Thesis/Project Final Defense Schedule

Join us as the School of STEM master’s degree candidates present their culminating thesis and project work. The schedule is updated throughout the quarter, check back for new defenses.

View previous quarter schedules

Select a master’s program to navigate to candidates:

Master of Science in Computer Science & Software Engineering

SUMMER 2026

Monday, July 20

Ankita Maria John

Chair: Dr. Geethapriya Thamilarasu
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Ankita Maria John’s online defense
Thesis: SCALES: Dual Information-Theoretic Approaches to Mitigating and Quantifying Prompt Injections in RAG Pipelines

Large language models in Retrieval-Augmented Generation (RAG) pipelines remain highly vulnerable to prompt injection attacks. Current defenses often fail against adaptive and multimodal attacks or rendered unusable due to high false-positive rates. This thesis resolves this usability-security trade-off through two parallel information-theoretic contributions. First, we present a modality-aware, three-layer defense cascade combining a DeBERTa pre-filter, KL-divergence semantic boundary chunking, and an Information-Theoretic Mixed-Objective Contrastive (IT-MOC) LoRA adapter. Across a 500-sample benchmark spanning five attack families, this cascade achieves a <10% Attack Success Rate while maintaining over a 90% Benign Pass Rate. Second, we introduce a graded leakage quantification framework that replaces binary success metrics with a Combined Leakage Index measuring lexical, semantic, algorithmic, and distributional leakage. This protocol mathematically proves a substantial reduction in mean leakage severity over undefended baselines, establishing information-theoretic principles as foundational for both active RAG defense and security benchmarking.

Tuesday, July 21

Gayatri Malladi

Chair: Dr. Geethapriya Thamilarasu
Candidate: Master of Science in Computer Science & Software Engineering
1:00 P.M.; Join Gayatri Malladi’s online defense
Thesis: GuardRAG: Adversarial Preference Training Against Indirect Prompt Injections in RAG Systems

Retrieval-Augmented Generation (RAG) systems have enhanced the performance of Large Language Models (LLMs) by effectively addressing challenges such as hallucinations and generating irrelevant responses. However, augmenting LLMs with retrieval introduces privacy and data-leakage risks through the retrieval pipeline. Indirect Prompt Injections (IPIs) occur when hidden instructions inside retrieved documents enter a model’s context as trusted evidence, exploiting weak guardrails and leading to unauthorized data leaks, policy overrides, or attacker-controlled behavior. Existing defenses rely on brittle delimiter heuristics or limited retriever adjustments, leaving RAG systems vulnerable to adversarial directives blended seamlessly into retrieved content.

This thesis studies how IPIs propagate across the lifecycle of untrusted retrieved content, from retrieval-time exposure to model behavior and persistent agent memory. We first introduce RIPE-II, a corpus-level benchmark that evaluates IPIs under a realistic content-poisoning threat model. RIPE-II contains 32k attacks across six corpora, four domains, and twelve carrier families, and measures retrieval exposure and generation compromise as separate stages. Results show that poisoned passages reach the model on most queries, reranking can amplify poisoning instead of filtering it, and even the strongest evaluated models follow injected directives on roughly half of queries and up to 89% on some corpora. Cosine-based scoring also underreports semantic compromise by more than 5x, motivating the need for calibrated judge-based evaluation. We then present GuardRAG, an adversarial training framework that converts these observed vulnerabilities into security-aware preference data. GuardRAG teaches the model to produce a span-grounded security report, making each refusal or acceptance decision auditable. On an 8B model that follows about 45% of stealth injections even when the security policy is included in the prompt, GuardRAG reduces behavioral attack success to 0.3% while preserving benign-query utility. The resulting 8B model also matches the defense quality of a 70B model nearly nine times its size on a single GPU. However, securing the immediate response does not prevent poisoned content from persisting in tool-using agents with long-term memory. We therefore introduce PRISM-Mem, a provenance-aware memory firewall that screens what may enter persistent memory, cutting agent attack success from 30.3% to 9.0% and cross-turn contamination from 98.7% to zero.

Together, RIPE-II, GuardRAG, and PRISM-Mem show that realistic evaluation, security-aware training, and provenance-governed memory can substantially improve RAG and agent security without sacrificing usefulness. This thesis establishes that a threat which grows with model capability can instead be held by a deliberately trained defense across the full retrieval lifecycle.

Monday, August 10

Dingyuan Xue

Chair: Dr. Kelvin Sung
Candidate: Master of Science in Computer Science & Software Engineering
1:15 PM.; Join Dingyuan Xue’s online defense
Thesis: Training-Free Feature-Based Ranking for Multi-Candidate Mask Selection

Recent advances in neural-network-based and promptable image segmentation have improved the accuracy, robustness, and flexibility of object-mask generation. Under ambiguous visual conditions, however, these systems may produce multiple plausible candidate masks for the same prompted target, whereas downstream applications generally expect only one mask. This creates a candidate-selection problem: determining which candidate most accurately delineates the intended target. Existing approaches often rely on learned quality predictors, additional supervision, calibration data, reference datasets, or model-specific modification. These dependencies increase implementation effort and limit portability when the candidate image segmentation system is pretrained, externally supplied, or otherwise fixed.

This thesis presents the ACES (Area, Center-location, Edge-proximity, and Silhouette) framework, a training-free, user-guided scoring system for inference-time candidate ranking. Inspired by classical binary-image analysis, ACES measures the four corresponding geometric and morphological properties of each candidate mask and compares them with a target reference vector to produce a compatibility score. The mask of the highest-scoring candidate is recommended as the final mask.

The ACES framework provides a versatile default reference vector derived through controlled experiments and allows the reference to be modified based on application knowledge and the characteristics of the target geometry. An existing candidate set can therefore be re-ranked under different target references without repeating candidate mask generation. Since the ranking requires only candidate masks and image dimensions, the scoring computation can be applied to outputs from different segmentation pipelines as a lightweight post-processing stage.

The ACES framework is evaluated on the DAVIS 2016 benchmark with candidate masks generated by a Grounding DINO and SAM 2 pipeline. Two settings are examined. The first applies the default reference vector to all sequences without target-specific adjustments. The second derives reference values based on the geometric characteristics of the intended target. In the default setting, ACES increases mean intersection over union from 0.6170 for upstream confidence-based selection to 0.7670. The framework is most effective when at least one candidate mask substantially overlaps with the intended target in scale, location, and shape and thus can be effectively encoded by the user defined reference vector where this overlap remains sufficiently stable across the evaluated frames. Under these conditions, ACES can suppress distractor masks that receive higher upstream confidence and recommend a candidate that more closely matches the intended target. These results demonstrate that reference-based geometric scoring provides an effective, interpretable, and portable basis for multi-candidate mask selection while preserving the existing segmentation pipeline.

Tuesday, August 11

Sicen Liu

Chair: Dr. Kelvin Sung
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Sicen Liu’s online defense
Thesis: LLM-Driven NPCs for Realistic Behavior and Collaboration in Games

Realistic behaviors of non-player characters (NPCs) are important in interactive games and simulations because NPCs influence how players understand, interact with, and experience virtual worlds. Large language models (LLMs) offer new opportunities to develop NPCs that are more believable and socially responsive than traditional scripted or rule-based systems. Existing LLM-based solutions have demonstrated believable social behavior and long-horizon interaction in simulated or game environments. However, these systems do not focus on NPC decision-making that is grounded in a continuously updated environment, or on inter-NPC interactions based on their visible states.

This thesis proposes an LLM-driven NPC framework that supports environment-grounded decision-making and multi-NPC responsiveness. The framework maintains a centralized Global World State that records the states of the environment and the player, as well as the observable states of the NPCs, providing NPC reasoning with explicit information about which entities exist in the simulation, where they are located, and what states they are in. Additionally, the system records and communicates observable NPC states that detail the externally visible conditions of NPCs, including facial expressions, current actions, and objects that the NPC is holding. An LLM-based decision-making pipeline driven by these structured representations generates actions, updates the world states, and maintains NPC behaviors that are consistent with the current environment and the general NPC states.
To evaluate the framework, a turn-based simulation system is implemented. The simulation is advanced in discrete time durations referred to as ticks. During each tick, the system processes the Environment, Player, and NPC Turns in a fixed order, allowing environmental changes, player actions, and NPC reasoning to update the shared world state. The system also records simulation traces and supports analysis through replay and scripted or manual inspections.

The simulation results show that the framework supports environment-grounded decision-making and actions, as well as collaborations among multi-NPCs. In the simulated world, NPC behavior remained consistent with current object availability, object states, locations, and ongoing environmental changes, while also adapting to other agents’ requests and ongoing actions. Together, these capabilities enabled multiple NPCs to maintain coherent coordination through multi-step collaborative activities in a dynamic environment.

Wednesday, August 12

Chandramathi Murugadass

Chair: Dr. Dong Si
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Chandramathi Murugadass’ online defense
Thesis: RNA Motif-Based Atomic Structure Prediction and Analysis Using Deep Learning Of Cryo-EM Density Maps

RNA molecules are central to numerous biological processes, including gene expression, viral replication, and the regulation of cellular machinery. Understanding RNA three-dimensional structure, particularly the recurring secondary structural motifs such as internal loops, bulges, and hairpins, is essential for drug discovery, vaccine design, and interpreting viral mechanisms of immune evasion.

This thesis presents a complete, three-phase deep learning pipeline for automated RNA structure prediction from cryogenic electron microscopy (cryo-EM) density maps, using secondary structural motifs as the primary intermediate representation. The first contribution is a large-scale, open-source dataset – the 3DEM-RNA-Motif dataset – containing over a hundred thousand motif-resolved cryo-EM density map segments paired with corresponding atomic models, spanning 25 classes of RNA secondary structural motifs sourced from the Electron Microscopy Data Bank (EMDB), the RCSB Protein Data Bank, and the RNA CoSSMos database. Each motif instance is stored as a standardised 3D voxel grid with atomic-level annotations propagated to voxel-level labels for the RNA phosphate-sugar backbone, ribose sugar, and nucleobase components. Dataset quality is validated through cross-correlation fitting scores and atom-level Q-score resolvability measures.

The second contribution is a motif classification pipeline. A 3D Convolutional Neural Network (Motif3DCNN) is trained on five consolidated motif classes – symmetric loops, asymmetric loops, hairpins, bulges, and unknown – achieving a peak validation accuracy of 96.8% and a specificity of 0.948 on high-resolution maps (1.5–2.8 Å).

The third contribution is an end-to-end inference pipeline for locating motifs in previously unseen global cryo-EM maps. Fast Fourier Transform (FFT)-based template matching first localises candidate regions across the full density volume; the Motif3DCNN classifier then assigns each candidate a motif class; and finally a MONAI UNETR transformer-based segmentation model predicts per-voxel backbone probability within each detected region. Post-processing converts these probability maps into phosphorus-atom PDB files via skeletonisation and chain ordering. The pipeline is validated on EMD-18332 (yeast 80S ribosome, 3.0 Å resolution).

Connor Wilding Leonie

Chair: Dr. Kelvin Sung
Candidate: Master of Science in Computer Science & Software Engineering
2:00 P.M.; Join Connor Wilding Leonie’s online defense
Project: Design and Prototype of an Adaptive Opponent for Real-Time Strategy Games

Artificial opponents in Real-Time Strategy games remain mechanical with static and deterministic behaviors. Methods exist that allow opponents to adapt, such as large-scale reinforcement learning. However, those methods demand resources beyond the reach of small development teams. This project seeks to develop an engagement-oriented adaptive opponent that can be trained within a limited budget.

The proposed solution is a hierarchical system comprising a modeling agent, a macro agent, and micro agents. Time constraints reduced that solution to a prototype controlling a single unit in Beyond All Reason. The prototype encodes the game state, maintains context via a Long Short-Term Memory network, and ranks candidate positions using weights that are trained on a hand-designed reward structure.

The prototype was tested via a scenario-based evaluation against an untrained agent and the base game AI. The prototype met its objectives in navigation, combat avoidance, and building when each was tested in isolation. However, combining those behaviors exposed friction in the reduced action-selection component.

The delivered system is a first step toward affordable adaptive opponents rather than a resolution of the gap. The prototype demonstrates that the training pathway is viable within a limited budget, and the evaluation identifies the components that would benefit from further development.

Thursday, August 13

Minnu Joseph

Chair: Dr. Min Chen
Candidate: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Minnu Joseph’s online defense
Project: Design and Development of a Machine Learning Application for Crime Prediction

Crime is a major concern that can have a wide impact on both communities and individuals. Knowledge of the crime rate in the neighborhood can be used by people who are concerned about crime for their awareness and to make informed safety decisions.

The project presents an application that utilizes historical crime data from the Seattle Police Department to build three crime prediction models using the XGBoost algorithm that can forecast the likelihood, types, and location of probable future crime. The models effectively predict high risk areas in the three frequently occurring crime categories – violent crime, property crime and other crime. In addition, the model’s prediction results are displayed on an interactive map of Seattle. The city is divided into a hexagonal grid, with each grid cell displaying the predicted probability of the different types of crime for the current day and the following day. The map also displays crime incidents reported during the past seven days.

Master of Science in Cybersecurity Engineering

SUMMER 2026

Wednesday, July 22

Ben Pearson

Chair: Dr. Geethapriya Thamilarasu
Candidate: Master of Science in Cybersecurity Engineering
1:15 P.M.; Join Ben Pearson’s online defense
Thesis: Optimized Multi-Agent Defense Pipeline Against Advanced Prompt Injection Attacks

Large language models deployed in production remain critically vulnerable to prompt-injection attacks. These attacks embed adversarial inputs that hijack the model into leaking data, taking unauthorized actions, or producing harmful content. Existing multi-agent defenses handle static, known attacks well but leave three gaps. First, they are not evaluated against adaptive attackers who optimize against the defense. Second, they do not handle indirect injection through external content or multi-turn distributed attacks. Third, they do not measure how the security layer behaves when the inference cache fills up.

I designed a multi-agent defense pipeline with three architectural contributions. The first is a white-box adaptive adversary. It attacks all defense components at once through a joint-loss optimization. I evaluated it across five configurations and 24 generations of iterative training. The second is a trust-role-aware cache compression scheme. It partitions cached tokens by trust role. It prevents the loss of security-relevant instructions when the cache fills up. The third is a multi-turn defense that tracks conversation history across turns. It aggregates four detector signals into a per-session trust score, resolved through a learned discriminative classifier.

The evaluation produced two headline findings. On an aligned Domain LLM, no configuration produced a verified attack. This held across four evaluation configurations and 24 generations of iterative training. On a vulnerable Domain LLM, the attack succeeded three-of-three times without the pipeline. With the pipeline in place, the attack failed five-of-five times end to end. The multi-turn defense caught every attack pattern on real-user dialogues at near-zero false-positive rate. An early design coupled the false-positive rate and the attack-catch rate through a single parameter. A learned discriminative classifier decoupled them. The cache-layer scheme preserved all protected token classes at full recall. A naive eviction baseline lost them entirely. The perimeter guards reduced attack-success rate by roughly 11×on one benchmark and 70×on another against an undefended baseline.

These results reframe what a multi-agent defense is for. On aligned Domain LLMs, the pipeline earns its value through orthogonal coverage of threat surfaces the alignment training does not address. On vulnerable Domain LLMs, the same pipeline becomes the binding direct-injection defense layer. The two roles are complementary rather than competing.

Friday, July 31

Hemin Rakesh Maru

Chair: Dr. Geethapriya Thamilarasu
Candidate: Master of Science in Cybersecurity Engineering
1:00 P.M.; Join Hemin Rakesh Maru’s online defense
Thesis: Gradient-Aware Adversarial Training for Resource-Constrained IoT Intrusion Detection Systems

In recent years, massive growth in the Internet of Things (IoT) has expanded the attack surface of connected devices. To defend against these attacks, deep-learning-based intrusion detection systems were introduced because of their ability to detect complex attack patterns. However, these deep learning models are vulnerable to adversarial attacks. Adversarial training is one of the most effective defenses against such attacks, but it requires generating adversarial examples which introduces high computational and memory requirements. These requirements pose a major challenge for IoT edge devices, where processing power, memory, and energy resources are limited.

This thesis proposes a Gradient-Aware Adversarial Training (GA-AT) framework that improves the efficiency of adversarial training while maintaining strong robustness and low computational overhead. GA-AT monitors the input gradient magnitude throughout the adversarial example generation process and incorporates three key observations to eliminate redundant computation, which saves CPU resources. Experimental evaluation in a simulated environment demonstrated a 70% reduction in adversarial training time, which also significantly reduced CPU utilization. These results indicate that GA-AT provides an effective balance between robustness and computational efficiency, making adversarial training more practical for resource-constrained IoT edge devices.

Master of Science in Electrical & Computer Engineering

SUMMER 2026

Wednesday, July 22

Jose Fernando Pagan

Chair: Dr. Harry Aintablian
Candidate: Master of Science in Electrical & Computer Engineering
4:00 P.M.; Discovery (DISC) 464 or Join Jose Fernando Pagan’s defense online
Thesis: System for Fine Control of Differential Drive, Over Uneven Terrain

Autonomous land vehicles and naval surface vehicles are a subset of systems that may benefit from improved course tracking and correction systems. Existing kinematic models for drive systems assume ideal conditions where travel occurs over smooth flat surfaces. Outside of the R&D laboratory, thes e vehicles will traverses surfaces that deviate from such ideal conditions. The work undertake n, in this thesis, examines the kinematics and behavior of a 2 wheel differential drive (2WDD) system traversing surfaces with discontinuities.

The 2WDD system was chosen because it is well known and offers a comprehensible baseline to correlate vehicle behavior with forward and in verse kinematics. Such vehicles display specific and predictable behaviors when traversing a surface with discontinuities, specifically, dips or bumps. Vehicle behavior is studied and translated to mathematical models that can be used to demonstrate behaviors in simulation (using MATLAB and Simulink).

Finally, this thesis examines the modeled behavior and proposes a control system to mitigate the undesired affects, of these behaviors. In particular, traversing a discontinuity produce a vehicle turn toward the discontinuity. This behavior is problematic when it is desired that the vehicle travel in a straight line. Thus, this thesis proposes a differential drive feedback control system designed to mechanism mitigate this undesired turning behavior and minimize deviations from the intended course of travel.

Thursday, July 30

Adam Loper

Chair: Dr. Kaibao Nie
Candidate: Master of Science in Electrical & Computer Engineering
3:30 P.M.; Discovery (DISC) 464 or Join Adam Loper’s defense online
Thesis: Specific Loudness Features for Sound Event Classification and Localization in Machine Learning

In sound event classification and localization, the mel spectrogram is the de facto standard input feature. It is, however, only a partial model of human hearing: it reshapes the frequency axis to match the ear but leaves the magnitude axis in raw physical units, ignoring the equal-loudness weighting, masking, and compressive growth that determine how loud a sound is actually perceived. This thesis investigates whether completing that perceptual transform—using the ISO 532 specific-loudness models from psychoacoustics—produces input features better matched to sound categories that are themselves defined by human listeners.

The central study concerns sound event classification. On the widely used ESC-50 benchmark, specific-loudness features significantly outperform both the linear-frequency spectrogram (STFT) and the mel spectrogram under an identical neural network, raising accuracy from 53.1% for mel to 63.0%. A controlled ablation shows that most of this improvement is recovered by a lightweight loudness transform applied directly to an ordinary mel spectrogram, at essentially no additional cost—identifying the perceptual treatment of the magnitude axis, rather than the frequency scale, as the source of the benefit. The effect replicates on two further datasets and extends even to acoustic scene classification. A second study applies the same features to spatial audio, where two-channel binaural loudness matches the localization-and-detection performance of a four-channel ambisonic format.

Together, these results support the central thesis: for sound classes defined by human perception, input representations that incorporate perceptual loudness align more closely with classification targets than representations based on physical signal properties alone.