I2S Masters/ Doctoral Theses

Accurate biological sequence annotation and privacy-aware clinical modeling are central challenges in modern computational biology and biomedical AI. This dissertation presents scalable and interpretable deep learning frameworks spanning protein family classification, metal-ion binding prediction, and privacy-preserving electrocardiogram (ECG) representation learning. First, we introduce GPCR-SLM, a lightweight transformer-based framework for high-resolution classification of G-protein coupled receptors (GPCRs), one of the largest and most pharmacologically important protein families, targeted by approximately 35% of FDA-approved drugs. Unlike traditional homology-based tools such as BLAST and HMMER, which struggle to distinguish closely related families with low sequence similarity, our knowledge-distilled small language model achieves 99% accuracy across 86 GPCR families. The framework significantly outperforms BLAST (86.4%) and HMMER (91%) while delivering a 33.5× computational speedup compared to large protein language models, enabling scalable functional annotation as protein databases continue to expand. 

Second, we present an end-to-end deep learning pipeline for protein–metal-ion binding prediction. Binding site annotation is traditionally labor-intensive and limited by handcrafted features or predefined residue sets. We systematically evaluate five state-of-the-art protein language models and incorporate positional encoding to capture long-range residue dependencies. Our approach achieves a Matthews Correlation Coefficient (MCC) of 0.89 with precision, recall, and F1 scores exceeding 95% for six major metal ions under 10-fold cross-validation, demonstrating robust predictive performance and improved biological interpretability. Finally, we address fairness and privacy in clinical AI through a variational autoencoder (VAE) framework for ECG representation learning. Because ECGs inherently encode sensitive soft biometrics such as sex, age, and race, we design a dual-discriminator architecture that suppresses demographic information while preserving clinically relevant signals. The reconstructed ECGs substantially reduce demographic identifiability while maintaining strong predictive performance for reduced left ventricular ejection fraction, left ventricular hypertrophy, and 5-year mortality. 

Collectively, this work advances parameter-efficient, scalable, and privacy-conscious deep learning methodologies for both molecular and clinical domains, bridging computational protein science and trustworthy biomedical AI. 

Motor difficulties show up in as many as 90% of people with autism, but surprisingly few, somewhere between 13% and 32%, ever get motor-focused help. A big part of the problem is that the tools we have for measuring motor skills either rely on a clinician's subjective judgment or require expensive lab equipment that most families will never have access to. This dissertation tries to close that gap with three projects, all built around the idea that a regular webcam and some well-designed deep learning models can do much of what costly motion-capture labs do today.

The first project asks a straightforward question: can a computer tell the difference between how someone with autism moves and how a typically developing person moves, just by watching a short video? The answer, it turns out, is yes. We built an ensemble of three neural networks, each one tuned to notice something different. One focuses on how joints coordinate with each other spatially, other zeroes in on the timing of movements, and the third learns which body-part relationships matter most for a given clip. We tested the system on 582 videos from 118 people (69 with ASD and 49 without) performing simple everyday actions like stirring or hammering. The ensemble correctly classifies 95.65% of cases. The timing-focused model on its own hits 92%, which is nearly 10 points better than a standard recurrent network baseline. And when all three models agree, accuracy climbs above 98%.

The second project deals with stimming, the repetitive behaviors like arm flapping, head banging, and spinning that are common in autism. Working with 302 publicly available videos, we trained a skeleton-based model that reaches 91% accuracy using body pose alone. That is more than double the 47% that previous work managed on the same benchmark. When we combine the pose information with what the raw video shows through a late fusion approach, accuracy jumps to 99.9%. Across the entire test set, only a single video was misclassified.

The third project is E-MotionSpec, a web platform designed for clinicians and researchers who want to track motor development over time. It runs in any browser, uses MediaPipe to estimate body pose in real time, and extracts 44 movement features grouped into seven domains covering things like how smoothly someone moves, how quickly they initiate actions, and how coordinated their limbs are. We validated the platform on the same 118-participant dataset and found 36 features with statistically significant differences between the ASD and typically developing groups. Smoothness and initiation timing stood out as the strongest discriminators. The platform also includes tools for comparing sessions over time using frequency analysis and dynamic time warping, so a clinician can actually see whether someone's motor patterns are changing across weeks or months.

Taken together, these three projects offer a practical path toward earlier identification and better ongoing monitoring of motor difficulties in autism. Everything runs on a webcam and a web browser. No motion-capture suits, no force plates, no specialized labs. That matters most for the families, schools, and clinics that need these tools the most and can least afford the alternatives.

Guava leaf diseases significantly affect crop yield and quality, making timely detection essential for effective disease management. This project presents an end-to-end software system for automated guava leaf disease detection using deep learning and transfer learning techniques. Multiple pretrained convolutional neural network (CNN) architectures, including ResNet, AlexNet, VGG, SqueezeNet, DenseNet, Inception-v3, and EfficientNet, were adapted through feature extraction and trained on a guava leaf image dataset.

The system allows users to either capture an image using a camera or upload an existing leaf image through a software interface. The input image is processed and classified by the trained deep learning model, and the predicted disease class is displayed to the user. The dataset was divided into training, validation, and test sets to ensure robust performance evaluation, and final test accuracy was used to measure generalization on unseen data.

Experimental results demonstrate that transfer learning enables accurate and efficient guava leaf disease classification. Among the evaluated models, the best-performing architecture achieved an accuracy between 97% to 99%. Overall, the developed software provides a practical and user-friendly solution for real-world agricultural disease diagnosis.

The rapid growth of Internet of Things (IoT) technology has brought unprecedented convenience to everyday life, enabling users to deploy automation rules and develop IoT apps tailored to their specific needs. However, modern IoT ecosystems consist of numerous devices, applications, and platforms that interact continuously. As a result, users are increasingly exposed to complex and subtle security and privacy risks that are difficult to fully comprehend. Even interactions among seemingly harmless apps can introduce unforeseen security and privacy threats. In addition, violations of memory integrity can undermine the security guarantees on which IoT apps rely.

The first approach investigates hidden cross-app privacy leakage risks in IoT apps. These risks arise from cross-app interaction chains formed among multiple seemingly benign IoT apps. Our analysis reveals that interactions between apps can expose sensitive information such as user identity, location, tracking data, and activity patterns. We quantify these privacy leaks by assigning probability scores to evaluate risk levels based on inferences. In addition, we provide a fine-grained categorization of privacy threats to generate detailed alerts, enabling users to better understand and address specific privacy risks.

The second approach addresses cross-app interaction threats in IoT automation systems by leveraging a logic-based analysis model grounded in event relations. We formalize event relationships, detect event interferences, and classify rule conflicts, then generate risk scores and conflict rankings to enable comprehensive conflict detection and risk assessment. To mitigate the identified interaction threats, an optimization-based approach is employed to reduce risks while preserving system functionality. This approach ensures comprehensive coverage of cross-app interaction threats and provides a robust solution for detecting and resolving rule conflicts in IoT environments.

To support the development and rigorous evaluation of these security analyses, we further developed a large-scale, manually verified, and comprehensive dataset of real-world IoT apps. This clean and diverse benchmark dataset supports the development and validation of IoT security and privacy solutions. All proposed approaches are evaluated using this dataset of real-world apps, collectively offering valuable insights and practical tools for enhancing IoT security and privacy against cross-app threats. Furthermore, we examine the integrity of the execution environment that supports IoT apps. We show that, even under non-privileged execution, carefully crafted memory access patterns can induce bit flips in physical memory, allowing attackers to corrupt data and compromise system integrity without requiring elevated privileges.

In resource‑constrained contested environments (RCCEs), communications are routinely censored, surveilled, or disrupted by nation‑state adversaries, leaving at‑risk populations—including protesters, dissidents, disaster‑affected communities, and military units—without secure connectivity. This dissertation introduces MeshBLanket, a Bluetooth Mesh‑based framework designed for low‑power, low‑throughput messaging with minimal electromagnetic spectrum exposure. Built on commercial off‑the‑shelf hardware, MeshBLanket extends the Bluetooth Mesh specification with automated provisioning and network‑wide key refresh to enhance scalability and resilience.

We evaluated MeshBLanket through field experimentation (range, throughput, battery life, and security enhancements) and qualitative interviews with ten senior U.S. Army communications experts. Thematic analysis revealed priorities of availability, EMS footprint reduction, and simplicity of use, alongside adoption challenges and institutional skepticism. Results demonstrate that MeshBLanket maintains secure messaging under load, supports autonomous key refresh, and offers operational relevance at the forward edge of battlefields.

Beyond military contexts, parallels with protest environments highlight MeshBLanket’s broader applicability for civilian populations facing censorship and surveillance. By unifying technical experimentation with expert perspectives, this work contributes a proof‑of‑concept communications architecture that advances secure, resilient, and user‑centric connectivity in environments where traditional infrastructure is compromised or weaponized.

Accurate real-estate price estimation is crucial for buyers, sellers, investors, lenders, and policymakers, yet traditional valuation practices often rely on subjective judgment, inconsistent expertise, and incomplete market information. With the increasing availability of digital property listings, large volumes of structured real-estate data can now be leveraged to build objective, data-driven valuation systems. This project develops a comprehensive analytical framework for predicting different types of properties prices using real-world listing data collected from 99acres.com across major Indian cities. The workflow includes automated web scraping, extensive data cleaning, normalization of heterogeneous property attributes, and exploratory data analysis to identify important pricing patterns and structural trends within the dataset. A multi-stage learning pipeline is designed—consisting of feature preparation, hyperparameter tuning, cross-validation, and performance evaluation—to ensure that the final predictive system is both reliable and generalizable. In addition to the core prediction engine, the project proposes a future extension using Retrieval-Augmented Generation (RAG) with Large Language Models(LLM’s) to provide transparent, context-aware explanations for each valuation. Overall, this work establishes the foundation for a scalable, interpretable, and data-centric real-estate valuation platform capable of supporting informed decision-making in diverse market contexts.

The rapid expansion of academic literature has made efficient information extraction increasingly difficult for researchers, leading to substantial time spent manually summarizing documents and identifying key insights. This project presents an AI-powered Academic Assistant designed to streamline academic reading through multi-level summarization, contextual question answering, and source-grounded traceability. The system incorporates a robust preprocessing pipeline including text extraction, artifact removal, noise filtering, and section segmentation to prepare documents for accurate analysis. After assessing the limitations of traditional NLP and transformer-based summarization models, the project adopts a Large Language Model (LLM) approach using the Gemini API, enabling deeper semantic understanding, long-context processing, and flexible summarization. The assistant provides structured short, medium, and long summaries; contextual keyword extraction; and interactive question answering with transparent source highlighting. Limitations include handling complex visual content and occasional API constraints. Overall, this project demonstrates how modern LLMs, combined with tailored prompt engineering and structured preprocessing, can significantly enhance the academic document analysis workflow.

This project presents Intellinotes, an AI-powered platform that transforms educational documents into multiple learning formats to address information-overload challenges in modern education. The system leverages large language models (GPT-4o-mini) to automatically generate four complementary outputs from a single document upload: educational summaries, conversational podcast scripts, hierarchical mind maps, and interactive flashcards.

The platform employs a three-tier architecture built with Next.js, FastAPI, and MongoDB, supporting multiple document formats (PDF, DOCX, PPTX, TXT, images) through a robust parsing pipeline. Comprehensive evaluation on 30 research documents demonstrates exceptional system reliability with a 100% feature success rate across 150 tests (5 features × 30 documents), and strong semantic understanding with a semantic similarity score of 0.72.

While ROUGE scores (ROUGE-1: 0.40, ROUGE-2: 0.09, ROUGE-L: 0.17) indicate moderate lexical overlap typical of abstractive summarization, the high semantic similarity demonstrates that the system effectively captures and conveys the conceptual meaning of source documents—an essential requirement for educational content. This validation of meaning preservation over word matching represents an important contribution to evaluating educational AI systems.

The system processes documents in approximately 65 seconds with perfect reliability, providing students with comprehensive multi-modal learning materials that cater to diverse learning styles. This work contributes to the growing field of AI-assisted education by demonstrating a practical application of large language models for automated educational content generation supported by validated quality metrics.

Designing a well-balanced exam requires instructors to review extensive course materials, determine key concepts, and design questions that reflect appropriate difficulty and cognitive depth. This project develops an AI-powered Question Paper Generator that automates much of this process while keeping instructors in full control. The system accepts PDFs, Word documents, PPT slides, and text files, extracts their content, and builds a FAISS-based retrieval index using sentence-transformer embeddings. A large language model then generates multiple question types—MCQs, short answers, and true/false—guided by user-selected difficulty levels and Bloom’s Taxonomy distributions to ensure meaningful coverage. Each question is evaluated with a grounding score that measures how closely it aligns with the source material, improving transparency and reducing hallucination. A React frontend enables instructors to monitor progress, review questions, toggle answers, and export to PDF or Word, while an ASP.NET Core backend manages processing and metrics. The system reduces exam preparation time and enhances consistency across assessments.

The rapid spread of misinformation across online platforms has made automated fake news detection essential. This project develops and compares machine learning (SVM, Decision Tree) and deep learning (LSTM) models to classify news headlines from the GossipCop and PolitiFact datasets as real or fake. After extensive preprocessing— including text cleaning, lemmatization, TF-IDF vectorization, and sequence tokenization—the models are trained and evaluated using standard performance metrics. Results show that SVM provides a strong baseline, but the LSTM model achieves higher accuracy and F1-scores by capturing deeper semantic and contextual patterns in the text. The study highlights the challenges of domain variation and subtle linguistic cues, while demonstrating that context-aware deep learning methods offer superior capability for automated fake content detection.