I2S Masters/ Doctoral Theses

Latency–accuracy tradeoffs are fundamental in real-time applications of deep neural networks (DNNs) for cyber-physical systems. In autonomous driving, in particular, safety depends on both prediction quality and the end-to-end delay from sensing to actuation. We observe that (1) when latency is accounted for, the latency-optimal network configuration varies with scene context and compute availability; and (2) a single fixed-resolution model becomes suboptimal as conditions change.

We present a multi-resolution, end-to-end deep neural network for the CARLA urban driving challenge using monocular camera input. Our approach employs a convolutional neural network (CNN) that supports multiple input resolutions through per-resolution batch normalization, enabling runtime selection of an ideal input scale under a latency budget, as well as resolution retargeting, which allows multi-resolution training without access to the original training dataset.

We implement and evaluate our multi-resolution end-to-end CNN in CARLA to explore the latency–safety frontier. Results show consistent improvements in per-route safety metrics—lane invasions, red-light infractions, and collisions—relative to fixed-resolution baselines.

Reliable cellular connectivity is essential for modern services such as telehealth, precision agriculture, and remote education; yet, measuring network performance in rural areas presents significant challenges. Traditional drive testing cannot access large geographic areas between roads, while crowdsourced data provides insufficient spatial resolution in low-population regions. To address these limitations, we develop an open-source UAV-based measurement platform that integrates an onboard computation unit, commercial cellular modem, and automated flight control to systematically capture Radio Access Network (RAN) signals and end-to-end network performance metrics at different altitudes. Our platform collects synchronized measurements of signal strength (RSRP, RSSI), signal quality (RSRQ, SINR), latency, and bidirectional throughput, with each measurement tagged with GPS coordinates and altitude. Experimental results from a semi-rural deployment reveal a fundamental altitude-dependent trade-off: received signal power improves at higher altitudes due to enhanced line-of-sight conditions, while signal quality degrades from increased interference with neighboring cells. Our analysis indicates that most of the measurement area maintains acceptable signal quality, along with adequate throughput performance, for both uplink and downlink communications. We further demonstrate that strong radio signal metrics for individual cells do not necessarily translate to spatial coverage dominance such that the cell serving the majority of our test area exhibited only moderate performance, while cells with superior metrics contributed minimally to overall coverage. Next, we develop several machine learning (ML) models to improve the prediction accuracy of signal strength at unmeasured altitudes. Finally, we extend our measurement platform by integrating non-terrestrial network (NTN) user terminals with the UAV components to investigate the performance of Low-earth Orbit (LEO) satellite networks with UAV mobility. Our measurement results demonstrate that NTN offers a viable fallback option by achieving acceptable latency and throughput performance during flight operations. Overall, this work establishes a reproducible methodology for three-dimensional rural network characterization and provides practical insights for network operators, regulators, and researchers addressing connectivity challenges in underserved areas.

Python's dynamic nature makes performance enhancement challenging. Recently, a JIT compiler using a novel copy-and-patch compilation approach was implemented in the reference Python implementation, CPython. Our goal in this work is to study and understand the performance properties of CPython's new JIT compiler. To facilitate this study, we compare the quality and performance of the code generated by this new JIT compiler with a more mature and traditional meta-tracing based JIT compiler implemented in PyPy (another Python implementation). Our thorough experimental evaluation reveals that, while it achieves the goal of fast compilation speed, CPython's JIT severely lags in code quality/performance compared with PyPy. While this observation is a known and intentional property of the copy-and-patch approach, it results in the new JIT compiler failing to elevate Python code performance beyond that achieved by the default interpreter, despite significant added code complexity. In this thesis, we report and explain our novel experiments, results, and observations.

Recent research in Electronic Health Records (EHRs) has enabled personalized and longitudinal modeling of patient trajectories for health outcome improvement. Despite this progress, existing methods often struggle to capture the dynamic, heterogeneous, and interdependent nature of medical data. Specifically, many representation methods learn a rich set of EHR features in an independent way but overlook the intricate relationships among them. Moreover, data scarcity and bias, such as the cold-start scenarios where patients only have a few visits or rare conditions, remain fundamental challenges in clinical decision support in real-life. To address these challenges, this dissertation aims to introduce an integrated machine learning framework for sophisticated, interpretable, and adaptive EHR representation modeling. Specifically, the dissertation comprises three thrusts:

1. A time-aware graph transformer model that dynamically constructs personalized temporal graph representations that capture patient trajectory over different visits.
2. A contrasted multi-Intent recommender system that can disentangle the multiple temporal patterns that coexist in a patient’s long medical history, while considering distinct health profiles.
3. A few-shot meta-learning framework that can address the patient cold-start issue through a self- and peer-adaptive model enhanced by uncertainty-based filtering.

Together, these contributions advance a data-efficient, generalizable, and interpretable foundation for large-scale clinical EHR mining toward truly personalized medical outcome prediction.

In the light of rapid advancement of global connectivity and the increasing reliance on multilingual communication, speech-based Machine Translation (MT) systems have emerged as essential technologies for facilitating seamless cross-lingual interaction. These systems enable individuals and organizations to overcome linguistic boundaries by automatically translating spoken language in real time. However, despite their growing ubiquity in various applications such as virtual assistants, international conferencing, and accessibility services, the security and robustness of speech-based MT systems remain underexplored. In particular, limited attention has been given to understanding their vulnerabilities under adversarial conditions, where malicious actors intentionally craft or manipulate speech inputs to mislead or degrade translation performance.

This thesis presents a comprehensive investigation into the security landscape of speech-based machine translation systems from an adversarial perspective. We systematically categorize and analyze potential attack vectors, evaluate their success rates across diverse system architectures and environmental settings, and explore the practical implications of such attacks. Furthermore, through a series of controlled experiments and human-subject evaluations, we demonstrate that adversarial manipulations can significantly distort translation outputs in realistic use cases, thereby posing tangible risks to communication reliability and user trust.

Our findings reveal critical weaknesses in current MT models and underscore the urgent need for developing more resilient defense strategies. We also discuss open research challenges and propose directions for building secure, trustworthy, and ethically responsible speech translation technologies. Ultimately, this work contributes to a deeper understanding of adversarial robustness in multimodal language systems and provides a foundation for advancing the security of next-generation machine translation frameworks.

As high-performance computing (HPC) systems achieve exascale capabilities, traditional single-objective schedulers that optimize solely for performance prove inadequate for environments requiring simultaneous optimization of energy efficiency and system resilience. Current scheduling approaches result in suboptimal resource utilization, excessive energy consumption, and reduced fault tolerance in the demanding requirements of large-scale scientific applications. This dissertation proposes a novel multi-objective optimization framework that integrates graph neural networks (GNNs) with reinforcement learning (RL) to jointly optimize performance, energy efficiency, and system resilience in HPC workload scheduling. The central hypothesis posits that graph-structured representations of workloads and system states, combined with adaptive learning policies, can significantly outperform traditional scheduling methods in complex, dynamic HPC environments. The proposed framework comprises three integrated components: (1) GNN-RL, which combines graph neural networks with reinforcement learning for adaptive policy development; (2) EA-GATSched, an energy-aware scheduler leveraging Graph Attention Networks; and (3) HARMONIC (Holistic Adaptive Resource Management for Optimized Next-generation Interconnected Computing), a probabilistic model for workload uncertainty quantification. The proposed methodology encompasses novel uncertainty modeling techniques, scalable GNN-based scheduling algorithms, and comprehensive empirical evaluation using production supercomputing workload traces. Preliminary results demonstrate 10-19% improvements in energy efficiency while maintaining comparable performance metrics. The framework will be evaluated across makespan reduction, energy consumption, resource utilization efficiency, and fault tolerance in various operational scenarios. This research advances sustainable and resilient HPC resource management, providing critical infrastructure support for next-generation scientific computing applications.

Remote attestation is a process of obtaining verifiable evidence from a remote party to establish trust. A relying party makes a request of a remote target that responds by executing an attestation protocol producing evidence reflecting the target's system state and meta-evidence reflecting the evidence’s integrity and provenance. This process occurs in the presence of adversaries intent on misleading the relying party to trust a target they should not. This research introduces a robust approach for evaluating and comparing attestation protocols based on their relative resilience against such adversaries. I develop a Rocq-based, formally-verified mathematical model aimed at describing the difficulty for an active adversary to successfully compromise the attestation. The model supports systematically ranking attestation protocols by the level of adversary effort required to produce evidence that does not accurately reflect the target’s state. My work aims to facilitate the selection of a protocol resilient to adversarial attack.

The used car market is characterized by significant pricing variability, making it challenging for buyers and sellers to determine fair vehicle values. To address this, the project applies a machine learning–driven approach to predict used car prices based on real market data extracted from Cars.com. Following extensive data cleaning, feature engineering, and exploratory analysis, several predictive models were developed and evaluated. Among these, the Stacking Regressor demonstrated superior performance, effectively capturing non-linear pricing patterns and achieving the highest accuracy with the lowest prediction error. Key insights indicate that vehicle age and mileage are the primary drivers of price depreciation, while brand and vehicle category exert notable secondary influence. The resulting pricing model provides a data-backed, transparent framework that supports more informed decision-making and promotes fairness and consistency within the used car marketplace.

Querying information from relational databases often requires proficiency in SQL, creating a steep learning curve for users who lack programming or database management experience. Text-to-SQL systems aim to bridge this gap by automatically converting natural language questions into executable SQL statements. In recent years, multi-agent frameworks have gained traction for this task, as they enable complex query generation to be decomposed into specialized subtasks such as schema selection based on user intent, SQL synthesis, and refinement of SQL queries through execution-based error correction. This work explores the integration of a human feedback component within a multi-agent Text-to-SQL framework. Human input is introduced after the selector agent identifies relevant schemas and tables, offering targeted guidance before SQL generation. The objective is to examine how such feedback can improve the system’s accuracy and contextual understanding of queries. The implementation leverages OpenAI’s GPT-4.1 mini and GPT-4.1 nano models as the underlying language components. The evaluation is carried out using a standard Text-to-SQL benchmark dataset, focusing on key performance metrics such as execution accuracy and validity efficiency scores.

Software vulnerabilities are widespread, often resulting from coding weaknesses and poor development practices. These vulnerabilities can be exploited by attackers, posing risks to confidentiality, integrity, and availability. To protect themselves, end-users of software may have an interest in knowing whether the software they purchase, and use is secure from potential attacks. Our work is motivated by this need to automatically assess and rate the security properties of binary software.

While many researchers focus on developing techniques and tools to detect and mitigate vulnerabilities in binaries, our approach is different. We aim to determine whether the software has been developed with proper care. Our hypothesis is that software created with meticulous attention to security is less likely to contain exploitable vulnerabilities. As a first step, we examined the current landscape of binary-level vulnerability detection. We categorized critical coding weaknesses in compiled programming languages and conducted a detailed survey comparing static analysis techniques and tools designed to detect these weaknesses. Additionally, we evaluated the effectiveness of open-source CWE detection tools and analyzed their challenges. To further understand their efficacy, we conducted independent assessments using standard benchmarks.

To determine whether software is carefully and securely developed, we propose several techniques. So far, we have used machine learning and deep learning methods to identify the programming language of a binary at the functional level, enabling us to handle complex cases like mixed-language binaries and we assess whether vulnerable regions in the binary are protected with appropriate security mechanisms. Additionally, we explored the feasibility of detecting secure coding practices by examining adherence to SonarQube’s security-related coding conventions.

Next, we investigate whether compiler warnings generated during binary creation are properly addressed. Furthermore, we also aim to optimize the array bounds detection in the program binary. This enhanced array bounds detection will also increase the effectiveness of detecting secure coding conventions that are related to memory safety and buffer overflow vulnerabilities.

Our ultimate goal is to combine these techniques to rate the overall security quality of a given binary software.