ABSTRACT. Mike Singleton is a Principal Engineer and software architect whose career includes leadership roles at SingleOps/Granum, OnSolve, Improving Enterprises, and Honeywell. He has modernized large legacy applications, designed high-performance micro-frontend platforms, accelerated CI/CD pipelines, launched multimillion-dollar products, and led company-wide adoption of AI-assisted engineering. Mike has also managed and mentored high-performing engineering teams while establishing standards that improve security, developer experience, and delivery speed. He holds bachelor’s degrees in Computer Science and Health Science from the University of Tennessee at Martin.

ABSTRACT. Rear Admiral Grace Hopper is well known as a pioneer in our field, especially for her outreach and public appearances. However, there is scant information exploring the source code which she produced. This project started as an effort to find, restore, analyze, and execute code written by Admiral Hopper. This necessitated a detailed study of the technology of the day, and the creation of a simulation of the Harvard Mark I computer. Some of the required code and documentation artifacts were available online in digital form but much of the source material only exists in physical media stored in the archives of the Smithsonian Institute’s National Museum of American History. Other physical artifacts reside at Harvard University and are either on display or in their collection of historical scientific instruments. In both cases, these archives contain details and code once thought lost. In this talk, we will discuss how we gained access to both sets of artifacts, the experience of using both archives, and what information we gained from the visit. We will also include a status report of our efforts thus far, and future directions of the project.

ABSTRACT. In this joint talk, we introduce a general framework for analyzing recursive algorithms. We employ recent techniques from discrete mathematics to derive analytical solutions to a broad class of recursive problems. These results provide a systematic approach to analyzing the computational complexity of recursive algorithms and offer new insights into the behavior and performance of such algorithms.

ABSTRACT. Universities across the United States increasingly leverage K–12 summer camps and outreach programs to promote computing education, broaden participation, and support student recruitment, with most initiatives primarily targeting middle and high school students. Since 2017, the Department of Computer Science and Information Technology (CSIT) at a regional public university has developed and sustained a continuous fall, spring, and summer coding camp and class model mainly serving students in grades 3–12. Supported initially by three Google grants, this initiative has evolved into one of the largest K–12 computing outreach efforts in the state, engaging approximately 400 students annually and around 2,500 participants over the past decade. The on-ground coding camp program is designed to be accessible, community-centered, and cost-effective, thereby lowering barriers to participation. The program has strengthened the department's community engagement initiatives while filling the computing education gap in the region. Additionally, raising awareness of CSIT programs among local students and families could indirectly support undergraduate enrollment efforts. This paper describes the program's design and evolution, reports preliminary observations, and outlines directions for future assessment and program evaluation.

ABSTRACT. Remote privileged access to data centers is a prime attack surface, and generative AI can now defeat static liveness checks with synthetic faces and cloned voices. Conventional privileged access management authenticates once at login and cannot confirm that the biometric capture is live or untampered. We present ZT-PAM, a three-layer Zero Trust design. Layer 1 fuses facial geometry, infrared depth, physiological signals, and synthesis-artifact analysis to verify a live subject. Layer 2 binds a digest of the liveness verdict and capture to the challenge of a TPM 2.0-signed FIDO2 assertion, giving the server cryptographic evidence of device-attested user presence. Layer 3 continuously scores session risk and triggers step-up re-authentication or termination. We define a threat model with four adversary classes (replay, real-time synthesis, OS-level compromise, insider session transfer), specify the attestation protocol, and compare client-server and edge-node deployments.

ABSTRACT. Nutrition tracking is important because it sheds light on the discrepancies between the nutrients we think we’re consuming and what we’re actually consuming. However, most existing nutrition apps don’t make this easy. They rely on manual food searches and repetitive data entry: you type in an ingredient, scroll through a dozen nearly identical database entries, guess at the portion size, and repeat that for every item on your plate. This summer, I worked alongside Dr. Volodymyr Kindratenko and Dr. Sharon Donovan at the University of Illinois Urbana-Champaign’s National Center for Supercomputing Applications to continue developing Visual Nutrition, a Flutter app with an AWS Amplify and GraphQL backend that lets users log meals by photo, barcode, voice, or search and tracks diet quality. Building on this foundation, I improved usability by redesigning the tab layout to make navigation clearer and expanded functionality by adding an “Add a Recipe” feature. I also refined the app’s AI components by comparing Illinois Chat and OpenAI models for our in-app chatbot, finding that although using Illinois Chat keeps user information within university infrastructure, its latency was much higher than that of many OpenAI models. These changes reduced the burden of nutrition tracking, making it easier for users to stay on track with their goals. We are currently testing the app with UIUC students and partnering with Design for America students to further refine its usability.

ABSTRACT. The increasing integration of electronic systems and connectivity features in modern vehicles has created new opportunities for digital forensic investigations. Vehicle forensics involves collecting and analyzing data from vehicle systems to reconstruct incidents, investigate accidents, and understand driving behavior. The On-Board Diagnostics II (OBD-II) port, widely used for vehicle diagnostics and monitoring, provides access to parameters such as accelerator position, brake position, and diagnostic trouble codes. This data may offer valuable insights into driving behavior and complement information obtained from other vehicle data sources, including event data recorders (EDRs). This project explores the collection and analysis of OBD-II diagnostic data for driver behavior classification. OBD-II diagnostic data will be collected from different vehicles driven by 5-10 participants recruited from friends and acquaintances. Then the collected data will be analyzed using machine learning algorithms to identify the data parameters most relevant to distinguishing between safe and reckless driving. The project will also examine how driving behavior assessments obtained from OBD-II data compares with the approaches used in insurance telematics programs. Finally, the project will evaluate the potential use of diagnostic logs as evidence in vehicle forensic investigations. The project is in progress. The data collection part will be partially completed in the next month. In the conference, a few samples of OBD-II data and their analysis and preliminary results will be demonstrated.

ABSTRACT. Artificial intelligence (AI) agents present potential to accelerate digital forensic investigations by automating investigation tasks and reducing investigative backlogs. However, their tendency to generate inaccurate or unsupported findings presents significant challenges, when forensic evidence must be handled securely and conclusions must be verifiable. Recent developments in AI-assisted forensic frameworks have demonstrated the potential of guardrails to reduce hallucinations. However, the effectiveness of these safeguards with locally hosted, open-weight language models remains insufficiently explored.

This project evaluates the reliability of locally hosted large language model (LLM) agents for digital forensics triage using the SANS SIFT Workstation. Three open-weight model, DeepSeek-R1, Qwen3, and Llama 3.1, will be evaluated across four harness configurations that progressively introduce forensic safeguards: shell access, read-only typed tools, mandatory citations to tool outputs, and an adversarial skeptic review. Using classroom-based and publicly available forensic investigation cases with established findings, the project will assess detection accuracy, evidence recall, precision, hallucinated-finding rates, citation validity, consistency, and susceptibility to prompt-injection attacks embedded in digital evidence.

The experimental evaluation will compare model performance across configurations to investigate the effects of forensic safeguards and model selection on the reliability of AI-assisted triage. A controlled prompt-injection experiment will further examine whether restricting access to typed, shell-free forensic tools reduces the risk of malicious instructions influencing investigative findings. All experiments will use reproducible configurations and documented evaluation procedures.

The project is in progress. The experiment results will be shared at the conference.

ABSTRACT. Quantifying rates of land use and land cover change matters because of its far-reaching effects on biodiversity loss and climate change. The New England Landscape Futures Project (NELF) brought together researchers as well as land use and conservation practitioners to co-develop scenarios of potential futures in New England based on different visions of how land use might change from 2010 to 2060. Since we now have land-use data from 2020, this study reviews the differences between the 2020 NELF simulations and the 2020 observed data. I investigated the extent to which observed changes in specifically development and conservation from 2010 to 2020 align with the NELF scenarios by using total disagreement, a local Moran’s I test, and the absolute difference between the simulated maps and observed data. I found that the "Go It Alone" scenario matched closest for rates and placement of conservation, whereas the "Connected Communities" scenario was closer in matching placement of development. It was difficult comparing new development because of differences in classification in the land use database for NELF compared to the classification system I used. Additionally, the NELF scenarios overestimated the correct amount of conservation due to large sections of Maine receiving conservation in the reference period of 1990-2010, but much less between 2010-2020. This study shows that development has become more clustered and that there is less conservation compared to recent trends.

ABSTRACT. Claims have been made that learning to use version control systems (VCS) is challenging and difficult. In the study presented in this paper researchers introduced a VCS early in CS2 to see how challenged students were and how the system was used. This paper presents a post-hoc study used to measure participant use of a VCS while developing programs in CS2. With only a brief (10-15 minutes) introduction and a single homework showing how to use it, participants were asked to use the repository for the remainder of the semester. Participants were not graded based on their repositories. By the end of the semester, all participants were using the VCS. Participants were logically grouped as either above average (a Final grade of A or B), average (C), or below average (D or F). When comparing the above average participants with the below average, the above average group had statistically more lines of code committed, more commits, and fewer lines of code per commit. This suggests that these two groups of participant types use their repositories differently. Additionally, it shows that participants will make effective use of development tools even when provided a minimal amount of instruction.

ABSTRACT. In the field of computational electromagnetics, specifically in the agricultural internet of things, bridging the gap between numerical engineering models and real world data is an ongoing challenge. Engineers and researchers often need to map a real world scenario to a computer simulation, and a popular method is to use pre-packaged general Finite Element Method (FEM) software. However, such closed source software introduces two major issues. First, verifying that the simulation is operating correctly is difficult and secondly, optimizing its performance is limited. The cause of both these issues is that the software is a black box, i.e., it does not expose its implementation. In this work, we introduce the algorithms and procedures required for simulating Through-the-soil wireless power transfer (TTS-WPT) open-region systems utilizing highly configurable, readily available, open source libraries, therefore allowing control, validation, and performance optimization of the complete system. We show how to solve a magneoquasistatic (MQS) potential at long ranges using the recommended algorithms and procedures in this domain, which necessarily differ from similar problem domains in the literature. Using our approach, our results are within a 2% NRMSE of an analytical single-dimensional point-source model. Moreover, the validation we provide can be used to characterize an MQS distribution within a lossy medium, such as soil.

ABSTRACT. The IETF Web Bot Authentication (Web Bot Auth) protocol uses HTTP Message Signatures to authenticate automated clients by allowing them to cryptographically sign HTTP requests. Clients include AI agents that acquire cryptographic identity. The protocol distinguishes between authentication and authorization: it authenticates the sender but does not determine whether a request is authorized. That boundary is deliberate, but what has attracted less attention is what security properties remain once the boundary is drawn. We build a conformant signer, verifier and key directory, and use them to ask the scope of a valid signature. A signature meeting the specification's minimum requirements is accepted on any request to the same host while it remains fresh. We then test the two mechanisms proposed for tightening this and find they answer a narrower question than the one that matters: both stop an observer who captures a signature, and neither touches a legitimate agent turned against its operator, because such an agent signs correctly over a connection of its own. What remains is one normative rule directed at operators rather than verifiers, and compliance with it is byte-identical to violation. A systematic reading of the working group's complete public record finds this has not been raised. We also measure what the cheaper mechanism costs, a figure asserted in the group's own discussions and never quantified.

ABSTRACT. Our project aims to implement real time head tracking and off axis projection to create a realistic parallax effect. Traditional implementations of this technology often require specialized hardware or setup; we will achieve this effect using a standard webcam. We will achieve this by using computer vision for head tracking, real time 3D rendering, and off-axis projection to adjust the perspective of the 3D environment based on the user's position. Computer vision libraries will allow us to detect facial landmarks on a user's face, and from those captured points we can convert the 2D points into a 3D vector in space. Using this calculated vector in space we can translate the user’s movements in real life into simulated movements in our created 3D environment using off axis projection. The final effect will simulate a window into a 3D environment, this would allow the user to look around objects as if they existed in the physical space. During the creation of this project we aim to discover both practical and educational applications using this technology, such as viewing a 3D print model before printing to get a better sense of scale or dynamically viewing anatomical models from different perspectives.

ABSTRACT. Transport-layer encryption is widely assumed to conceal user activity. However, it protects only the contents of network traffic, not the shape of said traffic. Prior research has shown that metadata can reveal sensitive user behavior, but the practical cost of defending against such leakage is rarely quantified. This project investigates whether a passive network observer can classify user actions within an encrypted mobile finance application using only timing and packet-size metadata. We evaluate how effectively common server-side mitigations reduce that leakage relative to the overhead costs in implementing them. We construct a test application using a Flutter mobile client, a FastAPI and PostgreSQL backend, and a dedicated capture setup that records packet-level traffic for offline analysis. Six user actions are captured across multiple encryption and mitigation configurations. Supervised classifiers are trained on extracted traffic features to measure how distinguishable each action remains. The study quantifies the resulting privacy–utility tradeoff, reporting the reduction in classification accuracy each mitigation achieves against the latency and bandwidth it imposes, and examines whether the underlying encryption scheme meaningfully affects that outcome. The findings clarify the extent to which encrypted mobile traffic leaks user behavior and provide developers with a comparative, cost-aware basis for selecting traffic-analysis defenses. This work contributes an empirical, reproducible evaluation framework linking side-channel attack feasibility directly to the measured effectiveness of its countermeasures.

ABSTRACT. The University of Tennessee Space Institute (UTSI) provided our team with useful imaging data from the Mach4 Ludwig Wind Tunnel, located at their campus. This project studies the problem of the effects when a particle in the freestream disrupts a shock. We intend to analyze the dataset to categorize any common patterns or irregularities from the particle-shock interactions. The data provided to us is read through a custom program developed using pyMRAW and OpenCV libraries in Python. This program allows the data to be viewed as a playable video using .cih header file coupled with the corresponding .mraw files. As part of our plan, we will run through each of the datasets provided and formulate the location and cause of particle shocks. After we have established consistency, we will graph the comparisons we made to demonstrate our classification. A summary showcasing our findings will also be provided. Once the base program of our project is completed, it will serve as a contribution to UTSI researchers. In conclusion, our hope is to provide a program that may be improved or expanded in the future as well as aid the future of air and spacecraft research development to withstand high speed winds.

ABSTRACT. This research evaluates the effectiveness of machine learning for the detection of phishing emails. Phishing attacks continue to pose a threat to millions of users around the world. Three different machine learning models will be trained and tested to determine whether an email is legitimate or phishing. The three models being tested will be Logistic Regression, Random Forest, and Support Vector Machine. Publicly available email datasets will be used to train and test the models, and libraries found in Python such as Scikit-learn and Pandas will be used to analyze the data. At the end of the testing, the three models will be compared using a number of metrics to evaluate their performance. These metrics will be accuracy, precision, recall, F1-score, and false positive rate. Ideal results will include a high success rate of classifying phishing emails while maintaining a low rate of false positives. With these findings, we hope they will be able to contribute towards improved phishing detection methods, decreasing the chance that a phishing email is successful in its attack.

ABSTRACT. With Artificially Intelligent systems becoming inevitable in every facet of modern life, the need to defend, deploy, and instruct on the use of such systems is more relevant than ever. The ethics, standards, best practices and rules of engagement continue to evolve. A couple of years ago I created a special topics course in using AI and ML in Cybersecurity in our Master of Science in Cybersecurity and Information Assurance program. Since that time Agentic AI has become dominant as both defense and adversary. The evolution of one such course will be discussed.

ABSTRACT. Although service-learning is widely recognized as a high-impact practice linked to improved student engagement, persistence, and professional identity, upper-division cybersecurity curricula often rely on isolated sandboxes or theoretical instruction, creating a gap in authentic, operational experience. Building on prior research examining short-duration service-learning for first-year computing students, this study addresses this gap through a required, five-hour service-learning initiative embedded within a primarily theory-based 300-level Secure Software Development course at a regional public university in Northwest Tennessee. In partnership with campus Information Technology Services (ITS), upper-division Computer Science and Cybersecurity majors designed and individually executed live vishing (voice phishing) simulations targeting campus personnel, observed real-time operational outcomes, authored executive summaries, and completed structured reflections. Using an exploratory qualitative case study design, we are examining student reflection responses collected via the Canvas LMS to evaluate shifts in professional identity, computing self-efficacy, sense of belonging, and career development. While qualitative analysis is ongoing, initial coding suggests that bridging theoretical concepts with authentic campus security operations helped students begin framing themselves as active cybersecurity practitioners. Early patterns also suggest gains in technical confidence and greater clarity regarding their interests and career pathways within computing, while students impact on their sense of belonging appear more varied. While bounded by a single cohort, these emerging findings offer preliminary insights into how operational campus partnerships can support professional identity development and career readiness in upper-division computing curricula.

ABSTRACT. The notion of geometric visibility plays a fundamental role in many applications such as robotics and shortest path computation in the presence of obstacles. One of the most fundamental data structures in visibility is the visibility graph (VG). We are given a graph G with labeled vertices v0, v1, . . . , such that the labeling corresponds with a Hamiltonian path H. G also may contain other edges. We are interested in determining if there is a terrain T with vertices p0, p1, . . . , such that G is the visibility graph of T and the boundary of T corresponds with H. G is said to be persistent if and only if it satisfies the so-called X-property and Bar-property. We show that there exists a persistent graph G that is not the visibility graph for any terrain T. We obtain this result by introducing a new style of reconstruction algorithm. We show that if one can compute a set of feasible x-coordinates for the points of the terrain, then the y-coordinates can be computed via linear programming (LP).

ABSTRACT. Large language model (LLM) use is growing rapidly across organizations, but commercial LLM services raise concerns about data privacy, cost per token, and energy use. Running open-weight models on local hardware is one alternative, but local models often run slowly. LLMs generate text one token at a time, and on local hardware each token requires reading every model weight from memory. As a result, single-user inference is limited by memory bandwidth rather than computation, leaving much of the processor idle. Speculative decoding addresses this bottleneck by having a small draft model propose several tokens that a large target model verifies in a single pass, with a mathematical guarantee that the output distribution is unchanged. Its benefit is governed by an analytic model relating speedup to the acceptance rate, the draft length, and the relative cost of the two models. This analytic model has been validated mainly on datacenter GPUs. Does it hold on consumer-grade hardware? This work tests it on an Apple Silicon workstation using open-weight models, measuring acceptance rates, draft-length tradeoffs, and speedups across task types, and comparing them with the theory's predictions. The results indicate how much AI work can practically shift to local machines.

ABSTRACT. Bioinformatics is a rapidly growing field of research, especially among undergraduate researchers, due in part to research initiatives like Characterizing Our DNA Exceptions (CODE) from HudsonAlpha Institute for Biotechnology. In the case of CODE, characterizing DNA exceptions takes a prohibitive amount of time and compute power to run the simulations needed to gather data towards the classification of a variant. While research is known to be time consuming, my independent research study on the COLQ gene has over 1600 unclassified variants to analyze. Extrapolate this across the entire human genome, and it's clear that efficiency is a priority. The goal of this project is to provide a tool for students to identify good candidate variants of uncertain significance (VUS) for study, to assist in the initial discovery phase, and head in the direction of fully classifying the variant. Using a machine learning suite originally developed for SERPINA1 as a baseline, this study has been expanded to provide results for any individual protein found in the ENSEMBL API. Combining the data from various predictive algorithms housed in the ENSEMBL database (CADD, PolyPhen, SIFT, etc.), as well as homolog conservation scores (ConSurf), the program uses multiple models found in PyTorch and scikit-learn to generate a list of prioritized variants for student research projects. The results of the program will include a justification for their significance and any contradictory results.

ABSTRACT. More than one billion people worldwide live with a mental health condition yet support and treatment reach only a fraction of those affected, making mental health one of the most under-resourced areas of modern health care. Within that population, only some conditions attract significant research funding and technology investment while others remain largely invisible, drawing disproportionately little research support relative to their real-world burden. At the same time, today's rapid advances in software platforms, real-time collaboration, AI, and privacy engineering make it newly feasible to build assistive technology for mental health, which makes the gap between what technology could do and what it actually serves more urgent rather than less. My solution, Team Calendula, is an independent software project created to narrow this gap: it develops technology for conditions that receive little mainstream attention, with its first focus being dissociative disorders. Its first product, Systemiser, supports this goal: designed around the unifying need for continuity when dissociative symptoms disrupt memory, identity, and the organization of daily life, it lets users represent the distinct identity states that characterize these conditions, track transitions between them, maintain a structured record of events, status, and functioning that bridges memory gaps, and share those details selectively through a granular privacy model. Systemiser is being built as an application ecosystem spanning desktop, mobile, and web, with a real-time collaboration layer and privacy-focused cloud-hosted data storage. It also plans gateways into third-party platform categories, including messaging and collaboration platforms such as Discord and Slack, and in-vehicle assistant platforms such as Apple CarPlay and Android Auto, serving as an example of how platforms that already shape daily digital life can be turned toward assistive purposes. Prototype testing has exercised four core feature cycles to date: identity-state representation and transitions, status and functioning tracking, privacy-gated sharing, and collaborative information management, with additional cycles under active development to address disorder-specific needs, including grounding techniques and crisis-support tools. I present Systemiser as the first case study in Team Calendula's broader trajectory: a growing family of products that address more under-represented conditions over time, closing the mental-health technology gap condition by condition.

ABSTRACT. This project proposes the development of an in-house beamforming array to demodulate audio and spatial transmissions over the frequency spectrum at a reduced cost through the repurposing of microcontrollers. The project aims to modify a recently discovered debugging function for WI-FI calibration in ESP32 microcontrollers to reconfigure the on-chip radio receiver to receive I/Q samples beyond and below the WI-FI bands. The proposed system will use an array of ESP32-C5 microcontrollers to capture spatial radiolocation and audio transmissions, while connected to a field-programmable gate array (FPGA) for demodulation, filtering, tuning, and analysis. The project will investigate whether an array of ESP32 microcontrollers can reliably replace radio and location arrays at a fraction of the cost. Digital signal processing (DSP) filters will be implemented to specify target bands, while software will control the receive only capabilities of the array of microcontrollers. The received signals will be processed through the programmable software and logic of the FPGA, and will be displayed or played using a graphical interface and a speaker. The system will be evaluated through dynamic analysis and figures of merit to determine the functional capabilities for spatial and radio signal intelligence. The expected outcome is a multi-modal phased array capable of demonstrating reliable reception capabilities in the frequency spectrum at a low cost.

ABSTRACT. Leaving the hospital is a risky moment for patients. Missed follow-up appointments, medication mix-ups and missing support at home can send patients back to the hospital within 30 days. This project tests whether an AI assistant can help hospital staff catch these problems before discharge. We built an AI system that reviews a patient’s discharge information step by step. It finds what is blocking discharge, lists the tasks still to be done, and suggests next steps. Safety rules built into the system require human approval for risky actions, block actions only a clinician should take, and record every decision. We compare this system with two simpler approaches: a fixed checklist and a single AI prompt with no safety rules. All three are tested on 150 realistic but made-up patient cases covering five conditions and three difficulty levels. We will measure how many problems each system finds, how complete its discharge plans are, and how well it follows the safety rules. No real patient data are used.

ABSTRACT. Cyberattackers usually launch attacks against a target system through a chain of intermediate hosts, for the purpose of hiding the origin of the attack. This type of intrusion is known as stepping-stone intrusion (SSI). Existing SSI detection methods using packet crossover focus on analyzing the crossover phenomenon in the whole connection chain. A recent approach to detect SSI computes the packet crossover ratio over a given period and compares it with a calibrated threshold to estimate how many connections lie in the downstream chain. This known algorithm simply matches a given Send packet with the first Echo, even when more follow-up Echo packets corresponding to the same Send may exist, which is possible due to high fluctuation caused by the intermediate routers on the Internet. In this paper, we propose an improved method to detect SSI more effectively and efficiently using packet crossover. Whenever there are multiple Echo packets corresponding to the same Send, our detection algorithm finds the last Echo packet to match the corresponding Send, which makes the calculation of packet crossover ratios much more accurate. As a result, our proposed SSI detection algorithm can more precisely estimate the length of a connect chain. We design and conduct network experiments on the Internet environment to verify the effectiveness, correctness and efficiency of our proposed SSI detection algorithm.

ABSTRACT. This project examines whether machine learning can estimate owner-reported residential property values while supporting interpretable feature-level explanations. The study uses 1,229,112 eligible owner-occupied one-family housing records from the 2020–2024 American Community Survey Public Use Microdata Sample (ACS PUMS) for California, Florida, New York, Tennessee, and Texas. Records from 2020–2023 were used for model development, while 268,930 records from 2024 were withheld for later-year evaluation. Elastic Net, Random Forest, CatBoost, and XGBoost were compared using leakage-controlled preprocessing. A development-only feature-reduction process reduced the selected XGBoost specification from 21 to 18 predictors while preserving nearly identical development performance. On the 2024 evaluation data, the reduced XGBoost model achieved a mean absolute error of $275,639, a median absolute error of $113,951, and an R² of 0.318. TreeSHAP was computed for all 2024 records. State-PUMA produced the strongest global SHAP attribution, while first-mortgage payment, household income, bedrooms, construction period, and non-bedroom rooms were also important. The same final model and its local TreeSHAP explanations were implemented in the Housing Value Explorer, where users can enter a housing profile, view a model estimate and its main feature contributions, and examine selected scenario changes. The project is intended for research and model interpretation rather than professional appraisal.

ABSTRACT. With the increasing use of artificial intelligence and machine learning, especially in areas such as healthcare and finance, protecting sensitive user data has become an important concern. Machine learning models use data to learn patterns and make predictions on new or unseen inputs. However, using these models may require sensitive information to be shared for processing. One way to address this concern is through homomorphic encryption, a technique that allows computations to be performed on encrypted data without revealing the original information during processing.

In this study, we use TenSEAL, a library built on Microsoft SEAL, with the Cheon–Kim–Kim–Song (CKKS) encryption scheme. CKKS supports approximate mathematical operations on encrypted data. This makes it useful for certain machine learning applications, but the approximations can introduce small numerical differences that may affect predictions. Our research investigates whether quantization-aware training (QAT), a technique that prepares machine learning models to work with reduced numerical precision, can help neural networks maintain their predictions during encrypted inference.

We conduct experiments using two publicly available medical datasets: Wisconsin Breast Cancer, containing 569 records and 30 features, and Cleveland Heart Disease, containing 303 records and 13 features. For each dataset, we train two neural networks with identical architectures, consisting of one hidden layer with eight neurons and a square activation function. The first model uses standard training, while the second uses the Brevitas library to implement QAT. Because CKKS and TenSEAL do not directly support the rounding operations required for hidden activation quantization, we create a QAT-informed version that retains the quantized weights but removes hidden activation quantization

We compare the models using plaintext inference and evaluate the Standard and QAT-informed models using three CKKS precision configurations. We examine both prediction accuracy and whether the encrypted models produce different predictions from their plaintext versions.

Our results show that the QAT-informed models achieve higher classification accuracy and f1 score in the selected experimental runs. These findings suggest that QAT-informed neural networks can be evaluated using CKKS. However, there is a need to evaluate the performance of QAT-informed models on larger datasets.

ABSTRACT. With the increasing number of cyberattacks, network intrusion detection systems (IDS) play an important role in identifying suspicious network activity and potential security threats. Machine learning has become a useful approach for intrusion detection because it can identify patterns in network traffic and distinguish normal activity from potentially malicious behavior. However, using a large number of features does not always lead to better performance. Some features may be more important for detecting suspicious activity than others.

This study explores how reducing the number of features affects the performance of a Random Forest-based intrusion detection system using the NSL-KDD dataset. Random Forest is a machine learning algorithm that combines the predictions of multiple decision trees to make a final classification. We developed and compared six models to classify network traffic as either normal or anomalous. The first model served as a baseline and used all 122 encoded features. Four additional models used the top 10, 15, 20, and 30 features, selected based on their importance scores from the baseline Random Forest. The final model used Principal Component Analysis (PCA), a technique that reduces the number of features by combining information from the original features into new components while preserving as much variation in the data as possible. In our experiment, PCA retained 95% of the variance and reduced the data to 88 principal components.

We compared their performance using accuracy, recall, F1-score, ROC-AUC, and training time. The baseline Random Forest achieved the highest accuracy of 99.91%. However, the model using only the top 30 features achieved a very similar accuracy of 99.89%. The PCA-based model achieved 99.75% accuracy but required considerably more training time.

Our findings suggest that reducing the number of features does not necessarily lead to a significant loss in intrusion detection accuracy. In particular, selecting the most important features allowed the Random Forest model to maintain performance close to the baseline while using substantially fewer inputs. However, PCA did not provide an improvement in accuracy or training time in our experiment. Overall, this study highlights the importance of examining how different feature reduction techniques affect machine learning models used for network intrusion detection.

ABSTRACT. Satellite imagery that provides very valuable information for environmental monitoring, geographic analysis, and many other applications. However, the sensitive information associated with these images can create many security risks and privacy concerns if misused, or disturbed. This project will investigate the design and development of a secure database system made for storing and managing satellite imagery and associated metadata while also restricting access to authorized users (RBAC). The system uses Sentinel-2 satellite imagery as a primary dataset and stores image data and metadata in the PostgreSQL database. A python-based application programming interface will provide access to the database, while role-based access control will be used to assign different permissions to users being administrators, analysts, and viewers. Sensitive images and metadata including geographic coordinates and timestamps will be encrypted. With encryption keys managed separately from the database to reduce the risks of a potential database compromise. The system will be evaluated through encryption and decryption tests, role-based access tests, unauthorized access attempts, and simulated database breaches. The expected outcome is a functional prototype that demonstrates how encryption, access control, and separated key management can work together to protect satellite data. This project contributes to a practical approach to bringing awareness and improving the security of satellite – image databases while maintaining controlled access for legitimate users.

ABSTRACT. Within the concentration of Dynamical Systems in mathematics, there is a category of systems called “attractors” that are constructed by a set of values that evolve over a period of time around a basin of attraction. Types of regular attractors include Fixed Point Attractors (finite) and Point Attractors (closed loop). Though, unlike regular attractors, there exists non-linear attractors defined by their chaotic sensitivity to minute alterations to initial conditions. These systems are called Strange Attractors. Using Python, attractors can be graphed and observed. This can be done by utilizing the Matplotlib visualization library and the NumPy numerical processing and scientific computational function library. Matplotlib provides many features including visual customization with color mapping and various graphical and plotting tools. My presentation will cover the graphing of Strange Attractors in Python and its extensions, creating beautiful representations of particle evolution and chaotic behavior.

ABSTRACT. Large language models can answer many cybersecurity related questions, but their responses are often inaccurate, not specific, or are unsupported by any source. This project explores whether Retrieval-Augmented Generation (RAG) can improve the answers of a small language model by augmenting it with cybersecurity related information. The system uses the Trendyol Cybersecurity Instruction Tuning Dataset from HuggingFace as its domain corpus. Records from the dataset are processed into documents, divided into chunks, converted into embeddings using DGE, and stored in ChromaDB. When the user asks a cybersecurity related question, the system retrieves the most relevant information from the corpus and provides that context to SmolLM2 before an answer is generated. The project will compare different chunking and retrieval strategies, including baseline dense retrieval and selected advanced retrieval methods. A held-out set will be used to evaluate answer quality, relevance, faithfulness, and the language model's ability to provide grounded responses. The project will also examine cases where retrieval or generation performs poorly. The goal is to determine which design choices are most useful for a small cybersecurity-focused RAG system.

ABSTRACT. Artificial intelligence (AI) is being explored for both offensive and defensive cybersecurity applications, including automated security testing and anomaly detection. This research will evaluate how effectively locally hosted AI agents attempt to gain access to a system and detect and respond to those attempted breaches in an isolated Red-Team/Blue-Team environment.       The experiment will use two connected virtual machines running locally hosted AI models, connected through a virtual network with no internet access. The Red Team agent will attempt to breach the Blue machine and create a harmless marker in the file system. The Blue Team agent will attempt to monitor security logs and system activity and take defensive actions to detect and prevent unauthorized access. Trials will begin under consistent conditions and be repeated over a predetermined number of times. The actions and decisions of both agents will be recorded for analysis.       Performance will be evaluated using Red's objective completion rate, Blue's detection rate, time to detection, the number of successful defensive responses and the false positive rate during benign activity. These findings will help show the capabilities and limitations of locally hosted AI agents in a controlled cybersecurity environment and inform further research into AI-assisted security.

ABSTRACT. An archivist appraises, preserves, organizes, and provides access to records, documents, and media of enduring historical, cultural, or legal value. Most U.S. municipalities however lack a dedicated archivist to process and preserve archival materials. As a result, many historically significant artifacts are managed locally, stored in local libraries, or remain in private hands. Currently, the National Archives allows the public to contribute to their catalog with a long list of meta-data requirements; and with Citizen Archivist program it allows people of interest to contribute to transcribe and enhance discovery the National Archives Catalog. These efforts only try to address the internal problem of the National Archives and do not solve the problem of (1) not having enough trained archivists nor (2) the meager public interests to address the large number of artifacts in the public domain. Although archival artificial intelligence (AI) has advanced in information extraction, retrieval, and metadata enrichment, it remains limited by the (3) absence of an OAIS-compliant agentic framework and structured mechanisms for public contribution. We proposed Crowd-to-OAIS Agentic Preservation Framework (COAP-Framework) addresses these challenges by combining cloud-native public intake, autonomous multi-agent processing, cross-repository semantic linking, dual-verification workflows, and cryptographic authorization into a unified digital preservation pipeline.

ABSTRACT. In 2023, lesson summaries were presented as an effective tool for data-driven teaching practices. Developed to improve student engagement after 2020, these brief post-lesson forms included self-assessment and summarization components. After generative AI tools became widely available in 2022, submitted summaries became increasingly AI-generated. In Spring 2026, one submission included the AI prompt followed by multiple pages of gibberish. Considering the growing body of research on the impact of generative AI on higher education, this experience is certainly not unique. Meanwhile, during the development of a neural networks course for the Fall 2026 semester, a prospective student asked, “What will the class be like?” The response was another question: “Where do we want to be at the end?” Given a rapidly changing computer science landscape, two priorities emerged beyond mastering course content. The first was the ability to communicate clearly about the subject, and the second was participation in undergraduate research resulting in an artifact. With these goals in mind, students work as a team to define an open-ended research question, identify appropriate datasets, apply disciplinary methods in specified computing environments, and iteratively refine their experiments. The project culminates in a poster suitable for submission to an undergraduate research venue. Several weeks into the semester, college-level initiatives involving Course-Based Undergraduate Research Experiences (CUREs) were introduced. The five essential characteristics are the use of scientific practices, discovery, broadly relevant or important work, collaboration, and iteration. This led to the realization that a course potentially qualifying as a CURE had been born.

ABSTRACT. What does it take to build and improve a substantial teaching application with an AI coding partner? This talk examines my ongoing experience developing CompileIt, a programming platform for intro CS students, through a summer and fall of AI-assisted development. Across roughly 1,550 prompts and 578 commits, the project expanded to include browser-based code execution, debugging, automated grading, classroom management, and tighter Canvas integration. Drawing on project history and concrete examples, I will discuss how the collaboration worked, where AI accelerated progress, how an AI mistake almost cost me $1,000 in hosting fees, and where human judgment, testing, and repeated debugging remained essential. The talk offers a practical look at the time, effort, possibilities, and limitations of using AI to develop software.

ABSTRACT. Healthcare administrative workflows often involve multiple systems, manual coordination, and time-consuming information retrieval, creating challenges for patients and healthcare providers. This work presents the design and production deployment of an intelligent healthcare workflow automation platform developed by GenXBots LLC to support patient onboarding, appointment scheduling, health plan comparison, and provider access to clinical information.

The platform employs a multi-agent architecture in which a Large Language Model (LLM)-based router identifies user intent and delegates requests to specialized agents responsible for knowledge retrieval and appointment scheduling. Developed using Python and FastAPI, the architecture supports API-based workflow execution and fallback routing for requests that cannot be assigned through the primary routing process. Its healthcare knowledge component incorporates Retrieval-Augmented Generation (RAG), FAISS-based semantic retrieval, Google Vertex AI embeddings, and the Gemini language model to generate context-informed responses from healthcare documentation. Authentication is implemented within the broader production application.

Reported production use cases include patient assistance with appointment scheduling, comparison of health insurance premiums, deductibles, and benefits, and provider access to clinical-record-based patient summaries. Qualitative operational observations suggest reduced administrative effort and improved accessibility of healthcare information; however, these benefits have not yet been established through controlled quantitative evaluation.

This work contributes an applied architectural case study of LLM-based agent orchestration integrated with healthcare administrative services and identifies opportunities for further evaluation of routing accuracy, workflow efficiency, clinical information reliability, and security.

ABSTRACT. Computer Science majors are often taught to create their own code from scratch, however they rarely have to learn to work backwards from someone else’s project. Thus never building the ability to review, critique, or understand systems that already exist when the original plans or source code isn’t available. This lack of reverse engineering experience creates a significant gap between academic training and industry demands, where software maintenance, security auditing, and legacy system updates dominate day-to-day engineering responsibilities. This presentation explores practical pedagogical strategies for introducing reverse engineering methodologies into the undergraduate computer science curriculum.

ABSTRACT. A honeypot is a tool that is used to lure in hackers and allow them to “break-in” so they can learn more about their behavior and attack patterns. The honeypot acts like a normal server with a fake file system and everything. Unfortunately, some hackers can see repeating patterns or find the end of the fake file system and realize that they have been tricked before they back out. Our goal is to add an AI through to it to try and come up with responses for commands that the system doesn’t accept. This will make the hacker stay in the server longer, allowing us to see more logs of what methods of attack they use. The honeypot we chose is Cowrie. It is an open-source honeypot where we will add an API for an AI to interact with it and generate those responses we need. We developed a middleware to handle these interactions, and it also adds some consistency to the experiment by making sure the simulated file system stays consistent through every command and it also allows for response-caching, allowing the same message to generate the exact same response to prevent self-exposure as a honeypot. We will run a vanilla baseline Cowrie model with our LLM powered model in parallel so that we can then track metrics such as depth, drop-off-point, and all the different commands they run. We hope this experiment will allow us to further our knowledge and let us better combat hackers in the future.

ABSTRACT. Machine learning is increasingly being incorporated into intrusion detection systems to improve the identification of malicious network activity. However, improved detection performance does not necessarily demonstrate that machine-learning models are reliable when attackers intentionally attempt to evade them. This research investigates the reliability and potential weaknesses of machine-learning-based intrusion detection when exposed to both conventional and evasive network attacks. The study will establish a baseline environment using a Kali Linux attacker, Windows target system, and Suricata intrusion detection system, while also training a Random Forest machine-learning model using the same attack categories. Both approaches will be evaluated against denial-of-service attacks, port scans, and brute-force activity to establish baseline detection performance. Modified versions of these attacks will then be introduced to evaluate how detection behavior changes under evasive conditions. The resulting detection rates and failure patterns will be compared to identify conditions in which machine learning may improve, reduce, or otherwise alter intrusion detection reliability. The expected contribution of this work is a clearer understanding of whether machine-learning models provide dependable security benefits when incorporated into network intrusion detection systems, particularly when facing adversarial or evasive behavior.

ABSTRACT. We delve into the research and standards for how computer science is taught at a lower education level. Specifically pointing our studies toward a 3rd grade level in order to better understand where these skills start developing. After developing a good understanding of what educators might find useful when teaching at this education level we developed a website that houses our ideas in which an educator can take and better implement into their own curriculum. These ideas consist of simplified interactive activities as well as animations for developing understanding. In a world that is constantly evolving with technology we hope to shine a light on the importance of developing computer skills as soon as possible.

ABSTRACT. Recent advances in speech synthesis have made audio deepfakes increasingly realistic, raising concerns about the reliability of existing detection methods. Although current detectors perform well on clean benchmark datasets, audio encountered in practice often undergoes various transformations before reaching a detector. Common operations such as compression, transcoding, noise corruption, and playback and re-recording can modify the acoustic characteristics of speech and interfere with the features used to distinguish genuine audio from synthetic speech. Understanding how these transformations affect detection is therefore important for assessing the practical reliability of audio deepfake detectors.

In this work, we present a systematic study of audio deepfake detection under real-world transformations and degradation. Using the WaveFake dataset, we first establish the performance of representative detectors under clean conditions and then examine their behavior across different audio processing and recording scenarios. Our evaluation considers both digital transformations, including lossy compression, resampling, filtering, and additive noise, and physical-world degradation introduced by speaker playback and microphone recording. Except from measuring detection accuracy, we analyze how different transformations influence classification errors and investigate whether transformation-aware robust training can improve detection reliability without changing the underlying model architecture.

Our experiments show that detectors with near-perfect clean accuracy can experience substantial performance degradation under transformed audio. Different transformations also produce distinct error patterns, with some causing synthetic speech to be classified as genuine and others increasing false alarms. Robust training improves detection under several digital transformations, but physical-world degradation remains challenging. These findings highlight the need to consider realistic audio transformations when developing and evaluating reliable deepfake detection systems.

ABSTRACT. Climate change is one of the most significant global challenges of today. Due to the constant influx of information, the long-term impacts of climate change can often be misconstrued or be hard to fully understand. This project presents a solution to this problem, by creating an interactive model utilizing historical data, 3D modeling, and a linear projection model. The goal is to provide an engaging and educational tool that clearly projects the future effects of climate change in a specific area, in this case the Mont Blanc. Using real historical data, long-term trends were identified and fed into a linear projection model, which was then used to predict possible future trends. These projections are integrated into a 3D environment that visually changes a model of the Mont Blanc and allows users to explore the potential changes over the next 50 years. By combining projected changes with an interactive visual representation, an effective tool is created that can educate a wide range of individuals and that almost anyone can use. This project contributes to climate change education and educational format that can be reused in numerous situations.

ABSTRACT. As XR technologies advance and become more accessible, it is imperative to explore their applications in multiple industries and technologies. Such industries include those that have high bars of entry, dangerous scenarios, or lack of qualified professionals to provide guidance. With the development of our XR experience Circuit Space, we want to provide an environment where users can learn, interact, and experiment with the foundations of circuitry through guided walkthroughs on components as well as a sandbox. The goal of Circuit Space is to lower the barrier of entry into circuitry by providing an accurate hands on experience providing all the tools necessary for basic circuit design. The software will be tested and developed on the Meta Quest 2 with the emphasis being on built in hand tracking. The Meta XR Standard Development Kit allows for the use of the headset’s built in features, mainly the use of hand tracking allowing for seamless control with intricate objects. Unity will handle complex systems like rendering, collision, and events to facilitate the experience of working up close to a breadboard. Success in this project will be the accessibility and accuracy of electrical components to their real world counterparts. XR will allow users with varying degrees of background the ability to design and experiment with the foundations of circuitry.

ABSTRACT. In this study we develop SCAIR, the Satellite Cyber AI Range, to provide a simulated Low Earth Orbit range for researchers to generate datasets for AI training, with the hope of enabling research into space cybersecurity and cyber intrusion. There are several studies into satellite cybersecurity, but very few publicly available datasets for use. This research provides the means to address that. We employ several techniques to make SCAIR as realistic as feasibly possible. We utilize commercially-off-the-shelf hardware from EdgeFlyte CubeSat V1 to develop a real hardware environment for satellite constellations. Utilizing NRF2040L1 32-byte radios, we develop a satellite constellation and ground station firmware with communication and basic defense capabilities. Said radios are equipped with typical radio packeting formats found on most satellites. We leverage an API to embed constellations with realistic telemetry datapoints from General Mission Analysis Tool (GMAT), a simulated-time orbital propagation system developed by NASA. Various domain-specific noise is embedded at the GMAT and firmware level. We analyze our next steps towards SCAIR development as well as next steps afterwards. Finally, we analyze a few important limitations in our work, most notably in lack of hyperrealism and the small number of defensive & attacking methods.

ABSTRACT. Identifying potential research collaborators can be challenging, particularly as the volume of scientific publications continues to grow. Previous research has investigated the role of expertise similarity and complementarity in research collaboration. Building on this foundation, this study explores a lightweight approach to identifying potential collaborators through semantic analysis of Computer Science publications.

The study focuses on a small dataset of publicly available research paper abstracts in a selected management area. Natural language processing techniques will be used to represent publication content and measure similarities among researchers' interests. The proposed approach will compare traditional keyword-based matching with semantic similarity using text embeddings and cosine similarity. Potential collaborators will be ranked according to their research alignment, and a sample of recommendations will be manually examined to assess their relevance.

The study aims to investigate the feasibility of using publication abstracts to support research collaborator discovery and to identify opportunities for improving recommendation methods in future work.

ABSTRACT. MindBridge: An Interactive System for Accessible Mental Health Support

Antania Bivens, Jaelynn Perry, Muhammad Rahman

Accessing the right mental health resources can be difficult for people trying to find a therapist, manage their own well-being, and keep up with ongoing support. As the awareness of mental health continues to grow, identifying suitable resources and maintaining consistent engagement with care can be difficult. Here we introduce MindBridge, an integrated application designed to give access to those in need to care for their mental health. It provides psychoeducational-based tools and resources to help users better understand their conditions and learn how to cope with them.

The application answers guided questions with an interactive chatbot that will query a list of therapists and available resources based on the user’s main grievances, listed prices, preferred schedule, and meeting times. Users will have access to a progress tracker letting them log day-to-day mood boards and complete daily tasks and activities that they and their therapists can also see. It also includes a community board to keep you connected and engaged with others around you.

This application is also a great tool for therapists whether you work for a specific company or have your own practice. They will be able to create a profile listing your specialties, prices, and availability, manage booking slots for your new/returning clients. They can create a dedicated care plan full of exercises and activities clients can practice in their own time or with the therapist themselves. It can be used as a tool for promotion and an easy way to build up a case load of clients.

With a database full of information and resources at the fingertips, systems set in place to keep user information secure and protected, and interactive activities and CBT tools, users will be able to start their mental health journey right here on their time and pace.

Keywords: digital mental health, database, chatbot, web application, psychoeducation, CBT, therapist matching