Weeds are a major burden in small and local farming communities in the United States due to the lack of technology, awareness, and education. Weed control is one of the biggest factors that affects crop production. Manual weeding gives maximum unique control of the weeds in the field. However, manual weeding has high labor intensity and high labor costs. This makes weed management difficult for small and local farmers in the state of South Carolina, resulting in loss of crop yield and poor quality production. In this work, the AI-based iOS mobile app is used to identify weed plants. In our work, we have successfully implemented an iOS mobile app to capture a weed image and identifies weed-type using CNN, CreateML, Xcode and Swift programming language. Our model was tested with our database of weed plants with a precision of 96%. Our results show that the iOS mobile app developed successfully identifies the weed plant. Our future work will include testing the iOS mobile app with more weed plant data to provide precise weed identification.

This paper presents a methodology for utilizing Large Language Models (LLMs) to create AI-driven scaffolding materials in Computer Science. As enrollment in online and hybrid courses grows, the capacity for instructors to provide individual guidance to learners diminishes. This study introduces the "A-H Pedagogical Framework," grounded in cognitive load theory and scaffolding principles. The methodology utilizes a Human-in-the-Loop AI workflow to analyze legacy educational materials (e.g., Jupyter Notebooks, PDFs), identify content and delivery method, and then propose improvements based on the "A-H Pedagogical Framework." Through a case study of GIS Data-Processing assignments, this framework demonstrates how AI can enhance student learning by providing step-by-step, scaffolded guidance while reducing instructors' workload. The AI-driven approach enables scalable education and fosters learning experiences that are otherwise unattainable in large online classes.

Fingerprints offer a reliable and unique means of identification and hence are crucial in fields such as law enforcement and personal identification. However, fingerprint images are hardly of good quality. They may be corrupted and degraded with elements of noise owing to many issues including deviations in skin and impression circumstances. We test some commonly available fingerprint image processing codes and show some experimental results to see the progress and challenges of fingerprint image enhancement, segmentation and minutiae extraction.

Using Generative AI in the classroom to improve student learning requires deliberate, structured integration into course activities and assessments. This paper explores the integration of a specialized Generative AI mentor, Dennis AI, into three sections of an undergraduate Computer Programming II course. The mentor was developed in conjunction with ibl.ai, a family-owned and operated company located in the technological hub of New York that specializes in building AI-driven, revenue-generating systems for the educational sector, serving learners from over 400 universities. The research evaluates a pedagogical shift from traditional passive learning to an active, generative model where AI serves as a "First Responder" for complex technical concepts like memory management and object-oriented design. By leveraging the ibl.ai platform to synthesize course-specific data, the study demonstrates how structured prompt-based inquiry and iterative code generation can bridge the "complexity wall" often encountered in mid-level computer science curricula[1]. Furthermore, the study illustrates how the integration of AI mentors and digital avatars---mapped to Bloom's Taxonomy 4.0---facilitates a scalable model for high-level architectural mentoring and diagnostic practice[2]. Ultimately, this study posits that the strategic deployment of specialized Generative AI tools creates a dynamic environment that encourages deep inquiry, competitive innovation, and a rigorous validation process that preserves academic integrity.

The rapid evolution of artificial intelligence (AI) and computing technologies has fundamentally reshaped workforce demands, necessitating a transformation in how computer science (CS) education is designed and delivered. Traditional CS curricula often struggle to keep pace with industry innovation, provide personalized learning experiences, and equitably support diverse learner populations. This paper presents a comprehensive framework for developing effective computer science program curricula integrated with AI-driven educational models to enhance student learning outcomes. The proposed approach aligns curriculum design with competency-based education, industry relevance, and adaptive AI technologies such as intelligent tutoring systems, learning analytics, and personalized content recommendation. Through curriculum mapping, instructional design models, and AI-enabled assessment strategies, the paper demonstrates how AI can improve student engagement, mastery, retention, and employability. A case-based implementation model and evaluation metrics are presented to guide institutions in adopting scalable, ethical, and inclusive AI-enhanced CS education.

Building a pathway from AI coursework to AI careers requires a clear understanding of what students experience between the classroom and the job market. This paper reports results from a feasibility study supported by the University System of Maryland Elkins Transformation Award at Bowie State University (BSU). The study examined barriers to student participation in AI-aligned experiential learning and assessed institutional readiness for scaling career-connected learning. The feasibility study gathered input from 200 students, 15 faculty members, and 10 industry partners to assess constraints, capacity, and partnership needs for expanding AI-aligned experiential learning. We describe the study methods and key findings, then discuss what they suggest about the importance of experiential learning centers as a strategy for strengthening AI workforce preparation at HBCUs. The work builds on several years of experiential learning work, including a Tech Pipeline Program featured in The New York Times describing how BSU developed employer partnerships to strengthen student career outcomes.

The CyberAICorps Scholarship for Service (CyberAI SFS), formerly the CyberCorps® Scholarship for Service (SFS), supports cybersecurity education in exchange for post-graduation government service. This paper describes Bowie State University’s CyberAI SFS program and shares early lessons from implementation since 2023. The program has supported 14 scholars and has strengthened recruitment by offering a funded, structured pathway into government cybersecurity careers. We summarize the student experience and program operations, including cohort support, mentoring, professional development, certifications preparation, research and conference participation, and structured preparation for internships and government employment. We report outcomes in aggregate, including internship placements and early cybersecurity-related government employment, and highlight the operational choices that helped students participate fully while meeting program requirements. The paper concludes with takeaways for HBCUs exploring CyberAI SFS participation and an invitation to an April 2026 virtual workshop for institutions interested in joining as mentoring partners and building toward future readiness.

The Right to Be Forgotten (RTBF) is a legal requirement that allows individuals to request the deletion of their personal data from digital systems. However, in modern machine learning environments, fully removing data is technically challenging once it has been incorporated into trained models. This research investigates whether machine unlearning can serve as an effective mechanism for supporting RTBF by removing the influence of specific data from a trained model. The study evaluates a pre-trained neural network using multiple forget set sizes and applies Membership Inference Attacks (MIA) to measure whether deleted data remains detectable after unlearning. Experimental results show that while machine unlearning preserves performance on retained data, it does not fully eliminate the influence of forgotten data, as residual information remains detectable across all tested configurations. These findings demonstrate that machine unlearning alone is insufficient to guarantee complete data deletion and highlight the need for stronger verification methods and complementary strategies to support RTBF compliance in AI systems.

SQL injections are one of the most common and dangerous vulnerabilities found in web applications, even though they have been well documented for decades. This paper explores the effectiveness of common prevention techniques against SQL injection attacks, including input validation, parameterized queries, and prepared statements. To demonstrate, a vulnerable web environment was created using Damn Vulnerable Web Application (DVWA) to simulate attacks and observe how each defense method withstands different injection attempts. The results will show the strengths and weaknesses of each approach when tested against real-world attack patterns. In addition to testing, this research highlights the relevance of SQL injections in today’s cybersecurity environment, shown by their inclusion in the OWASP Top 10 [4]. By demonstrating how easily unsecured applications can become victims of attacks and how effective proper countermeasures can be, this paper highlights the importance of implementing secure coding practices in modern web development.

As artificial intelligence increases and is used in daily life activi-ties, the need to understand and interact with artificial intelligence has become important and is now emphasized in undergraduate and graduate programs. Even though AI is being taught, some top-ics such as computer vision (DETR, YOLOv8, Faster R-CNN, and SSD) remain difficult for students to understand and incorporate into practice. Without a definite percentage indicating how many students are affected, current discussions continue to show that computer vision is one of the topics learners struggle to grasp [3]. To address these challenges, a structured framework of hands-on labs in computer vision can support students in strengthening their comprehension at both undergraduate and graduate levels. A hands-on lab is a structured learning activity in which students actively perform tasks, experiments, or problem-solving activities using tools, technologies, or data rather than relying solely on lec-tures. This experiential approach requires learners to interact with software, equipment, or real-world datasets to apply theoretical concepts in a practical context. Hands-on labs help students un-derstand complex AI and computer vision topics by transforming abstract concepts such as convolution, feature extraction, and ob-ject detection pipelines into concrete, interactive experiences that enhance understanding. By working directly with detection mod-els like YOLO, SSD, Faster R-CNN, and DETR, students develop stronger intuition, reduce cognitive overload, and build practical skills needed to apply these systems in real-world scenarios. Prior research in computing and engineering education indicates that project-based and hands-on learning approaches significantly im-prove student comprehension, engagement, and overall learning outcomes [4].

A significant challenge in Video Frame Interpolation (VFI) is reducing the processing time without significantly sacrificing visual quality. To address it, we developed a computationally efficient motion-focused VFI methodology, based on Google's FILM (Frame Interpolation for Large Motion) model. Our proposed approach, Motion-focused FILM, selectively interpolates only the most dynamic areas of the video to reduce the computational load and processing time. We also implemented a tensor bucketing strategy to reduce computational overhead. We evaluated our approach on the DAVIS 2017 dataset. The results show a 95% reduction in processing time compared to the full-frame method. It achieved a Peak Signal-to-Noise Ratio (PSNR) score that was approximately 88% of the baseline, indicating a discernible loss in pixel-level fidelity. However, it retained over 87% of the structural similarity (SSIM) test, suggesting that the overall structure of the interpolated image remains mostly intact. We also investigated the impact of video resolution on our approach's performance.

This paper presents a structured pedagogical framework for integrating blockchain decentralized application (dApp) development into cybersecurity education through experiential, project-based learning. The framework is implemented via a series of laboratory modules that immerse students in authentic, project-based activities grounded in real-world use cases, including blockchain-based diploma and transcript verification, cryptocurrency (Bitcoin) forensic analysis, and secure supply chain management. Through these activities, students gain practical exposure to distributed systems, smart contracts, and security mechanisms within blockchain environments. Preliminary classroom outcomes demonstrate increased student interesting in learning dApp development, enhanced conceptual understanding of blockchain and cybersecurity principles, and improved preparedness for industry and research-oriented roles. Overall, the proposed framework aims to (1) strengthen students’ skills and awareness of blockchain source code vulnerabilities, along with associated detection and mitigation techniques; (2) systematically integrate blockchain vulnerability concepts into information technology and cybersecurity curricula; and (3) prepare future IT professionals with a solid understanding of blockchain attack surfaces and defensive strategies in real-world contexts.

This research project explores a novel approach to bolstering digital forensics by integrating AlienVault, a leading security platform, with blockchain technology. By harnessing the capabilities of AlienVault for real-time threat detection and incident response and leveraging the immutable nature of blockchain for data integrity, this study proposes a framework for enhancing the reliability of digital forensic investigations. The tools used include AlienVault’s Open-Source Security Information Management (OSSIM) platform for security information and event management (SIEM). Ethereum’s blockchain-based ledger is used to log events detected by AlienVault OSSIM, ensuring each event log entry is time-stamped. Data sources for this study include a controlled setup network and the Open Threat Research (OTRF) Security Dataset of Windows event logs. These sources provide a comprehensive and realistic range of cyber-attack scenarios. By utilizing these datasets, the research evaluates how well the integrated system can detect and store threat information. The system’s performance is assessed based on its accuracy in identifying attacks, the speed of its incident response, and the reliability of its forensic data. The expected result is a blockchain-enhanced forensic framework that mitigates common challenges in digital forensics, such as data tampering and chain of custody issues.

The rapid expansion of artificial intelligence (AI), high-performance computing (HPC), and Science Gateway technologies in higher education has created new opportunities for experiential learning while introducing complexity for faculty seeking structured and policy-compliant integration. This paper presents the CLASS AlignED (Course Learning & Analytics Support System) Framework, a faculty-centered, governance-aware architecture designed to support the incorporation of AI, HPC, and Science Gateway technologies into undergraduate curricula. The framework connects institutional AI policy, structured instructional workflows, scalable computing infrastructure, and measurable learning analytics into a unified alignment model. Using a design-based research approach, we describe the conceptual model, system architecture, and planned prototype implementation. The framework proposes embedding compliance logic and cyberinfrastructure guidance directly within AI-assisted workflows to reduce faculty uncertainty and translate institutional policy into actionable instructional practice. CLASS AlignED offers a scalable model for institutions seeking to expand AI- and HPC-enabled instruction while maintaining governance, reproducibility, and equitable access to advanced cyberinfrastructure resources.

Business Email Compromise (BEC) attacks constitute one of the most financially damaging cyber threats, resulting in global losses exceeding 2.7 billion USD annually according to the FBI Internet Crime Complaint Center. Unlike conventional phishing attacks that deploy malicious payloads or URLs, BEC employs sophisticated social engineering via carefully crafted language, posing substantial challenges to traditional signaturebased detection systems. This work develops a robust machine learning framework for automated BEC detection, incorporating 58 specialized features extracted from email content, metadata, and behavioral attributes. We provide a formal mathematical formulation of the feature extraction process and evaluate five gradient boosting algorithms—XGBoost, LightGBM, CatBoost, Random Forest, and a stacking ensemble on the Kaggle Fraud Email Dataset (9,239 samples). The dataset undergoes an 80/20 stratified split to preserve class distribution. CatBoost attains the highest performance, with 97.29 percent accuracy, 97.29 percent F1 score, and 99.55 percent AUC ROC. We employ McNemar’s test to confirm statistical significance (χ2 = 7.52, p < 0.01) and utilize SHAP (SHapley Additive exPlanations) to isolate linguistic metrics specifically text entropy and readability—as primary discriminators. Furthermore, we present a computational complexity analysis demonstrating that our pipeline operates with O(L) linear complexity relative to email length, achieving sub 10 ms inference latency suitable for real time SIEM integration. The framework outperforms existing benchmarks by 8.8 percent in F1 score, establishing a new baseline for content centric threat detection.

Artificial Intelligence (AI) and automation has been one of the most popular topics of the century. It has been integrated into education, medicine, and transportation to make tasks easier and increase productivity. AI and automation is constantly being discussed for being a double edged sword. On one side, it makes tasks capable of being completed faster and oftentimes more accurate, but the other side argues that it lacks accountability. Before AI and automation is introduced to a system , it is integral to analyze the pros and cons of said implementation. This research focuses on the integration of AI and automation in sports and how its risks can affect athlete data, match outcomes, and device reliability. Past research has examined accuracy and health benefits; however, device security, data flow, and contingency planning in the event of an attack or breach has not been acknowledged. This study investigates these gaps by examining how AI-driven wearables and automated officiating systems function; how data flow is mapped; and what vulnerabilities affect confidentiality, integrity, and availability. Literature reviews and input from sports management and computer science professionals, athletes, and trainers will illustrate what protections are currently available and what they think of the trade-offs. The results will highlight key risks, identify potential solutions, and offer guidance for sports organizations and tech developers so that AI and automation in sports can be both innovative and secure.

Misinformation that spreads through social media platforms is becoming very important since artificial intelligence (AI) is helping the spread of misinformation. A lot of misinformation spreads through manipulated social media posts, and this can cause people to reduce their trust in what they see online, and some people may even be easily persuaded and believe what they see in these manipulated posts. Since artificial intelligence is growing and becoming the newest big thing, it’s being used to create posts that are manipulated, and this makes it hard for people to distinguish between real and manipulated posts. This research will analyze how well people can assess social media posts and what factors play a role in their ability to determine what posts are real from the ones that are manipulated. A survey was conducted where people had the opportunity to choose what post they believe is manipulated, and then they were asked what made them choose that choice, and then they rated their confidence level. After gathering all of the results from the survey, the results will be compared with tools that are already created for being able to detect misinformation that’s generated by AI. Comparing the human results from the survey with the detection tools will help determine if humans’ ability to spot misinformation is just as good as the detection tools. This research will highlight the difficulties of spotting misinformation in AI-manipulated posts, and it will also show how the cybersecurity side of things can help people continue to trust what they see online with the help of detection tools.

When it comes to the safety of autonomous vehicles using computer vision, we must first analysis the impact and risk that adversarial patches may present to its occupants, other drivers and property on the road. By seeing the result of the patch with the Ultralytics YOLO 11 model trained on things that an autonomous vehicle might encounter on the road we can perform an analysis of what could’ve been the impact without needing to run a simulation.

Using Claflin University’s robots, we will program them to support our outreach efforts in computer science. The goal is to engage potential students and visitors with our computer science department through interactive demonstrations. A series of programs has been developed to illustrate the importance and diversity of computer science using the NAO humanoid and Unitree Go2 robots. These activities aim to instill basic computer science principles such as sequencing, conditional statements, sensor-based decision-making, and human-robot interaction. During outreach events, the robots perform demonstrations to help participants see how code translates into physical actions, demystifying programming, and robotics. By providing an interactive and visually engaging experience, computer science can be made more accessible to individuals with varying levels of technical background. This project demonstrates how robotics can be an effective tool for computer science outreach, increasing engagement, encouraging curiosity, and fostering interest in computing disciplines.

As Artificial Intelligence (AI) becomes a powerful force within the technological field it is being integrated into all fields. As AI improves, optimizing it for everyday life is beneficial for the further development of technology. In this paper, we present our research findings and literature on adversarial examples and object detection. This research builds upon the previous work by investigating and optimizing an unmanned aircraft to be flown with the aid of Artificial Intelligence. We started with classifying and training AI to recognize certain objects on YOLOv11 custom trained models. Then a follow-up using the custom trained model with live drone footage to test the accuracy of the model evaluating how it can be utilized in Beyond Visual Line of Sight (BVLOS). Through this exploration it demonstrates the future of using unmanned aircrafts with support from machine learning.

This research focuses on developing a reliable methodology for detecting hidden MPLS (Multiprotocol Label Switching) tunnels using active network probing. Our proposed algorithm combines geographic anomalies, RTT behavior, hostname patterns, and backbone routing characteristics to distinguish MPLS tunnels from ordinary routing behaviors such as load balancing. Furthermore, our detection pipeline is implemented to automate tunnel identification, estimate tunnel boundaries, and visualize hop-level behavior. The purpose of this research is to create an accurate, reproducible framework for uncovering concealed MPLS structures in modern networks, enabling improved transparency, routing analysis, and network measurement accuracy.

The main problem with the conflict between online anonymity and accountability is that, although anonymity promotes privacy and free speech, it also makes harmful behaviors like trolling and scams possible. In order to understand how mandatory identification policies, like requiring government IDs or facial scans for site access, affect this balance, this study will review previous research. Privacy and security risks will be a major focus because private businesses collect extremely sensitive data, increasing the risk of data breaches, long-term identity theft, and user tracking. In order to make fair policy decisions, a survey will also be carried out to find out how the public feels about these intrusive verification techniques.

The integration of Artificial Intelligence (AI) into cybersecurity has significantly developed advanced threat detection and analysis. However, due to the deep learning nature, the inherent opacity that comes with these “black box” models creates doubts in the decisions of incident investigation. Explainable Artificial Intelligence (XAI) is the backbone of the future of this transparency gap, by utilizing visualization tools to make these decisions more interpretable. This paper examines two feature-visualization tools in cybersecurity: Partial Dependence Plots (PDPs) and Individual Conditional Ex- pectation (ICE) plots. We analyse the differences between PDPs, which are global explanations, by averaging the effects, and ICE plots, which offer local instance-level insights, to show heteroge- neous attack patterns. By evaluating these methods with the focus of improving cybersecurity intrusion detection and malware anal- ysis. This study highlights the necessary balance between clarity and depth to enhance operational reliability in AI-driven security systems.

Fairness and trustworthiness are principles that were implemented into the workplace to allow for the growth of a more competitive career marketplace in which all individuals have equal opportunity at a role regardless of external matters such as race, religion, gender, or sexuality [1]. However, despite the positive impacts these initiatives have had in the work place, in recent years there's been a push back against them. With multiple organizations even going as far as rolling back their programs supporting fairness in the workplace. To combat this we have developed a tool that allows for anonymous insider reporting of a company’s policies through the blockchain. Through the use of this tool we hope to allow consumers and potential jobseekers to make informed decisions about the companies they patronize.

Introduction: Peripheral Arterial Disease (PAD) is a common circulatory problem characterized by narrowed arteries that reduce blood flow to the lower extremities. In 2019 the global burden of disease study attributed over 74,000 deaths to PAD with over 113,000,000 individuals living with the condition globally. Despite its prevalence, accurate and timely diagnosis remains a challenge, often leading to severe complications such as muscular weakness, amputation etc. Study Objectives: This research investigates the use of machine learning specifically Convolutional Neural Networks (CNNs), to enhance PAD diagnosis from CT-Angiography (CTA) images. Methods: The study aims to determine the effectiveness of neural networks in detecting PAD, optimize model performance, and deploy a testing application for clinical use. We have generated PAD ML model by utilizing pretrained ResNet-50 architecture on PyTorch Framework with dataset splits of 80% training and 20% validation, a AdamW optimizer for 500 epochs. Results: Overall, the model shows stable validation performance with high accuracy, F1-score, and AUC. Metrics remain consistent in later epochs, and the selected checkpoint reflects strong generalization behavior. A validation accuracy of ~93–94%, precision of ~0.93–0.95, recall of ~0.91–0.94, F1-score of ~0.93, and an AUC of ~0.97–0.98 was observed from training (Figure 1). Discussion: By improving diagnostic accuracy, this project has the potential to facilitate early detection and treatment of PAD, reducing the risk of severe outcomes. The integration of machine learning models into clinical workflows represents a significant step toward more accessible and efficient PAD diagnostics, ultimately improving patient care and outcomes. Figure 1: Validation Area Under the Curve (AUC) for detecting PAD. Acknowledgment: This study is funded by National Science Foundation South Carolina Established Program for Stem Cooperative Research (SC EPSCoR), AI-enabled Devices for the Advancement of Personalized and Transformative Healthcare in South Carolina ADAPT, RII Track-1, Award Number: 2242812, Claflin University Sub-awardees Tisdale Verlie and Pawar Shrikant.

Introduction: Laser-based collagen biofabrication process is a novel approach for generating customizable 3D collagen structures as an in-vitro tissue scaffold and has the potential for in-vivo tissue engineering. The technology stems from the ability of a femtosecond (fs) laser pulse to optically generate a controlled and localized pH gradient via a two photon (2P) effect that facilitates the generation of collagen protein assembly, called fibrillogenesis, into fibers and bundles. Unlike all other collagen biofabrication techniques, this process utilizes the spatial and temporal precision attributed to light, which allows for accurate 3D modeling of native tissue structure. Tseng et. al, 2020 proposed a framework for a functional biopolymer that could alternate between the two β-sheet structures in response to pH changes. Chiba et. al, 2003 have extensively studied amyloidogenicity showing a significant correlation with the stability of the amyloid fibrils with PH, and little correlation with that of the native state. It has been proposed that the stability of the native state and the unfolding rate to the amyloidogenic precursor as well as the conformational preference of the denatured state is influenced by PH. The most immediate impact of collagen fabrication would be in the creation of custom, patterned cell culture scaffolds. Here, we propose an AI based innovative laser collagen alignment technique to be developed into an in-situ scaffold formation technology for producing next generation of tissue engineering-based implantable biomedical devices. Study Objectives: This research investigates if the use of trained convolutional neural net architecture (CNN) can effectively segment aligned collagen fibers deep inside tissue, this study set out to train a U-Net CNN architecture to accurately classify aligned collagen-positive pixels within a SHG image volume. This AI-enabled collagen image processing opens a new way to create real-time scaffold formation technique that mimic various in vivo tissue structures. Methods: To achieve this, we have used PyTorch, a python library to initialize the conventional U-Net CNN with four encoding and decoding units on 2000 images (Figure 1). The synthetic images were generated utilizing a Stable Diffusion 3.5 Large technique. A non-linear rectified non-linear unit (ReLU) layer comes after two 2D convolution layers in each encoding and decoding block. The network's weights and biases are modified by an adaptive moment (ADAM) optimizer. These findings will assess whether a trained CNN can accurately and precisely segment aligned collagen-positive pixels at a variety of imaging depths assisting our understanding of fibrillogenesis and subsequently collagen biofabrication. Results: Primary training found the model to have a training loss of 0.009 with a precision and recall of > 0.90. However, an overfitting has been observed with this run. To fix this, we intend to increase data (via augmentation), simplify the model by having fewer layers, try regularization (L1/L2), implement early stopping, or will apply K-Fold cross-validation to improve generalization. As an alternative, transfer learning can be used to modify the trained CNN for use in drastically distinct fiber networks or to retrain the network for images at significantly different magnifications. Discussion: Here, we propose an AI based innovative laser collagen alignment technique to be developed into an in-situ scaffold formation technology for producing next generation of tissue engineering-based implantable biomedical devices. This AI-enabled collagen image processing opens a new way to create real-time scaffold formation technique that mimic various in vivo tissue structure. These findings will assess whether a trained CNN can more accurately and precisely segment aligned collagen-positive pixels at a variety of imaging depths. Figure 1: Sample image for aligned and un-aligned collagen fibrils used in CNN training. Acknowledgment: This study is funded by National Science Foundation South Carolina Established Program for Stem Cooperative Research (SC EPSCoR), GAIN CRP Subaward (Grants for Applications in Industry and Networking Collaborative Research Program), Claflin University Sub-awardee Pawar Shrikant.

Abstract PeerConnect is an advanced mobile application designed to enhance academic support and peer-led learning by connecting students with compatible tutors based on their strengths, learning needs, and performance data. Addressing the challenges students face in finding effective academic assistance, the platform leverages predictive analytics and intelligent matching algorithms to recommend the most suitable tutors, optimize session scheduling, and monitor student progress, creating a personalized and data-driven learning experience. The application incorporates features such as user registration, detailed tutor profiles with performance statistics, progress tracking dashboards, session scheduling with conflict detection, and AI-driven recommendations for study materials and tutors. Built as an iOS application using Swift with a Python-powered backend and MySQL database, PeerConnect prioritizes accessibility, user experience, and responsive design, making it suitable for students across diverse academic disciplines. By integrating mathematical concepts such as weighted scoring, regression analysis, graph-based network modeling, and predictive probability, PeerConnect not only strengthens the peer tutoring process but also provides measurable insights into student performance and learning outcomes. The platform fosters a culture of collaborative learning, academic improvement, and knowledge sharing, offering a structured, reliable, and scalable alternative to informal or traditional tutoring services. PeerConnect ultimately aims to enhance student academic performance, engagement, and retention by providing a robust, technology-driven peer support system. Future development may expand AI-driven analytics, subject coverage, and integration with institutional learning management systems, highlighting the transformative potential of math-informed, data-driven educational technology in higher education

Intrusion Detection Systems (IDS) are important for identifying malicious network activity, but they rely heavily on large, labeled datasets for effective training. A major challenge in this field is that real network data often contains Personally Identifiable Information (PII), such as IP addresses and hostnames, creating significant privacy and security risks. This research aims to develop a pipeline that removes PII from network datasets while maintaining their research utility. The objective is to generate high-quality synthetic datasets that enable secure data sharing for IDS without compromising sensitive information. The study utilized the CIC-IDS-2017 dataset to reflect authentic network activity and diverse cyberattack types. The methodology involved a multi-stage process: first, PII detection was performed using pattern-matching for IPs, MAC addresses, and URLs. This was followed by anonymization through tokenization (e.g., replacing IPs with coded labels) and generalization (e.g., converting specific timestamps to dates). Finally, synthetic data was generated using Conditional Generative Adversarial Networks (CTGAN). The primary goal was to create synthetic records that mimic the statistical distributions of the original traffic without exposing individual identities. The effectiveness of the synthetic data was evaluated across three dimensions: Utility, Diversity, and Privacy. In terms of utility, machine learning models trained on synthetic data achieved significant performance improvement compared to models trained on real data. Diversity metrics confirmed that the synthetic datasets maintained consistent label distributions and generated enough unique samples to prevent model overfitting. Finally, privacy was validated by measuring the distance between synthetic records and their nearest real samples. The results showed a significant average distance of 23,952,242.9681 and a minimum distance of 2,583,688.0567, indicating that the synthetic records are not direct copies of the original data. These findings suggest that synthetic data generation is a powerful tool for IDS research, enabling secure data sharing and mitigation of cyberattacks while ensuring total privacy protection. Future work will focus on exploring Graph Neural Networks (GNNs), automated end-to-end pipelines, and techniques to ensure fairness across rare attack classes.

Healthcare systems often store large volumes of patient records in formats that are difficult to search efficiently using traditional keyword-based methods. These limitations can delay care, frustrate providers, and impact outcomes. A solution is needed that understands the context and meaning behind clinical documentation, enabling faster and more intelligent access to related cases. Semantic search is changing the way we manage healthcare information by helping us understand the meaning behind patient records instead of just looking for exact words. In this project, we used Google Colab to build a simple but powerful semantic search system that combines Sentence Transformers and ChromaDB. The goal was to make it easier to find similar patient cases or notes based on meaning. We used a pre-trained transformer model ("all-MiniLM-L6-v2") to turn sentences about patient data into numerical vectors. These vectors were saved and searched using ChromaDB, a lightweight vector database. All coding and testing were done in Google Colab. For example, when we searched for "high blood pressure treatment," the system returned a sentence about "medication for hypertension" - proving that it could understand medical terms even if they were worded differently. This kind of system can make it easier for doctors or medical staff to find relevant records quickly, especially in electronic health record systems. Overall, this project shows how machine learning tools like sentence embeddings can make healthcare data smarter and more useful.

Skin cancer is one of the most common and fastest-growing cancers worldwide. Early and accurate detection is critical to improving patient outcomes, reducing the need for invasive procedures, and lowering mortality rates. This project investigates the application of deep learning, specifically convolutional neural networks (CNNs), for the automatic classification of skin lesion images using the HAM10000 dataset. The CNN model was trained to recognize and differentiate between seven dermatological conditions: melanoma, basal cell carcinoma (bcc), benign keratosis-like lesions (bkl), actinic keratoses and intraepithelial carcinoma (akiec), dermatofibroma (df), melanocytic nevi (nv), and vascular lesions (vasc). To improve model performance and address class imbalance, preprocessing steps such as image normalization and data augmentation were applied. A custom batch data generator was also implemented to efficiently manage system memory and streamline training. Model development was conducted using Google Colab with TensorFlow and Keras, leveraging GPU acceleration for faster computation. The resulting CNN achieved a validation accuracy of approximately 85%, placing its performance within the reported dermatologist accuracy range of 62–80%. These results highlight the potential of AI-driven diagnostic tools in assisting with early skin cancer detection and supporting clinical decision-making. The model is especially valuable for deployment in underserved or remote areas where dermatology specialists may not be available. However, further validation is needed to assess its generalizability to real-world clinical settings, including diverse populations and non-dermoscopic images. With continued refinement, this system could significantly enhance screening, diagnosis, and treatment planning in modern dermatology.

Cardiovascular disease (CVD) remains one of the leading causes of death worldwide, posing a significant challenge to global health systems. Early detection and prevention are critical to reducing its impact, yet traditional diagnostic methods often fall short in identifying at-risk individuals before serious complications arise. This project focuses on developing a machine learning-based cardiovascular disease prediction model that can analyze key health indicators and predict individual risk with greater accuracy and efficiency. This aligns closely with the center’s mission to advance health equity through innovative, data-driven solutions. By combining medical knowledge with artificial intelligence, this work embraces the center’s commitment to improving public health outcomes, expanding access to preventive care, and empowering communities through technology-driven healthcare innovation. This project presents a machine learning-based approach for predicting the risk of cardiovascular disease using patient health data. By leveraging algorithms such as Logistic Regression, Random Forest, and Support Vector Machines, the system analyzes features including age, blood pressure, cholesterol levels, BMI, and lifestyle factors to assess cardiovascular risk. The dataset, preprocessed for missing values and scaled for consistency, is split into training and testing sets to evaluate model accuracy. Performance metrics such as precision, recall, and ROC-AUC scores are used to determine the most effective model. This predictive tool aims to support early diagnosis and preventive healthcare decisions, offering an accessible, data-driven method for identifying at-risk individuals.

In the current academic landscape, the bridge between undergraduate talent and faculty research projects is often built on informal emails or static forms. For professors, reviewing dozens of statements of interest to find specific technical overlaps is time-consuming and ineffective. For students, the lack of transparency in opportunities and how their skills align with them can be a barrier to entry. This project addresses these inefficiencies by providing a centralized, intelligent platform that digitizes recruitment and applies computational linguistics to assist in decision-making. This project follows a decoupled, three-tier architecture designed for scalability and separation of concerns. The frontend utilizes React.js for a dynamic User Interface (UI) that provides role-based dashboards. It handles state management for real-time application tracking and provides a responsive environment for both students and professors/Principal Investigators. The core of the project lies in the backend API, constructed with the Django REST Framework. Django manages the authentication logic, serves as the API gateway, and hosts the Machine Learning (ML) utility scripts. The relational database, MySQL, was chosen for its ACID compliance. This ensures that sensitive student data and application records remain consistent and secure. The core feature of this portal is the integration of the scikit-learn library to perform semantic analysis on text data. The matching process follows a three-stage pipeline: Pre-processing, Vectorization, and Similarity Scoring. To compare a professor's project description with a student's statement and Resume, the text must be converted into numerical vectors. This is accomplished by using Term Frequency-Inverse Document Frequency (TF-IDF) Where tf_{t,d} is the frequency of term t in document d, and the second term is the Inverse Document Frequency, which penalizes common words (like "the" or "and") while rewarding specific technical terms (like "Python" or "Microbiology"). Once the text is vectorized, the Cosine Similarity (2) is calculated to determine the distance between the two vectors. This measures the cosine of the angle between them in a high dimensional space. A score of 1.0 (100%) indicates a perfect keyword alignment, while 0.0 indicates no overlap. This project implements several key features to enhance the user experience. By using Role-Based Access Control, Users are automatically directed to either the Student or Professor dashboard upon login. Keywords in a student's statement of interest or resume are highlighted using the TF-IDF feature names. The system identifies the top 5 overlapping terms between the student and the project, providing a quick, at-a-glance description of student suitability. Applications can be sorted by their keyword alignment immediately, allowing professors to prioritize the most relevant candidates. By moving beyond manual review and adopting NLP-driven matching, this project reduces the administrative friction in undergraduate research. Future iterations of this work aim to incorporate Large Language Models (LLMs) such as Gemini or GPT-4 to provide nuanced summaries of student applications and support multilingual sentiment analysis for interdisciplinary projects.

Traditional baby monitoring devices serve as a cautionary tool for parents, guardians, and caregivers concerned about the safety and security of their child; however, they require constant and scrutinous supervision. With ever-so-persistent distractions and responsibilities, as well as the inevitable need for rest, humans do not have the luxury of around-the-clock surveillance. This preliminary study introduces the use of large language models (LLMs) to develop an artificial intelligence (AI)-enabled baby monitoring system that identifies when a child is engaging in unsafe behavior requiring adult attention. Forty-five use-case videos of children engaging in either dangerous or age-appropriate activity were fed into two video understanding LLMs: Qwen2-VL and Video-LLaVa. A sliding window approach was employed to simulate live video streaming. Both models were prompted to process the video input and return the first instance of danger detected. The outcomes were compared against human perception to identify maximal detection accuracy. Qwen’s accuracy matched human perception on 31/45 videos by the second and 29/45 videos by the frame, yielding a 67% average accuracy rate. LLaVa’s accuracy matched human perception on 23/45 videos by the second and frame, yielding a 51% accuracy rate. Results demonstrate that the use of AI in child danger detection is highly feasible, and Qwen2-VL presented itself as the superior model for this task. Future work aims to supply the model with context from previous windows, perform fine-tuning, and collect data from additional observers to further refine accuracy and precision in preparation for deployment to a fully operational device.

Object detection models are widely used in technologies such as surveillance, robotics, and autonomous driving. This literature review explores how models like YOLOv11 have shaped the field while also remaining vulnerable to adversarial attacks. These attacks, including adversarial patches, can cause models to misinterpret images in dangerous ways. By reviewing current research, this paper highlights how these vulnerabilities work, why they matter, and why further study is needed to improve the safety and reliability of modern object detection. Object detection models, particularly the "You Only Look Once" (YOLO) series, have seen rapid architectural evolution from YOLOv1 to the current YOLOv11. While deep learning models are traditionally considered vulnerable to adversarial attacks, this study identifies a significant shift in the effectiveness of these attacks within modern training frameworks. The rapid integration of object detection models into these safety critical domains has made the security of these systems a paramount concern. While previous iterations of the “You Only Look Once” (YOLO) architecture have been extensively studied, the transition to the anchor-free model YOLOv11 presents a new landscape for adversarial robustness. This study performed a comprehensive analysis of established adversarial patch methodologies: the Dynamic Adversarial Patch (DAP), which utilized Creases Transformation (CT) blocks to account for movement, and the Remote Adversarial Patch (IPatch) which seeks to manipulate model semantics from a distance. Based on the goal of securing autonomous vehicle environments, this study implemented a custom replication of the IPatch methodology, adapting it from images segmentation to object detection. The experimental framework utilized the BDD100K dataset and employed Expectation over Transformation (EoT) to ensure the patch remained effective across various angles and distances. Despite an optimization process spanning 1,000 epochs using the Adamax optimizer, the replication failed to achieve the intended adversarial suppression or false-positive results. A significant outcome of this study arose during subsequent robustness testing, where I evaluated YOLOv11's response to various digital perturbations. These experiments revealed that the model consistently and correctly identified the adversarial patches themselves as distinct objects (such as a "teddy bear" or "person"), effectively neutralizing the attack's stealth. Building directly upon these findings, our future research will pivot to achieving adversarial stealth. Our upcoming work will explore the development of patches designed to be semantically hidden from the detection head by incorporating similarity objectives that blend the patch into the environmental background, testing these stealthier boundaries through physical experiments to improve the security of high-stakes technologies.

For my final class demonstration I created an advertisement using a key frame of a video of me and then altered the image to create a personal and emotional advertisement to show the potential of AI generated personalized advertisements. This project explores the conceptual design of an AI-powered sports advertisement prototype that integrates generative visual techniques and adaptive personalization ideas to examine how artificial intelligence may influence future sports marketing communication. The motivating problem is the growing need for advertisements that resonate with diverse audiences while maintaining creative originality and ethical transparency. Sports marketing provides relevant context because visual engagement, identity, and performance symbolism strongly influence audience perception and brand interaction. The project materials consist of conceptual advertisement imagery and illustrative video examples presented in the slides, rather than structured quantitative datasets. Inputs therefore include visual design elements, inspirational themes centered on human potential and athletic progression, and exploring demonstrations of AI-assisted image and video transformation. Methods are limited to generative and AI-enhanced visual design concepts shown in the presentation, including frame extraction ideas, creative recomposition, and hypothetical personalization workflows. The system workflow emphasizes ideation, visual generation, transformation, and audience-targeting concepts rather than algorithmic optimization or performance evaluation. Outcomes are qualitative and prototype-oriented, focusing on narrative coherence, visual symbolism, and perceived adaptability of advertisements rather than numerical metrics. Ethical and feasibility considerations are central to the project discussion, including concerns about privacy invasion, financial and energy costs of large-scale AI generation, and the risk of hallucinated or inaccurate outputs. These factors highlight the importance of transparency and responsible deployment in advertising analytics. Future work would involve structured data collection and research into audience engagement and retention when viewing a personalized ad generated with AI. Also, I would research how measurable evaluation of engagement outcomes between a variety of viewers in different demographics. I would also explore how personalization strategies improve engagement and retention for Ads of varying products. Additional development would also examine computational efficiency, bias mitigation, and clearer frameworks for in order to keep AI use in sports advertising ethical.

Authors: Santiago Soto and Dr. Patrick Lucey Affiliation: Morehouse College and Stats Perform Email: santiago.soto@morehouse.edu Problem Statement: This ongoing research develops a predictive framework to analyze five-man lineup efficiency and identify player combinations that maximize Net Rating and win probability. By prioritizing collective unit performance over isolated individual metrics, the study provides actionable data for algorithmic coaching strategies and roster optimization. Data Description: The analysis leverages a tabular dataset of play-by-play basketball data featuring 9,437 unique lineups across 30 teams. To ensure statistical significance and minimize noise, a threshold of 20 sequences (possessions) per lineup was applied. Key features include Offensive/Defensive Ratings, Rebounding Percentage, and Steal Rate. Methods and Analytical Approach • Feature Engineering: Performance is normalized per 100 sequences to calculate Net Rating (Net = Off - Def). • Weighted Aggregation: Team strength is computed through a weighted average of lineup metrics based on frequency of use. • Predictive Simulation Engine: An interactive engine was developed using scaling factors to ensure projections align with league-wide targets. • Correlation Analysis: A linear regression model assessed the relationship between net score and win percentage. Key Findings or Outcomes • Statistical Correlation: Regression analysis confirmed a near-perfect correlation (r = 0.96) between a team’s net score and its winning percentage. • Efficiency Tiers: The model successfully classifies units into Above/Below Average Efficiency based on league-wide medians. • Defensive Edge: "Good" lineups are distinguished by their ability to consistently limit opponent scoring, whereas weaker teams allow more points despite higher offensive outputs. • Win Probability: The simulation engine generates win probabilities and projected final scores using weighted averages. Impact, Implications, and Future Work: This research develops a framework for optimizing lineups against opponent strength using historical data, though it cannot yet incorporate real-time variables like injuries. Future work will expand the simulation with more granular features and multi-season data to assess the long-term stability of high-efficiency lineup styles.

Golf is a sport associated with technical precision and long-term participation. However, access, instruction costs, and social barriers can limit entry for many new players. Perfect Path explores how computer vision and pose estimation can support individualized swing analysis while lowering barriers through accessible, mobile-based technology. We frame the problem as both a biomechanics challenge, helping users understand body positioning and rotational mechanics, and a design challenge, creating tools that make golf instruction more approachable and data-informed. Our system uses user-provided videos as its primary input. Through a structured video-to-image workflow, users upload their footage, select a specific key frame, and apply pose detection to extract body lines and joint positions. Using Python and OpenCV, the pipeline estimates joint locations, evaluates angles and rotation, and exports a processed skeletal overlay for visualization and comparison. This transforms raw golf footage into interpretable visual feedback focused on alignment, posture, and swing path. The prototype demonstrates a complete end-to-end workflow. First is video upload, second comes frame selection, third is automated pose detection, and lastly, an exportable visualization is provided to the user. Importantly, the system does not employ predictive modeling to forecast performance outcomes such as hot distance or scoring improvement. Instead, it prioritizes transparent, interpretable visual analytics that allow users to directly observe their mechanics. By avoiding predictions, the tool emphasizes general skill development through self-assessment rather than guiding users to improve on a specific course without raising their overall skill level. From an inclusion perspective, the system is designed to reduce intimidation often experienced by beginners in traditional golf settings. Instruction can be costly, socially hierarchical, and reliant on in-person correction. By enabling private, self-paced analysis on a personal device, users can identify and correct mechanical errors without the pressure of public critique. This approach may help create a more welcoming entry point for individuals who feel underrepresented in golf spaces. Future development for this project includes a real-time mobile application capable of live swing capture, immediate pose-based feedback, and side-by-side comparison with a built-in reference skeleton. These enhancements would extend usability while maintaining a clear, interpretable analytics framework.

We propose a unified, lightweight, and reproducible set of models for longitudinal valence–arousal (VA) modeling in a corpus of essays written over time by U.S. service-industry workers. Using only the official SemEval 2026 Task 2 data, we enforce user-disjoint splits to prevent leakage and ensure comparable evaluation. We decompose VA modeling into three complementary prediction views: (i) per-essay VA state estimation from text, (ii) short-horizon user-level VA change forecasting from recent history, and (iii) longer-horizon disposition-change prediction from aggregated user histories. For essay-level state estimation, we fine-tune a DistilBERT encoder with a lightweight regression head. For short-horizon forecasting, we pair ModernBERT-based text representations with trajectory-derived numeric features and blend a simple previous-delta baseline with a GRU sequence regressor over recent embeddings. For longer-horizon disposition modeling, we pool DeBERTa-based user-history embeddings, augment them with normalized summary features, and apply a compact MLP regressor. On the official evaluation, we obtain Subtask 1 (essay-level state) composite Pearson r = 0.665 (valence), 0.468 (arousal) (official baseline: 0.557, 0.299); Subtask 2A (short-horizon change) Pearson r = 0.597 (valence), 0.413 (arousal) (official baseline: 0.615, 0.670); and Subtask 2B (disposition change) Pearson r = 0.046 (valence), 0.348 (arousal) (official baseline: 0.434, 0.584). Across all settings, we prioritize strict split control and transparent inference pipelines to make results easy to reproduce and extend, providing a reliable foundation for future work on longitudinal emotion dynamics.

Batching multiple parts onto a single fused-filament fabrication build plate improves throughput, but existing packing algorithms optimize for geometric density rather than print time. We introduce a slicing-aware cost model that estimates print time from part geometry and placement without invoking toolpath generation, achieving strong correlation with slicer-reported times (Pearson r = 0.90, Spearman ρ = 0.96). Evaluating packing algorithms on synthetic production builds, we find that greedy polygon-based packing matches or exceeds Large Neighborhood Search at three orders of magnitude lower compute cost — a negative result we attribute to high initial packing density leaving minimal room for iterative refinement. Against PrusaSlicer’s default auto-arrange, our method achieves 5.7% throughput improvement (95% CI [4.7%, 7.3%], N = 329), with median savings of 19.5 minutes per build.

During natural disasters, access to timely and reliable information is critical. However, many emergency communication systems depend on continuous internet connectivity or one-way broadcast alerts that can become inaccessible during infrastructure disruptions. This work addresses the need for accessible, conversational access to structured disaster information across multiple delivery channels. We developed a dual-mode conversational chatbot that integrates large language model–driven dialogue with authoritative, structured data services. A GPT-based model interprets natural language queries and routes them to six disaster-relevant services via external APIs, including weather conditions, air quality, emergency alerts, road conditions, shelter availability, and geolocation. A local persistence layer supports session management and system coordination. The system provides both web-based and SMS-based interfaces, enabling access across different connectivity contexts. The current prototype grounds all responses in verified service outputs, incorporates location-awareness and emergency-detection workflows, and supports multi-intent query handling. Ongoing work focuses on improving system robustness, expanding fault tolerance, and preparing the platform for scalable deployment.

A number of studies have explored what makes jokes funny. Conversely, only a few have actually tackled generating them, mostly leaving humor-generation relatively unexplored. The SemEval MWAHAHA Challenge tasks participants with generating jokes under different constraints with the aim of pushing models beyond memorization towards genuine joke creation. In Subtask A, inputs were either keyword pairs or news headlines, and jokes had to incorporate both keywords or draw from the given headline. For headlines, we prompted the model to write reaction-style tweets, which produced more natural humor, while for the keyword pairs, we had the model adopt Dave Chappelle's comedic persona to create observational jokes about some everyday situations and disappointments. We experimented primarily with open-source models (Llama and Qwen) and ended up using Llama for our final submission. In our preliminary results, we found that persona-based prompting consistently outperformed generic prompting approaches. The Chappelle-style observational jokes for keyword pairs also seemed to elicit more reactions than the standard outputs, and the tweet-format jokes for headlines felt more natural and appropriate for the given context. We also observed that models struggled a lot more with keyword pairs than with headlines, most likely because combining two unrelated words into coherent humor may require a bit more creative reasoning. Our findings show that tailoring prompting methods based on input type, rather than applying a singular approach, decently improves humor generation. Our work also shows the value of using personas when guiding models toward more diverse and naturally funny jokes.

Early detection in Alzheimer’s remains challenging, as current diagnostic methods rely on clinical expertise, and cognitive testing, which can be costly and inaccessible. Because cognitive decline is reflected in speech through changes in vocabulary, readability, pauses, articulation, and prosody, spoken language offers a scalable, low-cost, and non-invasive signal for screening and monitoring. Prior work demonstrated good performance in predicting Mini-Mental State Examination (MMSE) scores using large acoustic-linguistic feature sets, with optimized LightGBM and ensemble models significantly reducing RMSE and achieving strong Alzheimer’s detection accuracy. In this work, we extend beyond text-focused modeling toward a multilingual, audio-driven model for dementia detection and severity assessment. We train Audio Spectrogram Transformer (AST) models across ADReSS-style datasets to establish robust audio-only baselines, apply multilingual automatic speech recognition to obtain transcripts, and develop multimodal fusion models that integrate acoustic embeddings with linguistic features. This work lays the foundation for scalable and multilingual speech-based tools that can support early dementia detection. By using both audio and linguistic signals, it moves towards a more accessible approach to identifying cognitive decline.

This project addresses aircraft exterior inspections performed between flights to identify cracks, dents, corrosion, and paint erosion. Today, these walk-through inspections require significant manpower, can vary by inspector, and may expose technicians to safety risks when accessing elevated or hard-to-reach areas. A.N.T.S. (Autonomous, Navigation, Technician, Swarm) proposes an operator-supervised autonomous drone inspection system to improve repeatability, documentation quality, and inspection efficiency while maintaining technicians' full decision-making authority. Our approach uses stable, close-range quadcopter flights around a stationary aircraft to capture high-resolution imagery (with optional depth sensing). Collected data is processed on resource-constrained edge hardware using an embedded computer-vision pipeline (YOLOv8n) to detect defect candidates and automatically generate maintenance-ready inspection reports. To ensure flight safety and real-world reliability, the system includes software-controlled inference throttling to manage power and thermal limits during inspection. So far, we have defined the inspection concept of operations as a decision-support workflow, selected the baseline perception model (YOLOv8n) and edge-inference strategy, and outlined validation needs for close-proximity navigation near reflective surfaces, varied lighting, and different materials. Next steps focus on assembling test imagery, integrating sensors, and running controlled flight and mock-panel defect trials.

Abductive event reasoning in the wild requires selecting plausible explanations for an event from noisy, partially relevant multi document context. We present an evidence guided abductive scoring pipeline for SemEval 2026 Task 12 that separates evidence selection from explanation scoring. For each topic, we chunk documents and retrieve option conditioned evidence using dense embeddings, then apply a cross encoder reranker to form compact evidence packs per option. A constrained large language model scorer evaluates each option using only its evidence pack and outputs structured signals capturing evidence support, explanatory directness, and contradiction. We then apply deterministic decision rules to produce single or multi label predictions, including robust handling of none of the above style options through semantic detection rather than reliance on option position. This modular design reduces distraction from irrelevant documents, improves comparability across options, and enables controlled calibration for multi answer outputs. Our approach demonstrates that retrieval focused evidence compression combined with disciplined, signal based scoring can effectively support abductive reasoning without explicit knowledge graphs or end to end prompting over full document context.

Self-supervised speech learning extracts transferable representations from unlabeled audio, which is essential for multilingual settings where transcripts and reliable language annotations are limited. We present a masked spectrogram modeling (MSM) system for the Unsupervised Speech in the Wild Challenge (UPS dataset), Open Filtering sub-track, with an emphasis on data selection strategies that reduce language imbalance during training. The UPS audio pool is large and highly multilingual, but the distribution is skewed toward high-resource languages. To prevent training from being dominated by a small number of languages, we enforce manifest-level balancing for the non-English portion by grouping examples by language and applying a per-language quota (cap) before filling a fixed multilingual budget. In our current setup, no single non-English language contributes more than 4,000 examples to the multilingual manifest, ensuring consistent representation of lower-resource languages throughout training. We train a BiMamba2-based MSM model on log-mel spectrogram segments (10-second chunks), reconstructing masked time-frequency regions to learn phonetic and speaker structure without supervision. The current training phase uses 200 hours of filtered audio (100 hours English, 100 hours non-English) aligned with the challenge’s Few-Shot ASR, Zero-Shot Language ID, and Speaker Clustering objectives. Preliminary results show stable optimization, with validation loss decreasing from approximately 12.0 at early checkpoints (1k–2k steps) to 4.68 at 145k steps, indicating substantially improved reconstruction under multilingual balancing. We plan to scale to larger filtered subsets and evaluate downstream transfer on the challenge metrics.

A Statistical Fairness Evaluation Using Error and Semantic Similarity Metrics Kennedy Gregg HCAI Institute, Howard University, kennedy.gregg@bison.howard.edu Saurav Keshari Aryal HCAI Institute, Howard University, saurav.aryal@howard.edu Gloria Washington HCAI Institute, Howard University, gloria.washington@Howard.edu This study evaluates eight commercial automatic speech recognition (ASR) systems Google Cloud Speech, Microsoft Azure AI Speech, IBM Watson Speech, Deepgram, Amazon Transcribe, OpenAI Speech, AssemblyAI, and Speechmatics across three speech conditions: clean Common Voice (CV), African American Vernacular English (AAVE), and noisy CV. Outputs were compared to human transcripts using word error rate (WER), character error rate (CER), Levenshtein distance, and semantic similarity metrics (BERT cosine similarity, Euclidean distance, Jaccard similarity). Two-sample t-tests (Student’s or Welch’s, selected via Levene’s test) with Benjamini Hochberg correction assessed performance differences. All vendors showed statistically significant disparities across all metrics (adjusted p ≈ 0 to < 10⁻⁵⁸). AAVE speech produced higher error rates and lower semantic similarity than noisy CV, indicating reduced transcription accuracy linked to dialectal variation rather than noise. These findings highlight the need for dialect-aware benchmarking to ensure equitable ASR performance. REFERENCES [1] Manu Edavakandam, “From Audio to Words: A Python Guide to Measuring Transcription Accuracy,”Medium,2023. https://medium.com/@manuedavakandam/from-audio-to-words-a-python-guide-to-measuring-transcription-accurracy-f9dd9e70651f

The automatic recognition of multilingual medieval manuscripts remains a challenging task due to the wide variety of languages and writing styles, making it difficult to achieve consistent analysis. Historical documents exhibit grammatical irregularities such as inconsistent spelling and nonstandard grammar, degraded manuscript conditions that have compromised legibility over time, multiple languages written in diverse alphabets, and paleographic complexities, thus rendering conventional OCR(Optical Character Recognition) systems ineffective. Using the CMMHWR26 dataset as a source of multilingual medieval manuscript data, we acquire and preprocess manuscript images and their corresponding transcriptions to ensure proper normalization and formatting. Our method focuses on training open-weight multimodal language models that integrate visual and textual data to improve handwritten text recognition performance. By leveraging both image-based features and language modeling capabilities, the system aims to achieve robustness and generalization across diverse medieval scripts and languages, including related and linguistically different language families. This work contributes to the development of multilingual historical handwritten text recognition by exploring the use of open-weight multimodal language models in medieval manuscript settings, supporting the creation of more robust systems capable of handling real-world retrodigitization challenges and improving large-scale digitization efforts in the digital humanities.

Historical newspaper archives are critical for preserving Black press history, yet many digitized newspapers exist as fragmented image strips that are difficult to access and study. This fragmentation hinders researchers' ability to efficiently navigate and analyze these invaluable historical documents. Working with the Black Press Archive at Howard University, this project addresses this challenge by developing an automated pipeline to reconstruct complete newspaper pages from image strips. Using OpenCV computer vision library, the approach vertically concatenates sequential newspaper strips into complete documents. The system then employs brightness analysis to detect dark horizontal gaps between pages, automatically identifying page boundaries and splitting the stitched image into individual pages. This eliminates manual segmentation work that would otherwise be time-intensive and error-prone. The implemented pipeline successfully processes multiple newspaper strips, generates intermediate stitched outputs for quality verification, and produces organized individual page files. This automation significantly improves the accessibility and usability of the Black Press Archive's digitized newspaper collection for historical research and preservation.

The aim of this project is to assess the ability of a generative AI to produce reasonable and believable images based on the same original images at varying scales to simulate differences in distance from the object. The objects used in this project were varying types of ships. The generated images were assessed in three different ways, all on a scale of high medium and low to measure levels of quality. Firstly, background quality which refers to the realism and consistency of the generated image. Subject dimensions which refers to how well the generated image preserves the scale and zoom of the ship in the original image. And subject integrity, which referred to the features of the ship and how consistent and believable they were. Additionally, we noted whether the original image contained the ship as a whole without any ‘cut off’ parts, and if so, being labelled as “Incomplete source”. And lastly was “Artifact present” which stated if the generated image had had hallucinations or objects added that were not present in the original image. Space was also made for comments to be made on the generated images.

This paper presents a robust framework for Top-View RGB-D Person Re-Identification (TVRID), addressing the specific challenges of the ICPR 2026 competition. Our approach integrates three specialized tracks: an RGB track utilizing part-based attention and ResNet backbones (e.g., ViT, Swin) to improve robustness to occlusion; a Depth track focusing on body-shape cues learned from 1-channel depth with attention and metric losses; and a Cross-Modal track employing dual-stream fusion with optional cross-attention and cross-modal metric losses. By combining identity (ArcFace) and metric (batch-hard triplet, center) losses within a PyTorch Lightning framework, our method achieves strong discriminability across same-camera and cross-passage scenarios.

Adverse Drug Events (ADEs) remain a leading cause of preventable morbidity and healthcare expenditure, yet their identification from electronic health records (EHRs) remains challenging due to the unstructured and context-dependent nature of clinical narratives. While annotated corpora such as the n2c2 dataset support benchmark evaluation, their limited scale and annotation cost constrain generalizability. We present a knowledge-grounded ADE detection framework that integrates clinical note extraction with structured pharmacovigilance knowledge from SIDER. Our approach first applies large language model (LLM)-based entity and relation extraction to identify drug–event mentions within clinical text. We then incorporate a context-engineering layer that cross-references extracted drug entities against curated side-effect profiles in SIDER, enabling structured validation of candidate ADE associations. This integration reduces spurious drug–event pairings while preserving sensitivity to plausible associations, thereby improving precision without sacrificing recall. By combining biomedical knowledge bases with LLM-driven extraction, this work demonstrates a scalable and biologically informed strategy for pharmacovigilance from real-world clinical data. The proposed framework contributes to translational health informatics by bridging curated molecular–drug knowledge and patient-level clinical evidence.

This work presents the development of a modular bipedal humanoid platform focused on tightly integrating onboard artificial intelligence with multimodal sensor perception for data-driven autonomy. Built upon the Berkeley Humanoid Lite architecture, the system emphasizes the unification of mechanical design, embedded compute, and learning-based control within a fully self-contained humanoid framework. We design and implement a multimodal sensor array incorporating LiDAR, RGB and depth cameras, and microphone inputs, enabling rich environmental perception across spatial and acoustic domains. Sensor data are synchronized and fused through a ROS 2-based middleware pipeline to produce real-time state estimation, semantic scene understanding, and context-aware environmental representations. These perception outputs inform higher-level decision-making modules, allowing the robot to adapt its locomotion and manipulation strategies based on sensed environmental conditions rather than pre-scripted behaviors. Locomotion and task policies are trained in GPU-accelerated simulation and deployed to onboard embedded AI hardware for low-latency inference. This architecture enables closed-loop, perception-driven autonomy without reliance on external compute infrastructure. We evaluate system performance in terms of perception latency, decision consistency, and behavioral adaptability in dynamic indoor environments. The resulting platform demonstrates a scalable approach to embedding AI directly within humanoid robotic systems, enabling real-time, sensor-informed decision-making and advancing the integration of learning-based intelligence in embodied agents.

Large language models (LLMs) are increasingly used in high-stakes decision support to summarize situations, propose actions, and communicate rationale. While these systems often produce fluent and plausible responses, such outputs can obscure uncertainty, weaken grounding, and invite over-reliance by human decision-makers. We present Project Comprehension, a forensic evaluation framework that examines LLM outputs as decision-relevant artifacts rather than isolated answers. The framework combines operationally grounded scenarios with human-centered annotation to assess plausibility, uncertainty signaling, grounding transparency, comprehension support, and actionability. Across empirical testing, we find that surface-level response quality is only weakly predictive of grounding transparency: a non-trivial subset of responses appear clear and actionable while providing limited justification or source signaling. These patterns highlight an interpretive risk that is not captured by accuracy-focused evaluation alone. We discuss how forensic evaluation can support trust calibration, assurance practices, and the design of language-enabled decision support systems that better align with human judgment in high-stakes contexts.

We present a lightweight, reproducible submission for the TidyVoice 2026 cross-lingual speaker verification challenge implemented in the WeSpeaker toolkit under single-GPU Google Colab constraints. Our primary system, S1, uses the official SimAM-ResNet34 checkpoint pretrained on VoxBlink2 and VoxCeleb2 and fine-tuned on TidyVoiceX, which we further fine-tune for five epochs with large-margin classification. In parallel, we implement a secondary self-supervised system, S2, using a frozen WavLM-Base frontend with a compact statistics pooling speaker head, trained for four epochs. Both systems use standard speech augmentation during training with MUSAN noise and RIRS reverberation, while inference uses clean embeddings and cosine scoring. To combine systems, we perform score-level fusion calibrated on a labeled Tune-S development split. We z-normalize each system’s Tune-S scores using their mean and standard deviation, grid-search a convex fusion weight alpha in the range 0 to 1 with step 0.01 to minimize EER, and apply the frozen normalization and alpha to fuse Eval-A (Task 1) and Eval-U (Task 2) score files for submission. On Tune-S, S1 substantially outperforms S2, so the selected fusion weight is alpha equals 1.0.

We present a hybrid ensemble approach for SemEval 2026 Task 5 (AmbiStory), which requires predicting the plausibility of homonym senses in literary narratives on a 1–5 ordinal scale. The task challenges systems to handle nuanced contextual ambiguity in creative writing, where traditional Word Sense Disambiguation methods often fail. Our methodology combines the complementary strengths of large language models and fine-tuned transformers through a multi-stage pipeline. First, we employ diverse LLM prompting strategies including few-shot learning, contrastive reasoning, and chain-of-thought prompts across multiple model providers and temperature settings, with optional retrieval-augmented generation to surface relevant training examples. Second, we fine-tune RoBERTa-large for ordinal regression using contextualized example sentences, deploying a multi-seed ensemble to reduce prediction variance. Third, we apply ensemble voting with median score aggregation across LLM outputs to improve prediction robustness. Finally, we integrate LLM ensemble and RoBERTa predictions through hybrid combination methods: weighted averaging, confidence-based weighting, and a calibrated meta-learner that learns optimal blending strategies from development data. Our approach achieves 80% exact-label accuracy and 0.74 Spearman correlation on the development set, substantially outperforming baseline methods. The results demonstrate that structured reasoning from LLMs, when combined with learned transformer representations, ensemble aggregation, and proper calibration, effectively captures the ordinal nature of sense plausibility in complex literary contexts. Ablation studies reveal that ensemble voting reduces individual model variance while the meta-learner successfully exploits complementary error patterns between LLM and RoBERTa components. Future work will explore refined retrieval mechanisms and ensemble optimization strategies.

This paper investigates whether personality type influences a user's ability to bypass AI safety guardrails. We construct psychologically grounded personas using Gaussian Mixture Models (GMM) applied to the Big Five personality dimensions (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism), drawing from the Statistical "Which Character" Personality Quiz (SWCPQ) dataset of 2,125 fictional characters rated on 500 personality traits. Characters are clustered per Big Five dimension independently to avoid high-dimensionality issues, with optimal K determined via the elbow method, silhouette scoring, AIC, and BIC, and validated through bootstrap resampling. Each cluster's most representative character becomes the blueprint for a persona, yielding P unique AI personas encoded as LLM system prompts. These personas systematically probe AI models with restricted prompts, measuring compliance rates and manipulation strategies across personality types. Preliminary clustering with K=7 produced balanced groupings with exemplars from works including 10 Things I Hate About You and The Great Gatsby, though AIC and BIC disagreement motivated the shift to Big Five-aligned clustering. Next steps include completing the Big Five trait mapping, constructing the full persona matrix, and running probing experiments to determine which personality profiles most effectively bypass AI safety guardrails.

African languages represent one of the world’s most linguistically diverse regions, yet they remain critically under-resourced in Natural Language Processing, limiting the development of equitable and effective language technologies. Sentiment analysis for these languages is particularly constrained by scarce labeled data, limited representation in pretrained models, and heavy reliance on translation-based pipelines that introduce cultural and semantic distortion, especially in code-switched contexts. This work extends the AfriSenti benchmark with sentiment data from 38 additional African languages and examines how linguistic relatedness, captured through language family structure, can be leveraged to improve multilingual sentiment modeling. We evaluate two complementary approaches: extended task- adaptive pretraining on a large and heterogeneous multilingual dataset, and direct language and family conditioning through explicit input-level metadata. Experiments conducted at the overall, language family, and individual language levels show that task-adaptive pretraining provides strong gains in large and noisy multilingual settings, while direct language and family conditioning is most effective on smaller and cleaner benchmarks such as AfriSenti. Together, these results provide empirical evidence for the language family effect, demonstrating that both implicit and explicit modeling of genealogical relationships improves robustness and generalization for African languages that are underrepresented or absent in standard pretrained models. Our findings highlight the value of linguistically grounded and data- efficient approaches for building more inclusive and sustainable NLP systems for African languages.

Technology revolving around remote sensing is being utilized to effectively identify crop diseases in agricultural areas. Crop diseases such as fungal, bacterial, and other infections impact agricultural productivity, which can reduce plant growth and nutritional value in everyday food. Throughout past research, advances in digital imaging have led researchers to developing methods for seeking potential in crop disease diagnosis through RGB imagery, hyperspectral data, and multispectral analysis. However, the vast number of spectral bands and relationships in remote sensory can pose a challenge to effectively select features and extract data. That is where we come in as we are currently in the process of creating deep learning algorithms to negate those situations. The approach I have come to take is through developing a complex file discovery and reading pipeline to accurately read agricultural data based on the image and file type. My script for containing specific files such as TIF, PNG, and CSV finds the file, analyzes the image, and then extracts the data to give statistics and previews. I have made this happen through utilizing OpenCV, pandas, rasterio, and pathlib to accomplish this.

While large language models increasingly serve as chat-based assistants, grounding their responses in retrieved evidence across multi-turn conversations remains a significant challenge, particularly when questions reference earlier turns, when the system must recognize unanswerable queries rather than hallucinate, and when relevant passages shift as the conversation evolves. We address these challenges on the MTRAG benchmark across four domain-specific corpora: ClapNQ (Wikipedia), Cloud (technical documentation), FiQA (financial), and Govt (government web pages). Our system employs a hybrid retrieve-then-rerank architecture. Queries are first augmented through LLM-driven query rewriting, breaking down entities and query itself, and generating hypothetical embeddings (HyDE) for semantic matching. Results from dense vector search and lexical matching are then fused via Reciprocal Rank Fusion and reranked though cross-encoder. Llama-3.3-70B-Instruct then generates responses based strictly on the most relevant text passages. The system achieves an nDCG@5 of 0.4098 on passage retrieval, a harmonic mean of 0.7462 on reference-grounded generation, and 0.5796 on end-to-end RAG.

The digital divide has historically separated individuals with reliable access to technology from those without it. As Artificial Intelligence (AI) becomes increasingly integrated into education, healthcare, business, cybersecurity, and workforce development, this divide is expanding into what can be described as an “AI divide.” Individuals in lower socioeconomic communities often lack the technological infrastructure, high-speed broadband access, digital literacy skills, and institutional support necessary to effectively utilize AI tools. In contrast, higher-income communities benefit from stronger internet connectivity, updated hardware, and educational programs that introduce students to emerging technologies. AI systems depend on fast data processing, cloud computing resources, and stable broadband connections to function efficiently. However, high-speed internet services are disproportionately available in urban and suburban areas, while many rural and underserved communities experience slower and less reliable connectivity. This infrastructure gap directly limits access to AI-powered applications used for learning, job training, and economic advancement. Educational disparities further intensify the problem. Underserved schools often lack funding for advanced computer science courses, AI literacy programs, and updated technological equipment. Without early exposure to AI concepts and skills, students may be unprepared for a workforce increasingly shaped by automation and machine learning. As AI expands across industries and begins to automate certain job functions, individuals without access or training risk long-term economic disadvantage. This study examines how infrastructure inequality, limited AI education, and insufficient institutional support may widen socioeconomic gaps in the era of artificial intelligence. It also explores policy and educational strategies designed to promote equitable access to AI technologies and workforce opportunities.

The goal of the VISTAC challenge is to improve visual object tracking under adverse weather conditions. While advanced tracking technologies perform well in controlled and well-lit environments, their performance decreases significantly in challenging conditions such as haze and rain. This limitation is critical because real-world applications, like traffic monitoring systems and autonomous vehicles, must operate reliably in unpredictable environmental settings. The challenge aims to develop robust tracking algorithms capable of maintaining accuracy and consistency in harsh environments. This research investigates the effectiveness of different image denoising and filtering techniques for improving object tracking performances in degraded visual conditions. Using annotated video data containing hazy and rainy scenes, multiple preprocessing methods are applied to enhance frame clarity before tracking. The impact of each denoising technique is evaluated based on visual quality, feature preservation, and tracking performance metrics such as Qualitative Precision (QP) and effective frames per second.

The transition from high school to college can be challenging for many students. College classes are more demanding, schedules are less structured, and students must learn to manage their time independently. Many students also feel nervous about meeting new people and adjusting to a new academic environment. To support this transition, many colleges offer pre-college bridge programs that prepare incoming students before their first semester begins. Bridge programs focus on building important skills such as time management, study habits, and critical thinking. In our bridge experience, Artificial Intelligence (AI) tools were also introduced as part of academic preparation. We explored how to use AI responsibly to assist with studying, brainstorming ideas, understanding assignments, and practicing problem-solving. Learning how to use AI as a support tool could help freshmen feel more prepared for college-level coursework and more confident in completing assignments. The program bridge provided opportunities to connect with other incoming freshmen and mentors, which could reduced anxiety and helped us build early friendships. In this study, we reflect on how the bridge program—along with guided AI use—improved my academic readiness, confidence, and sense of belonging during our first year. Overall, this experience showed that combining traditional bridge support with responsible AI integration can strengthen both academic success and student confidence during the transition to college.

Large language models (LLMs) continue to underperform for culturally diverse and linguistically underrepresented communities, limiting their applicability in multilingual and code-switched environments. This work introduces a culturally aware Mixture of Specialists (MoS) framework coordinated by a Model Control Protocol (MCP) server to dynamically route user inputs to language- or region-specific models based on linguistic proximity, cultural relatedness, and data availability. When a dedicated specialist exists, it is used directly; otherwise, a hierarchical fallback strategy selects a linguistically related model, then a culturally proximate variant such as a West African English–tuned specialist, and finally a multilingual backbone augmented with lightweight regional adapters. As part of a multi-phase research program, this paper presents the first stage of the system, focusing on the routing architecture, cultural metadata extraction, and region-aware prompting components while specialist model training is ongoing. To support future specialization, we prepare parameter-efficient fine-tuning pipelines (LoRA and QLoRA) using openly licensed corpora rich in local context, including OSCAR, mC4, BigScience ROOTS, Tatoeba, African StoryBooks, and Global Voices, with thorough deduplication, filtering, and native-speaker validation. Evaluation on the BLEnD benchmark from SemEval 2026 Task 7 across 26 languages and 30 regions demonstrates that culturally grounded routing signals, regional metadata, and language-specific constraints yield substantial gains in contextual accuracy, robustness in low-resource settings, and cross-regional generalization. These Phase-1 results provide early empirical evidence that linguistic relatedness and cultural proximity can meaningfully enhance multilingual model performance even before full specialist integration. Overall, this work establishes a scalable foundation for developing globally adaptive and culturally grounded NLP systems.

Automatic speech recognition (ASR) systems are widely used in voice assistants, transcription services, and accessibility tools, yet prior research suggests they perform unevenly across dialects. This project investigates performance disparities in commercial ASR systems for African American Vernacular English (AAVE). We curated and transcribed over 200 hours of question–response style AAVE speech data and split it into training, validation, and test sets. Using an automated benchmarking pipeline, we evaluate systems including OpenAI Whisper, Amazon Transcribe, and Deepgram. Performance is measured using Word Error Rate (WER), with statistical analyses such as Welch’s t-test and Shapiro–Wilk tests applied to assess significance and distributional assumptions. Preliminary findings indicate that several commercial systems exhibit elevated WER on AAVE speech relative to reported general benchmarks. To address this gap, we are fine-tuning models on the curated dataset and observing reductions in WER, though this phase remains ongoing. By combining systematic benchmarking, statistical rigor, and dataset development, this work contributes toward more equitable and representative speech recognition technologies

Adversarial patch detection represents a critical frontier in cybersecurity defense. As artificial intelligence systems assume greater responsibility in safety-critical and security-sensitive applications, the ability to detect and neutralize adversarial attacks becomes paramount. As artificial intelligence systems become increasingly woven into the fabric of critical infrastructure—impacting areas such as autonomous vehicles, facial recognition technologies, medical diagnostics, and financial fraud detection—their vulnerability to adversarial patch attacks takes on a new level of significance, posing a considerable and escalating cybersecurity threat. Adversarial patches are intricately designed perturbations, whether physical objects or digital modifications, crafted with precision to deceive AI vision systems. These deceptive alterations can manipulate the system's perception and decision-making processes, resulting in erroneous classifications. Such manipulation can empower malicious actors to bypass established security measures, take control of autonomous operations, or elude detection mechanisms, potentially leading to catastrophic consequences. This research carefully examines the urgent necessity for adversarial patch detection as an essential component of a comprehensive defensive strategy within the cybersecurity landscape. It explores the increasing sophistication of current adversarial attack methodologies, which often exploit subtle vulnerabilities in AI algorithms with alarming effectiveness. Moreover, the study investigates the capabilities of emerging detection frameworks that aim to identify, analyze, and mitigate these sophisticated threats. By exploring the dynamic relationship between advancing adversarial tactics and the evolving defense mechanisms, this work seeks to illuminate strategies to bolster the resilience of AI systems against these insidious attacks, thereby enhancing the safety and reliability of critical infrastructure in a rapidly evolving digital landscape.

Deep learning-based object detection technologies, such as YOLOv5 and Faster R-CNN, are being increasingly applied in safety-critical areas, including self-driving cars, surveillance systems, and smart transportation infrastructure. Although these models show remarkable performance under standard conditions, they are vulnerable to physical adversarial patch attacks. These attacks involve the careful placement of specifically designed printed perturbations in a scene to provoke misclassification or to obscure objects. In contrast to digital attacks that take place in controlled settings, physical adversarial patches operate under real-world conditions, where variables such as changing light, distance, angle, and occlusion can greatly influence their effectiveness, making them a significant danger. This study explores detection-based defense mechanisms designed to identify physical adversarial patch attacks using object detectors. We assess various methods, including confidence-score analysis, monitoring of bounding-box instability, feature-distribution anomaly detection, and ensemble-based detection strategies. To simulate realistic deployment scenarios, researchers collected a controlled dataset of both clean and patched objects under various environmental conditions. The detection performance is evaluated using precision, recall, F1-score, mean Average Precision (mAP), and inference latency. Early investigation of research studies suggests that the use of ensemble detection, along with tracking confidence distributions, significantly improves the detection rates of adversarial patch attacks while keeping performance near real-time levels. This research focuses on identifying, rather than stopping, physical adversarial patch attacks using object detection-based defense strategies.

The PsyCoMark 2026 shared task emphasizes modeling the psycholinguistic structure underlying conspiracy belief expression rather than relying solely on topical cues. In this work, we address Subtask 2, which requires classifying Reddit comments as conspiracy-related or non-conspiracy-related across diverse domains. We implement a syntactically augmented transformer-based classifier using ModernBERT-base. To introduce lightweight structural information aligned with PsyCoMark’s theoretical framing, we extract Part-of-Speech (POS) tags using spaCy and concatenate the resulting syntactic sequence with the original comment text via a separator token. This approach allows the model to jointly encode lexical semantics and shallow grammatical structure without architectural modification. The model is fine-tuned for binary classification using cross-entropy loss, with early stopping applied to reduce overfitting. Preliminary experiments on the official development split yield a macro-averaged F1 score of 0.46. While performance remains modest, these results establish a functional baseline for structurally augmented classification and provide initial insight into the contribution of shallow syntactic signals for conspiracy detection in topic-diverse online discussions. This work contributes an empirically grounded starting point for exploring psycholinguistically informed transformer models within the PsyCoMark framework.

Advances in neural text-to-speech technologies have allowed for realistic voice cloning, however, determining how “natural” synthetic voices are remains a challenge. As voice cloning becomes more integrated into applications such as accessibility systems, entertainment platforms, virtual assistants, and more, the quality of synthetic speech becomes increasingly significant. Subtle differences can shape listeners' attitudes towards usability, reduce trust, and affect overall user experience. Therefore, identifying factors that contribute to human-like speech and establishing reliable methods to evaluate perceived naturalness is essential in advancing speech synthesis systems. This study investigates the relationship between training data quantity and perceived human-likeness in cloned voices using a neural voice cloning pipeline based on Coqui XTTS. An end-to-end voice cloning system was implemented, with voice models trained on speech samples of varying lengths. Synthetic speech outputs were generated from each model and evaluated through a qualitative listening study. The generated voice samples were assessed by three human evaluators: self, familiar, and unfamiliar listeners. Evaluators rated each sample on a five-point Likert scale for perceived naturalness. Ratings were compiled into a structured dataset for comparative analysis. Additionally, different approaches for evaluating speech naturalness were explored to inform the study design. Thus far, the end-to-end system has been implemented, models have been trained under varying data conditions, and listening evaluations have been conducted to support ongoing analysis. Together, these components establish a systematic framework for examining how training data quantity relates to perceived naturalness in synthetic speech, offering a foundation for further refinement in voice cloning applications.

Radiomics allows for the extraction of features from medical images for predictive modeling, but due to data privacy regulations, training models on multi-institutional datasets is not feasible. We propose a cross-silo federated learning framework that allows hospitals to jointly train radiomics models without sharing patient data. Our solution utilizes the Flower framework and the radMLBench benchmark (50+ radiomics datasets) to create a multi-hospital federated learning setting. Each hospital is given unique datasets, as seen in real-world settings. We handle feature heterogeneity by using intersection-based feature alignment, making the model compatible across hospitals. The framework utilizes a four-layer neural network architecture that is trained locally at each hospital. The central server combines the model parameters using Federated Averaging (FedAvg), weighted by the number of samples. We compare the federated learning method with a centralized approach that trains on aggregated data using accuracy as a metric. Although the centralized approach has the advantage of direct access to all the data, the federated approach is able to attain similar results while preserving the privacy of the data, thus proving that collaborative learning can come close to the centralized approach without breaching the privacy constraints. The framework is applied to brain-related disorders (gliomas and glioblastomas) using MRI images, thus illustrating how different institutions can work together to collaboratively enhance the performance of the model while being HIPAA compliant.

This paper describes the PolarNLP system submitted to SemEval-2026 Task 9, Subtask 1, which focuses on detecting political polarization in multilingual text. The task spans 22 typologically diverse languages and poses challenges related to domain shift, class imbalance, and cross-lingual generalization. We explored two modeling strategies: (i) a weakly supervised teacher–student approach that uses a large language model (LLM) to generate pseudolabels for training a multilingual classifier, and (ii) direct LLM-based inference augmented with language-agnostic stylistic features. While the teacher–student approach achieved reasonable in-distribution performance, it failed to generalize to the heldout test set, collapsing toward the majority class. Consequently, our final submission relies on direct LLM inference. We present a detailed analysis of both approaches, highlighting the limitations of weak supervision for polarization detection and the relative robustness of direct LLM reasoning in multilingual settings.