School of Computing & Analytics

Title: An AI-Driven Conversational Context-Aware Scam Detection

Authors: Trey McGonnell and Dr. Shahid Noor

Department: School of Computing & Analytics

Abstract: Instant messaging and dating apps such as Messenger, Tinder, and WhatsApp are becoming increasingly popular, but they also serve as central targets for scammers. Email providers such as Gmail and Microsoft Outlook protect users using scam filters. However, these filters may not detect extended conversations, in which a scammer gradually builds trust with the victim by emotionally engaging the victim and personalizing the scam. To address this issue, we present an AI model trained to assess the relevance of messages exchanged between victims and scammers. Additionally, our model analyzes the strategies scammers use during extended conversations and generates a score based on message relevance and the likelihood of a scam. For future work, we plan to validate our model using a larger dataset and further experiments to demonstrate its effectiveness and feasibility.

Title: Mapping Food Access Inequities in Northern Kentucky: A GIS Analysis of LILA Areas and Grocery Store Accessibility

Authors: Gabe Potter, Abbott Rauch, Chase Yunker

Department: School of Computing & Analytics

Abstract: Many Northern Kentucky residents live in areas classified as Low-Income/Low-Access (LILA) by the U.S. Department of Agriculture, where limited access to affordable, healthy food contributes to obesity and diet-related chronic diseases. In early 2026, the grocery chain Publix expanded into Northern Kentucky by opening stores near existing grocery retailers. This study evaluated whether the Publix locations would improve food access for LILA residents. Geographic Information Systems (GIS) were used to create a LILA-obesity risk index, identify high-risk census tracts, and perform a distance analysis to Publix locations. Results showed the Publix stores are not located within LILA tracts and provide limited benefit to populations with the greatest need. The analysis suggests that locating a Publix store within one of Boone County's two highest-risk LILA tracts would improve access to healthy food for underserved residents. These findings can inform future grocery site selection and support equitable access to food.

Title: Mapping Residential Lead-Based Hazards and Vacancies in Northern Kentucky

Authors: Tiffany N. Frederick

Department: School of Computing & Analytics

Abstract: Residential lead hazards remain a major public health concern in Northern Kentucky, where approximately 60% of homes were built before 1978 (U.S. Census, 2024). Combined with a severe shortage of affordable housing, especially in urban communities such as Covington, these hazards lead to a critical need for targeted intervention. This study used ArcGIS to map residential lead hazards and housing vacancies using records of childhood elevated blood lead levels (EBLLs) reported to the Northern Kentucky Health Department from January 2025 through July 2026, along with department data on vacant lead-hazard properties. Spatial analysis identified a high concentration of hazardous vacant properties in Covington’s 41011 ZIP Code. The resulting maps provide a decision-support tool for prioritizing HUD-funded lead abatement and affordable housing initiatives, demonstrating how GIS can guide evidence-based public health planning and direct resources to communities at greatest risk.

Title: PolyDomain Privacy-Preserving Multi-Model Architecture

Authors: Dharmikkumar Patel and Rasib Khan, Ph.D.

Department: School of Computing & Analytics

Abstract: This study introduces a privacy-preserving and user-centric AI model integration architecture for connecting diverse domain-specific models. The architecture safeguards user identities, proprietary metrics, and sensitive intellectual property that organizations must keep confidential. An intelligent orchestrator routes queries to specific domain models, restricting data exposure to the minimum required per task. A central aggregator then combines these outputs into a single, cohesive response. For robust governance, an AI-powered Retrieval-Augmented Generation (AI-RAG) policy engine automates user-level access control, authentication, and role-based privacy compliance. Unlike rigid rule-based systems, this engine automatically retrieves, interprets, and applies new regulatory or organizational policies without manual re-engineering. Ultimately, this multi-model architecture minimizes sensitive data exposure, enforces strict privacy-protected interactions, and scales across diverse industry domains to deliver dependable, secure knowledge exchange.

Title: Prompt Injection in AI-Assisted Peer Review: Evaluating Defenses and Model Robustness

Authors: Nathan Feiler, Linh Nguyen, Dr. Nicholas Caporusso, Dr. Nazmus Sadat

Department: School of Computing & Analytics

Abstract: The increasing use of large language models (LLMs) introduces security risks like prompt injection, where instructions in content may be misinterpreted as commands. This is especially risky when LLMs evaluate scholarly manuscripts, as hidden malicious instructions could influence reviews and acceptance recommendations. One mitigation is using defensive prompts warning about prompt injection, but its effectiveness in manuscript evaluation remains unclear. This study examines whether defensive system prompts reduce the influence of adversarial instructions. We also assess if newer LLMs are more resistant to prompt injection. We evaluated models from Claude, Gemini, ChatGPT, and DeepSeek using standard and defensive system prompts. Our results show that defensive prompts vary in effectiveness across model families; although they improve detection, they can also increase false positives, reducing reliability. Newer models also provide some inherent resistance to prompt injection. These findings highlight the potential and limits of prompt-based defenses and the need for stronger safeguards.

Title: ThinkPilot: An Adaptive Decision Framework for LLM-Assisted Programming Education

Authors: Mackenzie Glaser, Junxiu Zhou, Ph.D

Department: School of Computing & Analytics

Abstract: Large language models (LLMs) have become an increasingly prevalent tool in programming education, raising concerns that their ability to generate immediate, high-quality answers may discourage the development of students’ critical thinking and problem-solving skills. ThinkPilot is an adaptive framework that promotes active learning by guiding students through the reasoning process rather than immediately providing complete solutions. Within a controlled programming-learning environment, students interact with an LLM while solving programming tasks. ThinkPilot analyzes students’ questions, responses, and interaction history to identify observable dialogue states and determine the most appropriate pedagogical action. Depending on student progress, it provides guiding questions, Socratic prompts, conceptual hints, worked examples, or direct answers only when appropriate. By adapting assistance to each learner’s needs, ThinkPilot encourages independent reasoning, fosters critical thinking, and supports effective human–LLM collaboration in programming education.