrial systems. Using simulations in GNS3 and Cisco Packet Tracer, and packet analysis via Wireshark, the study compares unmanaged and managed switch configurations. Redundancy protocols, particularly STP and RSTP, are evaluated under failure scenarios. Results show that RSTP offers faster recovery times, making it more suitable for time-sensitive environments. Additionally, managed switch features like VLANs, port security, and MAC filtering significantly reduce vulnerabilities and improve network segmentation. The findings highlight the importance of incorporating both cybersecurity and redundancy in industrial network design as systems move toward …

‘Waves Of Imagination’ Unconditional Spectogram Diffusion Using Diffusion Architecture., Rahul Vanukuri

Computer Science and Engineering Theses - Archive

The swift evolution of wireless communication technologies,particularly in the field of rf signals or in CBRS bands,demands increasingly sophisticated signal processing techniques to ensure efficient transmission, reception, and spectrum management.Traditional approaches to signal generation and reconstruction, although effective in controlled environments, often struggle to cope with the challenges presented by real-world noisy conditions, hardware constraints, and limited access to large-scale datasets. In response to these limitations, this thesis explores the application of diffusion models—a class of generative models known for their ability to produce high-fidelity samples—to the domain of spectrogram generation for communication signals.

Different from conventional strategies to simulate …

Introducing Catalizer: A Framework For Prototyping Models Of Biological Systems, Andy M. Day

Honors Theses

Modeling the light response system of Nannochloropsis oceanica brings

a set of challenges that make modeling difficult. Notably, potential models

may contain a large number of chemical species. A large number of

species creates a quadratic explosion in the number of potential pathways.

In addition, mathematically defining these pathways is error prone, yet

follows a surprising simple set of rules. We seek to create a domain specific

language which can precisely define these chemical reaction networks.

Once the networks have been defined, they can be exported as procedures

defined in popular programming languages for further analysis.

Vision-Language Models For Future Image Caption Prediction: Methods, Applications, And Adversarial Vulnerabilities, Md Ishak

Wayne State University Theses

Image captioning has traditionally focused on generating descriptions for individual static images. However, predicting future events from visual information is a fundamental challenge in this domain. While existing methods primarily describe current visual content, the ability to anticipate and generate captions for future events remains largely unexplored. We propose a novel approach for future caption prediction by leveraging the capabilities of Vision Transformer (ViT), Generative Pre-trained Transformer 2 (GPT-2), and Text-to-Text Transfer Transformer Model (T5) architectures.

Our method includes two complementary strategies: a two-stage pipeline where ViT-GPT2 generates captions for current images and T5 analyzes these captions to predict future …

Indigenous Technology Futurisms: Reclaiming Mackinac Island In Virtual Reality, Madeline A. Gupta

Library Map Prize

In Indigenous Technology Futurisms: Reclaiming Mackinac Island in Virtual Reality, Madeline A. Gupta introduces a place-based digital storytelling project that uses interactive mapping as the foundation for cultural reconnection and community wellness. Centered on Mackinac Island—ancestral land of the Anishinaabe people—the project features a custom-designed map by an Anishinaabe artist, allowing users to navigate a VR website by selecting specific locations across northern Michigan. Each map point opens into a spatial video experience paired with traditional audio content, including oral histories, poetry, and songs.

The project reframes digital cartography through an Indigenous lens, challenging Western colonial mapping practices that …

Artificial Intelligence For Digital Deception: A Study On Detection, Generation, And Evaluation, Tasnim Akter Onisha

College of Graduate Studies: Theses & Dissertations

The rapid advancement of artificial intelligence has significantly influenced digital media, enabling both the detection and generation of synthetic content. This thesis, titled Artificial Intelligence for digital deception: A Study on Detection, Generation, and Evaluation, explores AI’s role in digital deception through three distinct studies focused on facial expression analysis for deepfake detection, machine learning-based spam classification on cloud platforms, and the evaluation of generative AI state-of-the-art text to video models. The first study investigates the effectiveness of facial expression analysis in distinguishing between deepfake and genuine videos. Using Noldus FaceReader 7, participant’s emotional responses were analyzed while viewing deep-fake …

Mastering Enterprise Networks (2nd Ed), Mathew J. Heath Van Horn

OER Main

Mastering Enterprise Networks, is a comprehensive guide to building, defending, and attacking enterprise networks. It covers a wide range of topics, from network fundamentals to advanced security concepts. The book is well-organized and easy to follow, making it a valuable resource for both beginners and experienced network professionals.

One of the strengths of this book is its focus on hands-on learning. The book includes 50 chapters of labs that allow readers to practice the concepts they have learned. These labs are a great way to reinforce your understanding of enterprise networks while developing practical skills.

This book covers security from …

การศึกษาแบบแผนการเต้นของหัวใจระหว่างการวิ่ง โดยใช้การเรียนรู้ของเครื่องแบบไม่มีผู้สอนเพื่อการตรวจจับความผิดปกติ, ชวิน หังสสูต

Chulalongkorn University Theses and Dissertations (Chula ETD)

การศึกษานี้มีวัตถุประสงค์เพื่อวิเคราะห์รูปแบบการเต้นของหัวใจระหว่างการวิ่ง โดยใช้เทคนิคการตรวจจับความผิดปกติแบบไม่มีผู้สอน (Unsupervised Anomaly Detection - UAD) เพื่อการตรวจจับความผิดปกติ ปัญหาหลักในงานวิจัยนี้คือข้อจำกัดทางจริยธรรมในการรวบรวมข้อมูลภาวะวิกฤตเพื่อใช้สอนแบบจำลอง และความแปรปรวนของข้อมูลสรีรวิทยาที่สูงมากระหว่างบุคคลเพื่อแก้ไขปัญหานี้ งานวิจัยนี้ได้พัฒนาระบบเว็บแอปพลิเคชันสำหรับการรวบรวมและกำกับข้อมูล โดยผู้เชี่ยวชาญ และได้สร้างชุดข้อมูลอ้างอิงจากกลุ่มตัวอย่างนักวิ่ง 5 ท่าน จากนั้น ได้ดำเนินการศึกษาเชิงเปรียบเทียบแบบจำลอง UAD จำนวน 7 แบบจำลอง โดยใช้กลยุทธ์การเรียนรู้แบบจำเพาะบุคคล ภายใต้เงื่อนไข การกำหนดค่าไฮเปอร์พารามิเตอร์แบบคงที่ เพื่อจำลองสถานการณ์การใช้งานจริงแบบเริ่มต้น โดยปราศจากการปรับแต่งค่าล่วงหน้า และประเมินผลด้วยค่า F0.5-Score เพื่อเน้นความแม่นยำและลดการแจ้งเตือนที่ผิดพลาดผลการทดลองพบว่า เมื่อไม่มีการปรับจูนพารามิเตอร์ ประสิทธิภาพโดยรวมของทุกแบบจำลองลดลงอย่างมีนัยสำคัญ ซึ่งสะท้อนถึงความท้าทายในการสร้างแบบจำลองสากล อย่างไรก็ตาม แบบจำลอง Matrix Profile (MP) ในกลุ่มคลาสสิก พิสูจน์ให้เห็นถึงความทนทานสูงที่สุด โดยมีประสิทธิภาพเฉลี่ยสูงสุด (F0.5-Score 2.71%) และสามารถตรวจจับความผิดปกติได้ดีในรายบุคคล ในขณะที่แบบจำลองพื้นฐานล้มเหลวโดยสิ้นเชิง (0.00%) สำหรับกลุ่มการเรียนรู้เชิงลึก (USAD, LSTM-AE) พบว่ามีข้อจำกัดในการใช้งานแบบเริ่มต้น (Cold-start) โดยไม่สามารถตรวจจับความผิดปกติได้แม้จะมีปริมาณข้อมูลมาก หากปราศจากการปรับแต่งค่าพารามิเตอร์ ผลลัพธ์นี้ยืนยันว่าระบบเฝ้าระวังสุขภาพในอนาคตจำเป็นต้องมีกลไกการเรียนรู้แบบปรับตัว ร่วมด้วยเพื่อให้สามารถนำไปใช้งานจริงได้อย่างมีประสิทธิภาพ