Campbellsville University

We address the problem of empirical privacy auditing for differentially private stochas- tic gradient descent (DP-SGD) under the hidden state threat model, where adversaries only observe the final model parameters. Our work introduces a gradient-crafting framework that enables tighter auditing by allowing adversaries to pre-specify worst-case gradient sequences without access to interme- diate training checkpoints. We demonstrate that when a data point is used at every optimization step, hiding intermediate models provides no privacy amplification beyond standard composition bounds. For less frequent data usage patterns, we identify regimes where privacy amplification may occur for non-convex problems, though the effect is weaker than in convex settings. Our findings clarify the actual privacy loss in practical DP-SGD deployments and provide foundational insights for improved privacy accounting in the hidden state model.

The convergence of artificial intelligence (AI) and sustainable finance represents a paradigm shift in how financial institutions approach environmental, social, and governance (ESG) analytics. This paper proposes a novel interdisciplinary framework, the Resource-Aware AI for Sustainable Finance (RAISE) framework, that optimizes resource management and cost efficiency while maintaining analytical precision for ESG investment decisions. The framework integrates machine learning techniques for ESG data analysis with resource-aware algorithms that dynamically allocate computational resources based on priority tasks. Our approach addresses the significant energy consumption and computational costs associated with AI implementations in finance while improving the accuracy and transparency of ESG scoring. Experimental results demonstrate that our framework reduces computational resource utilization by 32% and decreases energy consumption by 28% compared to conventional AI approaches for ESG analysis, while maintaining 96.3% of predictive accuracy for climate risk assessment. The research contributes to both sustainable finance and green AI literature by providing a scalable, efficient, and transparent methodology for AI-driven financial analysis that aligns computational efficiency with sustainability objectives.

In modern industrial systems, unplanned downtime leads to significant operational and financial losses. To address this, we propose a unified predictive maintenance pipeline that integrates generative anomaly injection, explainability, and reinforcement learning (RL) within a digital twin framework. Our system uses sensor data from real-world datasets (XJTU-SY, AI4I, and C-MAPSS) for Remaining Useful Life (RUL) prediction, fault simulation, and proactive maintenance policy generation. First, a GAN-based fault simulation module injects rare but realistic failure patterns using low discriminator confidence. Next, SHAP-based explainability is applied for feature attribution and drift monitoring. LSTM and Prophet models are employed to forecast RUL trends, which are fed into a Q-learning-based RL scheduler to optimize prescriptive actions like maintenance or load reduction. These actions are sent to a virtual command layer and closed-loop feedback is achieved via MQTT/OPC-UA to simulate interaction with the physical twin. We quantitatively evaluate the synthetic fault realism using t-SNE, PCA, and classifier generalization. The overall pipeline improves prediction accuracy by 12.6% and reduces failure downtime by 19.4% on average. SHAP analysis reveals interpretable patterns across synthetic and real failures. Our modular architecture demonstrates a self-optimizing, explainable, and simulation-augmented predictive maintenance system for Industry 4.0 environments. This work highlights the practical synergy of generative models, RL, and digital twins for scalable, real-time industrial reliability solutions.

Breast cancer continues to be a significant cause of mortality among women worldwide, underscoring the importance of accurate and early detection to enhance survival rates. This research explores the effectiveness of several supervised machine learning models in predicting breast cancer using a specialized dataset. Five algorithms were evaluated: Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), and Gradient Boosting. The dataset includes 569 samples with 30 continuous features derived from digitized images of fine needle aspiration (FNA) of breast tumors. The models were trained and assessed based on key metrics such as accuracy, precision, recall, and F1-score. Among the models, Random Forest and SVM delivered the best results, both achieving a 97.20% accuracy, which highlights the strength of ensemble and kernel-based methods in this application. Logistic Regression also performed commendably, reaching an accuracy of 95.10%, indicating that the feature space exhibits a high degree of linear separability. While Decision Tree and Gradient Boosting models demonstrated good performance, their results were slightly lower in comparison. These findings emphasize the capability of machine learning, particularly Random Forest and SVM, in supporting breast cancer diagnosis. Furthermore, the consistently high accuracy across models reflects the robustness of the features derived from FNA images for breast cancer classification.

An effective product strategy requires a systematic framework to prioritize features and match development with organizational objectives. The MoSCoW Prioritization framework provides a flexible, qualitative approach for classifying characteristics into Must Have, Should Have, Could Have, and Won’t Have categories, facilitating targeted decision-making. To assess its efficacy, five principal attributes—Stakeholder Alignment, Business Value, Customer Demand Level, Implementation Effort, and Strategic Relevance—were examined. Data supplied from productlift.dev facilitated a comparison of MoSCoW Prioritizations with the Analytic Hierarchy Process (AHP), Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), and the Weighted Scoring Model. Included materials were ranking templates, prioritization instruments, and result indicators. MoSCoW Prioritization regularly surpassed alternatives, exhibiting enhanced flexibility, interpretability, and strategic alignment. Quantitative study verified a 98% accuracy in prioritization throughout the implementation of roadmap scenarios. The findings validate the model’s potential to facilitate agile product planning and roadmap optimization, making it a pragmatic option for rapidly changing business contexts and significant product initiatives.

Cloud automation offers transformative potential for healthcare organizations facing increasing administrative burdens and patient experience challenges. By integrating Electronic Health Record systems with cloud platforms like Salesforce Health Cloud, providers can streamline operations across the entire patient journey. These technologies enable digital patient onboarding, intelligent appointment scheduling, and automated claims processing while reducing manual workloads. AI-powered engagement tools further enhance patient interactions through virtual assistants that provide 24/7 support. Despite implementation complexities involving legacy system integration and staff resistance, healthcare organizations can achieve significant operational efficiency and improved patient outcomes by following established best practices. As these technologies continue to mature, their strategic implementation promises substantial benefits for healthcare delivery in an increasingly digital ecosystem.

Intrusion Detection Systems (IDS) are essential in identifying and reporting potential network attacks. IDS are classified into Host-based IDS (HIDS), which monitor internal computer threats, and Network-based IDS (NIDS), which monitor network-level attacks. IDS can also function as anomaly-based, using Machine Learning (ML) and Deep Learning (DL) to recognize unfamiliar attack patterns, or as rule-based, which relies on historical data-driven rules. Notably, anomaly-based IDS can detect zero-day attacks. Therefore, this study proposes a Deep Neural Network (DNN) model for anomaly-based IDS, focusing on accurate and efficient attack detection and categorization using the CIC-IDS-2017 dataset. By optimizing dataset preprocessing, the DNN architecture aims to minimize computational demands while maintaining high accuracy. Comparative evaluation with other models demonstrates the proposed model's effectiveness in attack detection despite using a simpler, more lightweight architecture than those in other studies, where more complex, less efficient approaches are often employed.