Other Computer Engineering Commons^™

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 …

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 …

Photoproof: A Mobile Application For Verifying The Authenticity Of Images, Pruthviraj Urankar

Master's Projects

The easy access to artificial intelligence (AI) technologies, such as deepfakes and generative adversarial networks (GANs), has facilitated the creation of highly realistic artificial images, thereby undermining the authenticity of photos in today’s digital age. Misinformation and manipulation are key dangers to digital content due to this advancement. Therefore, the need for reliable methods of photo verification and authentication has become increasingly important. This report presents a decentralized iOS app that uses blockchain to ensure photo authenticity. The app leverages Ethereum smart contracts and cryptographic hashing to securely log image metadata. When a user takes a photo, the app hashes …

Optimizing Analytics Storage Strategies For Search Engines And Wiki Platforms, Sujith Kakarlapudi

Master's Projects

Modern search engines and wiki platforms generate vast quantities of user inter-
/="/">action data such as page views, edits, clicks, and session events. This data must be stored and aggregated efficiently to enable scalable analytics and responsive querying. Yioop, an open-source search engine framework, serves as our primary case study, processing millions of such events to power its indexing and recommendation features. This report explores a shift from the conventional database storage based model to a log based model, in order to improve scalability and write efficiency. A size-limited, append-only logging facility was provided to log analytics events: the …

Image-To-Text Transcription: Analyzing And Describing Visual Content, Zixiao Fan

Master's Projects

Image captioning, which provides a textual understanding of visual content, is the fundamental support for the advancement of Human-A.I. Interaction technology. In the hope of exploring the application of such technology, this project focuses on two specific goals. One is to directly explore the application of the image informationretrieving abilities, and the other is to dive into the specifics of the pipeline and components of image captioning models. As a result, this project presents a working app that exploits the text retrieval functionalities to enable image storage with functions like tagging and transcription. It also supports search functionality with a …

Galora: A Lightweight Graph-Aware Llm Framework For Node Classification On Text-Attributed Graphs, Mayur Choudhary

Master's Projects

With the exponential rise of language models (LMs) and their potential to understand semantic relationships, large LMs are being used across a wide range of applications. Text-attributed graphs (TAGs) are one notable example where LLMs can be combined with Graph Neural Networks (GNNs) to enhance node classification results. TAGs associate textual content with each node and are commonly seen in various domains such as social networks, citation graphs, recommendation systems, etc. Effectively modeling TAGs would enable deeper insights into different aspects of the graph and improve decision-making in relevant domains. We present GaLoRA, a parameter-efficient framework to integrate structural information …

Physiotrack: A Gamified Physiotherapy System, Pranavi Chaturvedula

Master's Projects

Traditional physiotherapy methods tend to be non-interactive and provide little to no personalized instruction, even though physiotherapy is critical to stroke recovery. This thesis explores a fully adaptive, sensor-based, feedback architecture intended for stroke patients which remotely supervises movement and personalizes exercises enabled by multimodal sensors. The system uses filtering and windowed segmentation of accelerometer and skeletal data to compute features like jerk, speed, and joint movement angular range. A game engine applies accelerometer and skeletal features together with optimized, lightweight ML models to drive adaptive feedback, scoring, and difficulty adjustment. The architecture supports responsive continuous sensor streaming within the …

Optimization Of Permutation Flowshop Scheduling Using An Island Genetic Algorithm For Makespan Minimization, Sahil Salim

Master's Projects

The Permutation Flowshop Scheduling Problem is a well-known NP-hard combinatorial optimization problem that involves the sequencing of n jobs across m machines in the same order to minimize the total makespan value. This project proposes a Heterogeneous Island Genetic Algorithm framework (HIGA). Each island represents a group of solutions that evolve in parallel using different initialization heuristics, crossover and mutation operators, and adaptive parameters. A dynamic, stagnation-based migration strategy is proposed to maintain targeted communication between the islands. The proposed HIGA approach was compared against the basic Standard Genetic Algorithm (SGA) and a more advanced Niche-based Genetic Algorithm (NEH-NGA) on …

Extraction Of A Knowledge Graph Of Biomedical Relationships, Brian Tran

Master's Projects

Rapid release in biomedical literature poses a challenge in linking information. This thesis aims to extract data from expanding datasets to identify and form meaningful relationships between biomedical entities. Large language models (LLMs) enable us to learn at a rapid pace. Creation of LLms from scratch are impractical. This thesis aims to collect a small dataset, containing biomedical papers, and use it to train large language models (LLMs) to extract entities from the text and learn the relationships between these entities. The experiment will be divided into two stages and utilize EU-ADR and ChemProt dataset. Starting with named entity recognition …

Domain-Specific Graph Rag Pipelines: Optimized Approaches For Building Efficient Personal Knowledge Repositories, Omkar Yadav

Master's Projects

Managing personal data, including notes, calendar events, to-do lists, and other personal information, has become increasingly complex and challenging. In response to this issue, I propose a framework using RAG that enables a large language model (LLM) to efficiently query this data without requiring training on the personal data itself. Conventional retrieval systems, including those leveraging vector-based retrievalaugmented generation (RAG), are effective at handling basic queries but struggle to deliver coherent global abstractions, integrate diverse knowledge sources, and account for temporal nuances. This research explores domain-specific Graph based RAG frameworks that incorporate a knowledge graph to better model relationships, thereby …

Enhancing Recommender Systems Using Graph Neural Networks, Long Short-Term Memory And Textual Embeddings, Tianxiang Chen

Master's Projects

Recommender systems surround us. They shape what we watch, how we buy, and even what our future might look like next. The Amazon Review Dataset and Movielens, those two datasets will help this project explore how to improve recommender systems through the user’s preferences. Two methods were combined: sequence-based models and graph-based models. Sequence models, such as LSTMs and Transformers, look at the order of user actions to find patterns by their sequence. On the other hand, Graphbased models focus on relationships between users, items, and their attributes. Textual embeddings added depth and context. Both methods offer something special, according …

Large Language Model Powered Etl Pipeline Development A Project Report, Aditya Chandras

Master's Projects

This projects discusses the use of Large language model to simplify and automate Extract, Transform, Loading (ETL) development. Data engineering tasks oUen need technical experDse, thereby challenging non-experts and even industry professionals for such Dme intensive operaDons. The project tackles these challenges by employing quanDzaDon of a base Lllama-2-7b-Chat model using QLora technique. The lowered precision of the base model allow further execuDon on a limited hardware resource. The project further performs Supervised fine-tuning trainer (SFT) to perform fine-tuning, specializing the model for generaDng script for ETL tasks. The fine-tuned model is evaluated through a series of micro and comprehensive …

Optimizing Integrated Access And Backhaul Topology Using Monte Carlo Tree Search, Srimanth Reddy Ummenthala

Master's Projects

Integrated Access and Backhaul (IAB) plays a central role in enabling scalability and high-throughput in wireless networks, especially in areas where wired backhaul is not feasible. Optimizing the topology of IAB networks is a complex task, involving trade-offs between path loss, node capacity, and link quality. To address this, this project investigates a Monte Carlo Tree Search (MCTS)-based approach to improve overall network performance by maximizing the capacity of network, considering Signal-to-Noise Ratio (SNR) on wireless links. MCTS provides a guided search mechanism to construct topologies efficiently under the constraints and is evaluated against baseline topologies. Experimental results show that …

Dynamic Pricing For Revenue Maximization In 5g Networks With Elastic Network Slicing And Reinforcement Learning, Naisheel Shah

Master's Projects

Network slicing is a concept that enables a diverse array of network applications with lots of unique service requirements. For the scope of this research, we delve into elastic network slicing, which has a potential benefit for both the providers and users through cost-effective resource utilization. Dynamic pricing of these slice resources is a method for operators to realize the balance of different types of network slices by implicitly communicating the current network state to slice users. We have designed a custom pricing scheme using Deep Reinforcement Learning for elastic network slices that maximizes the revenue of slice providers while …

Link Failure Localization In Hierarchical, Multidomain Optical Networks, Martin Mihailov Bojinov

Master's Projects

Optical networks transport data encoded on light signals over optical fiber cables. Individual optical networks are managed by domain administrators, such as service providers, vendors, or regional bodies. Multidomain optical networking explores the possibility of enabling seamless data transmission across domain boundaries. In this project, we explore how we can perform network monitoring in a hierarchical multidomain optical network while preserving domain security and autonomy. This is done by allocating dedicated monitoring trails across various broker abstractions of our network topologies. We propose three heuristic functions that expedite trail selection. Of the three heuristics, our least monitored algorithm performs the …

Performance Comparison Of Machine Learning Across Metal, Cuda, And Neuromorphic Frameworks, Ryan Saini

Master's Projects

Machine learning’s computational demands necessitate optimal performance and utilization. This research compares Apple Silicon M3 Pro with MPS, NVIDIA RTX 3070 GPU with CUDA, and neuromorphic computing for machine learning methods. We provide a cross-platform and cross-architecture performance analysis of machine learning methods to identify optimal configurations for training and inference scenarios. On traditional neural networks, Apple Silicon with MPS delivers superior energy efficiency at the cost of longer processing times for training and inference. NVIDIA with CUDA offers faster computation in training and inference at higher energy costs. Convolutional spiking neural networks perform competitively on event-based data, particularly on …

Suicidal Ideation Detection On Reddit Using Llm-Annotated Data And Graph Neural Networks, Ikbal Singh Gurdev Singh Dhanjal

Master's Projects

Suicide is the fourth leading cause of death among people aged 15-29. More than 720, 000 people commit suicide every year. During the COVID-19 pandemic, we saw an increase in people seeking out mental health support on anonymous forums like Reddit. These anonymous forums allow people to express their suicidal ideation without judgment and give them a support structure that not everyone has. The aim of this project is to detect suicidal ideation using Reddit. In this work, we propose SIRGEL (Suicidal Ideation on Reddit using Graph Embeddings and LLMs), a dual-pipeline approach that combines large language model (LLM)- based …

Malware Opcode Embedding And Quality Assessment Of Generative Sample Embeddings, Atishay Jain

Master's Projects

Malware is software used to damage and disrupt computer systems with the intent to cause damage to the victim. Malware detection and classification into malware families is a crucial problem for cybersecurity researchers. One of the major bottlenecks in improving these systems is the shortage of good quality labeled malware data, especially for malware families with scarce samples. Researchers have utilized generative models to generate malware data to address this issue. Malware embeddings encode patterns within a malware file, which can be used to detect and classify malware. Recently, encouraging results have been obtained in generating malware embeddings using generative …

Transforming Chatbot Text: A Sequence-To-Sequence Approach To Human-Like Text Generation, Natesh Reddy

Master's Projects

With the advancements in Large Language Models (LLMs) such as ChatGPT, the boundary between text written by a human and that created by AI has become blurred. This poses a threat to systems designed to detect AI-generated content. In this work, we adopt a novel strategy to adversarially transform GPT-generated text using sequence-to-sequence (Seq2Seq) models, to make the text more human-like. We concentrate on improving GPT-generated sentences by including significant linguistic, structural, and semantic components that are typical of human-authored text. The goal is to then use this transformed data to train a more robust detection model. Our experiments show …

Human Vs. Llm: Interpreting Emotion From Neutral Faces In Contextually Charged Backgrounds, Sai Mounika Peteti

Master's Projects

In this study, we investigate the intersection of cultural context, visual perception, and implicit bias in interpreting neutral human expressions. Specifically, we explore how emotionally charged backgrounds can shape viewers interpretations of neutral facial expressions. Using an experimental setup, participants are shown neutral human portraits paired with varying background types ranging from emotionally neutral or pleasant to evocative scenes. Participants are tasked with selecting which background best matches the emotional state of the person depicted in the portraits. Eye-tracking data is collected to analyze visual attention patterns and cognitive processing. In parallel, we also evaluate how large language models (LLMs) …

Clustering Organ Cell Types, Venkata Satya Swathi Mattaparthi

Master's Projects

The Human Cell Atlas (HCA) created a reference map of all human cells. My project uses the Tabula Sapiens dataset, developed under HCA and based on single-cell RNA sequencing data, to explore cell type and tissue diversity. I performed experiments using the Elbow method and a formula based on dataset observations to determine the number of clusters, then applied k-means clustering on two representative subsets of the All Cells dataset. Clusters were selected for analysis using Shannon’s Diversity Index and Pielou’s Evenness. A novel algorithm based on the cell differentiation tree was used to validate the biological coherence of the …

Comparative Analysis Of Adversarial Permeability In Cpu-Native, Qat And Onnx-Based Quantized Transformer Models, Fahad Siddiqui

Master's Projects

This work explores securing and optimization of Transformer-based time series forecasting models. We employ several quantization techniques, including quantization-aware training (QAT), and tested the robustness of quantized models by adversarially attacking them. The preliminary results of this work, in our controlled setup, indicate that quantized models outperform the full precision model in terms of robustness against adversarial attacks. They achieved this robustness while showing a very minimal decrease in their forecasting performance.

Framework For Identity Privacy Through Gender Based Skeletonization, Harrison Hwang

Master's Projects

The protection of one’s privacy and sensitive information is becoming increasingly difficult in the modern age full of surveillance and data collection. Through the use of image based object detection machine learning models trained for human and facial recognition, people can be identified and tracked to a terrifyingly accurate degree. On the other hand, the information present in surveillance media can play a key role in security and law enforcement. This presents a problem of how to preserve key information without compromising the privacy of any individuals present in the video. In this research project, Computer Vision techniques and a …

Enhancing Code Review Automation With Large Language Models Using Qlora Fine-Tuning And Rags, Sumukh Naveen Aradhya

Master's Projects

In this technological era where Artificial Intelligence and Machine Learning are revolutionizing various domains, Large Language Models (LLMs) are emerging as a very powerful tool. In the software development lifecycle, it is extremely important to have reliable code reviews to ensure security and maintain code quality. This project aims to survey various existing methodologies to aid creation of efficient code review automation agents and also research on ways to make this process more efficient. Parameter Efficient Fine-Tuning (PEFT) methodologies such as LoRA and QLoRA have been explored with an additional focus on a hybrid model that combines adaptive QLoRA with …

Large Language Models For Bacterial Genomic Analysis, Manvendra Chavan

Master's Projects

Identification of bacterial gene sequences with agricultural applications has the potential to transform agricultural biotechnology. These genes can be used in environmentally friendly pest control strategies. One such use case is identifying genes with potential insecticidal properties. With an increasing number of genomic information and decreasing numbers of available annotated sequences, finding new insecticidal genes has become more challenging.The traditional methods relying on sequence alignment and annotated databases are not effective in detecting functionally relevant genes lacking close homology to known cases. This project investigates the data-driven classification of genes by sequence modeling. This research is focused on learning DNA …

Genegate: Genetic Gating In A Mixture-Of-Experts For Real-Time Multi-Objective Traffic Signal Control, Rashmi Vishwanath Bhat

Master's Projects

Urban traffic signal control often needs to juggle between competing goals. It needs to minimize delays, reduce emissions, prevent crashes, and prioritize emergency vehicles all while the demand is constantly fluctuating. Traditional fixed-time or statically blended policies cannot reallocate priorities quickly when conditions change. We introduce GeneGate, a mixture-of-experts framework that uses a lightweight genetic gate to fuse four specialist controllers (throughput, emissions, safety, emergency) and adjusts their weights in real time. A short offline genetic search produces a robust initial blend, and an online micro-evolution step refines it every few cycles based on live traffic feedback. GeneGate’s adaptive gating …

Advanced Knowledge Extraction With Biomedical Data Using Llms, Akshat Krishna

Master's Projects

The rapid growth of biomedical research has led to an overwhelming volume of unstructured textual data in the scientific literature. This has necessitated the development of an automated approach for knowledge extraction and integration. In

this project, we present a comprehensive pipeline for constructing a unified biomed- ical knowledge graph by combining two well-known datasets: CHEMPROT [1],

which captures chemical–protein interactions, and EU-ADR [2], which annotates drug–gene–disease relationships. In order to identify important biomedical entities and interactions from CHEMPROT dataset, we perform Named Entity Recognition (NER) and relation Extraction (RE) using state-of-the-art biomedical models like BioBERT [3], BioGPT [4] and …