2026 IEEE International Conference on AI and Data Analytics
(ICAD 2026)
June 11 – 12, 2026
Boston, Massachusetts
Room 115 abstracts and final papers
35: Synthetic Data for Training & Evaluation in Financial Fraud Detection: A SMOTE-Powered KAN-XGBoost Framework with SHAP Interpretability presented by RAJESH LINGAM (Independent Researcher)*
The fast growth of digital financial platforms has further exacerbated the challenge of identifying fraud in transactional data. Rare nonlinear fraudulent behaviors associated with highly imbalanced financial data cannot be captured by any of the existing machine learning models. The above-mentioned gap motivated us to propose KAN-XGBoost, a hybrid model incorporating the expressiveness of the Kolmogorov-Arnold Networks regarding feature transformation and the robustness of XGBoost associated with gradient-boosted classification. With the synthetic financial datasets for fraud detection-PaySim, along with a carefully designed preprocessing pipeline ranging from filtering fraud types and balancing features to tokenization and rebalancing using SMOTE, we train and evaluate our model under realistic conditions. According to experimental results, KAN-XGBoost outperforms its natural competitors—LR, RF, LightGBM, and vanilla XGBoost—by a significant margin, achieving an accuracy of 99.99%, precision of 1.00, recall of 1.00, F1-score of 99.99%, and an AUC of 0.999. Domain-consistent insights, such as the effect of dynamics in feature balance and transactional anomalies on predictions, are clearly conveyed through SHAP-based interpretability. This work improves the use of synthetic data and hybrid architecture in financial security systems by providing a highly accurate, interpretable, and scalable solution for fraud detection.
74: Hidden Stylistic Schools of Pre-Islamic Poetry presented by Salem Othman (Wentworth Institute of Technology)*; Youssef Qranfal (Wentworth Institute of Technology)
Pre-Islamic Arabic poetry is an early and sizable record of Arabic that survives in written form. In tribal society poets served as spokespeople and informal historians recording events in verse. The poems preserve values, social norms, and a distinct literary taste from that era. Yet the digital study of this corpus remains limited. This paper presents a computational exploration of 2,328 curated pre-Islamic Arabic poems from 260 poets, scraped from the Aldiwan repository. Four embedding models were evaluated, overall showing that an Arabic-specific sentence-transformer (Arabic-SBERT) shows the clearest poet separation among the models we tested. We further compared four pooling strategies for aggregating verse embeddings into poet-level representations, and we find that hierarchical verse-to-poem-to-poet averaging captures structure that simple concatenation tends to miss. Using UMAP dimensionality reduction and HDBSCAN density-based clustering, we recover twelve stylistic clusters and test their robustness using randomization (p < 0.01 for cluster count), parameter stability analysis, and comparison against TF-IDF, character n-gram, and topic model baselines. The results reveal stylistic diversity within this highly formulaic tradition and demonstrate that a poet’s stylistic centrality bears essentially no relationship to their productivity. Code and data are available to support replication and follow-up studies.
5: Using Deep Learning and Bayesian Optimization for Construction of Optical Coatings for PV Applications presented by Nishikant Sonwalkar (AdaptiveWaves Inc.)*; Siddha Karjee (AdaptiveWaves Inc.)
Multilayer thin film interference coatings enable precise control of optical transmission and reflection, with applications in photovoltaics, filtering, and laser protection. Optimizing these coatings is challenging because of the highdimensional design space defined by material selection and thickness. Traditional design methods based on the Fresnel equationsand optical constants defining the quarter-stack approach require significant effort to optimize desired optical performance for PV applications. In this study, we compare the AI directed
optimization approach using advanced computational strategies in two targeted cases. In the first, a Bayesian Monte Carlo optimizer was used to refine the thicknesses of MgF2/SiO2 bilayerstacks, demonstrating rapid convergence and strong performance within a constrained material system. In the second, a deep learning surrogate model optimized ZrO2/SiO2 alternating highcontrast coatings for space applications, achieving comparable best-case performance while maintaining design diversity and robustness. Transmission was evaluated using simulations of the transfer matrix in 380–1080 nm. These case studies highlight how physics-informed optimization and data-driven learning provide complementary strengths. Bayesian Monte Carlo excels at efficient convergence, while deep learning surrogates broaden
the search space and capture diverse high-performing solutions.
93: Predicting St. Louis Housing Prices with Machine Learning on Market and Assessor Data presented by Brian Adler (University of Missouri-St. Louis)*; Anne Brown (Washington University in St. Louis)
Housing markets are more complex than a simple supply-demand relationship. Prices are set by complex market and spatial neighborhood dynamics. Certain cities like St. Louis, MO have experienced dramatic population decline marked by extreme vacancy and abandonment. Amidst its population decline, St. Louis simultaneously demonstrates neighborhoods with sharp housing shortages and competition alongside others with entrenched vacancy and disinvestment mere blocks away from one another. We use supervised machine learning models to predict housing prices with a diverse feature set that incorporates spatial aspects of vacancy among other traditional housing amenities in St. Louis. Our results show how proximity to vacancy may impact a home’s value even more than its number of bedrooms. These findings, we expect, may prompt policymakers to combat vacancy even more urgently to maintain neighborhood market stability.
132: On the Geometric Limits of Transformer Defenses against Obfuscation Attacks: Latent Embedding Collapse & Performance–Robustness Gap presented by Becky Mashaido (University of the Pacific); Tapadhir Das (University of the Pacific)*
Prompt-injection attacks pose significant risks to language model safety, yet existing defenses are typically evaluated using classification performance. We show that high detection performance does not imply representational robustness. Specifically, multi-operator obfuscated prompts (combining homoglyphs, zero-width characters, and punctuation or emoji noise) can partially collapse onto the embedding manifold of clean prompts, a phenomenon we term latent embedding collapse. Results indicate that across multiple BERT-family encoders with varying depth and capacity, detectors achieve near-perfect classification performance, yet the minimal clean–obfuscated margin δ = 1.02, indicating near-overlap of obfuscated and clean embeddings. Obfuscated embeddings further exhibit elevated intra-class variance (3.33 ± 6.23), indicating severe latent-space instability despite high performance. These results reveal a substantial performance–robustness gap, demonstrating that standard evaluation metrics fail to capture latent embedding collapse and underlying geometric fragility. Our findings show that increasing model capacity does not eliminate latent embedding collapse, motivating geometry-aware robustness analysis as a necessary complement to performance-based evaluation for prompt-injection defenses.
144: Self-Evolving AI Systems: Stability–Plasticity Trade-Offs and Regret-Aware Learning in Deployed Machine Learning presented by RAJ KUMAR MYAKALA (CVS Health)*; Vinithya reddy Podduturi (World bank); AKHIL REDDY JAGIRAPU (university of north texas)
157: Multi-modal Causal Reasoning from Sparse Image Datasets: A Synthesis and Discovery Framework presented by Atul Rawal (Towson University)*; Adrienne Raglin (DEVCOM Army Research Lab
Causal discovery from observational data typically assumes access to complete, multi-modal datasets. However, many real-world scenarios often provide sparse and incomplete visual observations such as medical imaging without patient records, historical photographs without contextual metadata, or sensor degraded autonomous systems. This position paper proposes a framework for inferring causal relationships from sparse image datasets through cross-modal synthesis and uncertainty-aware discovery. It synthesizes missing modalities from visual data using foundation models, quantifies epistemic uncertainty in generated content, and performs causal structure learning that accounts for information incompleteness. We describe the system architecture, and discuss validation strategies. Our proposed framework enables causal reasoning in data-scarce environments across various domains such as medical imaging, robotics, and tactical autonomy.
209: Quantum-Hybrid Deep Learning Ensemble for Credit Card Fraud Detection presented by Tanay Patel (Arizona State University)*; Todd Hodges (American Express Co.); Glen Uehara (Arizona State University); Sutapa Samanta (American Express Co.); Dagen Wang (American Express Co.); Andras Ferenczi (American Express Co.); Andreas Spanias (Arizona State University)
Credit card fraud remains a persistent challenge in the financial sector, leading to significant monetary losses if not effectively mitigated. The growing volume of online transactions and the use of AI to mimic genuine spending patterns make fraud detection tasks increasingly more difficult. In this paper, we propose a hybrid ensemble pipeline in which transformers, graph attention networks, and variational autoencoders serve as deep feature extractors over classical and quantum-enhanced representations. A tree-based meta-learner (XGBoost) aggregates these feature embeddings for the final decision. The impact of quantum feature map representations when integrated with classical ensemble learners remains limited. To this effect, we examine and evaluate quantum circuit design strategies across different number of qubits, Pauli feature maps and data reuploading patterns under statevector simulations. We compare model configurations including tree-based models, standalone neural networks, and the proposed ensemble. Even without the quantum-enhanced features, the ensemble improves detection accuracy, achieving 2% higher averaged F1-scores over the classical tree-based baseline. Introducing quantum-enhanced features in the ensemble pipeline attains comparable performance over classical-only inputs, maintaining an AUC of approximately 98% on the European Credit Card Fraud Dataset. Simulator-based quantum noise causes slight degradation in averaged F1-scores, while averaged AUC is maintained relative to the statevector baseline across varying noise types and levels.
210: When Does Quantum Computing Provide Advantage for Malware Detection? Structural Complexity and the Intermediate Complexity Window presented by Nicholas Carducci (Monmouth University)*
Quantum malware-detection studies often compare one quantum method with one classical method on one dataset. That setup is useful for algorithm comparison, but it does not show where quantum discovery helps once the underlying malware representation changes and strong classical baselines are allowed. We address that question by defining an 11-rank representational-complexity scale. The paper makes three contributions. First, it gives an explicit experimental protocol for comparing quantum discovery with rank-appropriate classical baselines across three flows and two passes. Second, it shows that the measured advantage is concentrated in an intermediate representational window: at the co-occurrence and local-order sets, the quantum workflow retrieves malware where classical retrieves none (4.29% vs. 0%), and improves Recall by +150% (4.38% vs. 0.27%), while classical methods remain stronger for the lower-complexity statistical sets and the higher-complexity sets. Third, we demonstrate that this pattern remains after parameter tuning, indicating that the result is not an artifact of one favorable setting. The practical implication is limited but useful: evaluate quantum discovery for representations dominated by local pairwise structure, and prefer specialized classical methods where mature inductive biases already exist.
251: Short-form Text Rewriting with Phi Silica presented by Divya Tadimeti (Microsoft)*; Shawn Pan (Microsoft); Sameera Lanka (Microsoft); Chenghui Zhou (Microsoft); Sadid Hasan (Microsoft)
Short-form text rewriting is a constrained variant of paraphrasing in which limited context and high semantic density leave little room for variation. While large language models perform well on general paraphrasing, small language models (SLMs) often struggle with semantic fidelity and hallucination robustness in short-form settings. In this work, we present an empirical study of adapting an SLM, Phi Silica, for short-form rewrite through dataset curation, prompt distillation, parameter-efficient fine-tuning, and evaluation. We curate a dataset of short presentation-style text from public slide decks and use GPT-5-Chat both to generate rewrite supervision and to conduct LLM-as-a-judge evaluation. Our results show that finetuning improves semantic fidelity, reduces hallucinations, and increases preference win rate against GPT-5-Chat rewrites. The findings suggest that targeted adaptation for SLMs can substantially narrow the gap to cloud models and provide practical guidance for adapting SLMs to precision-critical rewrite tasks.
228 Operational Hallucination and Safety Drift in AI Agents presented by Shasha Yu, Fiona Carroll, Barry L. Bentley (Cardiff Metropolitan University)
Large language models (LLMs) serving as planners in tool-using autonomous agents introduce dynamic reliability risks in multi-turn execution. While single-turn safety mechanisms are relatively mature, extended interactions reveal structural vulnerabilities where initial alignment degrades over time. This paper empirically characterizes two observed failure modes across multiple state-of-the-art LLMs: Safety Drift, the gradual erosion of declared safety intent leading to constraint-violating actions (e.g., textual refusal followed by reconnaissance and unsafe execution), and Operational Hallucination, persistent repetitive tool calls indicative of flawed state perception (e.g., livelocks even in legitimate tasks). Through controlled multi-turn evaluation on high-stakes ethical dilemmas, malicious requests, and benign controls, we quantify these phenomena using declaration-action gap and livelock metrics, demonstrating their cross-model prevalence under direct execution protocols. Root-cause analysis attributes the instabilities to the decoupling of reasoning context from execution state in current agent loops. We propose an Action-Aware Supervision Layer—a lightweight, plug-and-play architectural blueprint incorporating intent-action consistency checks, runtime state tracking, and forced termination primitives. Post-hoc simulation on captured failure trajectories shows the layer can intercept observed violations without false positives on benign cases. This work advances agent reliability by shifting focus from linguistic safeguards to enforceable architectural mechanisms for responsible agentic AI.
ROOM 117 ABSTRACTS AND FINAL PAPERS
143: A Unified MLOps Architecture for Reliable Deployment of Generative AI Systems presented by Vasanth Rao Jadav (EPAM Systems)*; Shravya Bussari (HCL Tech)
Generative AI systems are increasingly deployed as core components of production software, enabling capabilities such as intelligent search, summarization, and conversational interfaces. Unlike traditional machine learning models, these systems produce nondeterministic,
human-facing outputs whose behavior may change due to variations in data, prompts, or model versions. In practice, many deployments rely on ad-hoc engineering practices, offering limited support for versioning, monitoring, rollback, and continuous validation, which
leads to unstable behavior in production. This paper presents a unified MLOps architecture for the reliable deployment and operation of generative AI systems. The proposed architecture integrates model and prompt versioning, orchestration, runtime monitoring, automated
recovery, and lifecycle management into a single engineering framework.We evaluate the architecture using operational reliability metrics, including failure detection latency, recovery time, and output stability under controlled changes. Experimental results show that treating generative behavior as a first-class production artifact significantly improves system robustness and operational transparency with minimal overhead.
170: Applied Machine Learning for Clinical Risk Assessment of Obstructive Sleep Apnea Using Electronic Health Records presented by Manoj Purohit (Marquette University)*; Octavian Ioachimescu (Department of Medicine Medical College of Wisconsin); Praveen Madiraju (Marquette University)
Long COVID has emerged as a persistent public health concern following the COVID-19 pandemic, with postacute symptoms that frequently overlap with clinical manifestations of obstructive sleep apnea (OSA), including disrupted sleep, abnormal breathing patterns, and increased metabolic and cardiovascular risks. While prior studies have primarily examined OSA as a pre-existing risk factor for COVID-19 outcomes, considerably less attention has been given to the reverse question whether individuals diagnosed with Long COVID face an elevated risk of subsequently developing OSA. In this study, we analyze longitudinal electronic health record (EHR) data from the Froedtert Health Medical System in Wisconsin to investigate the risk of OSA onset following a Long COVID diagnosis. Using a combination of feature selection and supervised machine learning techniques, we develop predictive models to identify individuals at elevated risk and to characterize the clinical, demographic, and encounter-level factors associated with postCOVID OSA development. Recursive Feature Elimination (RFE) is employed to identify the most informative predictors, while class imbalance is addressed using the Synthetic Minority Oversampling Technique with Edited Nearest Neighbors (SMOTEEN). Multiple classifiers are evaluated using cross-validation, including Gaussian Naive Bayes (GNB), Random Forest (RF), and Support Vector Machine (SVM). Among these, GNB achieves an area under the ROC curve (AUC) of 0.967 with recall and precision of 0.928 and 0.942, respectively. RF and SVM also demonstrate strong performance, achieving AUC values above 0.96. The resulting models consistently identify features such as encounter type, employment status, body mass index, insurance payer, age, sex, and race as key contributors to OSA risk following Long COVID. Beyond predictive performance, these findings highlight the potential of longitudinal EHR analytics to support earlier identification of patients at risk for post-COVID sleep disorders. While the primary contribution of this work lies in applied predictive modeling, we also outline future directions toward agentic AI–enabled clinical decision support systems that continuously observe patient data, reason over evolving risk trajectories, and surface actionable, context-aware insights within routine clinical workflows.
113: From Belief to Behavior: Modeling and Tracking Moral Alignment in Autonomous Agents presented by Soraya Partow (Georgia Southwestern state university)*; Satyaki Nan (Georgia Southwestern State University)
Ensuring moral alignment in autonomous agentic
systems remains a central challenge in AI safety. Current approaches such as Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI often struggle to balance transparency, adaptability, and value stability. This paper presents the Value-Aligned Structured Tracking (VAST) framework, a computational architecture designed to model and monitor the moral reasoning of agentic systems. VAST introduces a tripartite belief model, (π, κ, J), representing probabilistic credence, epistemic confidence, and moral justification. Belief updates occur through a Justification-Weighted Moral Consistency (JWMC) operator that preserves normative coherence while incorporating new evidence. Decision-making is guided by an Ethical Expected Utility with Cascading Constraints (EEUCC), which integrates soft ethical priorities into rational optimization. Finally, a Four- Gauge Alignment Monitoring System evaluates calibration, normative alignment, coherence, and reasoning quality in real time. Experimental evaluations across ethically sensitive scenarios demonstrate that VAST outperforms baseline alignment models (RLHF, Constitutional AI, and Value Learning) by 15–20% in aggregate alignment score while maintaining interpretability and low computational overhead. These findings highlight the potential of structured moral inference for developing trustworthy, transparent, and value-stable agentic AI systems.
127: Mitigating Risks in Agents and Robots presented by Mark Maybury (Lockheed Martin)*
The rapid advancement of both AI agents and humanoid robots is introducing a new set of risks characterized by novel attack surfaces and failure modes. This paper analyzes recent inappropriate use of and attacks against weaknesses in agents and robots to identify impacts on confidentiality, integrity, and availability and resultant harms. We present a set of countermeasures that help address underlying vulnerabilities, essential to ensure confidence and trust in these important capabilities to protect their potential to offer personalization, productivity, and prosperity.
130: Privacy Awareness in Large Language Models: Input Regurgitation and Prompt-Induced Sanitization for HIPAA and GDPR Compliance presented by Aravinda Jatavallabha (Independent Researcher)*; Venkatesha Matam (Independent Researcher); Vidisha Vijay (Independent Researcher)
Large Language Models (LLMs) are indispensable in privacy-sensitive fields like healthcare and hiring because of their remarkable natural language processing capabilities. However, the potential for input regurgitation and sensitive information in outputs raises serious privacy concerns, especially when it comes to regulations like GDPR and HIPAA. This study examines these privacy concerns by contrasting seven innovative models – GPT-3.5, GPT-4, GPT-4 Turbo, GPT-5.1, GPT-5.2, Gemini-2.5-pro, and Gemini-2.5-flash – and their propensity to retain and regurgitate PHI and personally identifiable information. We test these models on a combination of prompt-induced sanitization techniques and synthetic datasets to determine their ability to generate privacy-compliant outcomes. Our findings reveal that newer models, particularly GPT-5.2 and Gemini-2.5-pro, demonstrate enhanced privacy-preserving capabilities, with GPT-5.2 achieving near-zero leakage (0.0%) even with minimal prompt adjustments, while maintaining utility. Surprisingly, we discover that some of the advanced models have built-in privacy consciousness, in that sensitive information is sanitised even without instructions. This paper contains some useful recommendations and tips on increasing privacy compliance in LLM applications.
136: Beyond Artificial Certainty: Synthetic Socratic Inquiry for Trust Calibration in AI-Augmented Higher Education presented by Zomana Majid (ReliSource Inc.)*; Wasim Chaudhuri (ReliSource Inc.); Vijay Kanabar (Boston University)
The rapid integration of Generative Artificial Intelligence (GenAI) into higher education is enabling unprecedented scale in knowledge retrieval and synthesis. However, maximizing this utility necessitates a redefinition of the human role, specifically shifting from passive verification to active conceptual governance. Synthetic Socratic Inquiry (SSI) is introduced in this paper as a framework for systematically embedding ‘constructive challenge’ within the user-AI workflow. Through the integration of Modular Prompting architectures with the R.E.D. (Recognize, Evaluate, Draw) Model, the operational dynamics of GenAI are reconfigured from a statistical generation tool to those of a structured ‘sparring partner.’ Additionally, the Trust Calibration Index (TCI) is presented as a diagnostic tool for evaluating validation behaviors in learning exchanges. An ‘Efficiency Paradox’ is indicated by preliminary results, wherein the imposition of structured challenge extends procedural demand but substantially improves reasoning fidelity and error detection. This paper demonstrates that the proposed approach counters uncritical acceptance, not by restricting tools, but by enforcing a verifiable reasoning workflow that fosters critical thinking. Finally, evidence suggests that this human-in-the-loop optimization promotes superior innovation and substantive reliability compared with unstructured prompting alone.
138: From Scripts to Prompts: How Large Language Models Are Transforming Penetration Testing presented by Eman Alatawi (University of Tabuk); Umar Albalawi (University of Tabuk)*
Recent advances in Large Language Models (LLMs) have introduced new opportunities for automating of Penetration testing (PT) tasks. However, existing LLM-based frameworks continue to struggle with unstable long-horizon reasoning, context drift, inconsistent command generation, and limited ethical control. To address these challenges, this paper presents a structured autonomous multi-agent LLM-powered penetration testing framework that integrates prompt-engineered cognitive agents, Recon Strategist, Exploit Planner, and Failure Analyzer through a persistent knowledge graph and history engine. The methodology incorporates a detailed behavioral analysis of recent LLM-based PT systems, establishing taxonomies for prompt design, reasoning mechanisms, and self-reflective correction. Results demonstrate that model selection directly affects execution latency, reasoning stability, and operational cost, with GPT-5 models providing strong multi-step reasoning performance at modest cost, while open-source models offer low-cost or offline alternatives with varying latency trade-offs. The proposed architecture remains portable and stable across all backends, successfully completing autonomous PT workflows in a controlled virtual environment. Overall, the study contributes a reproducible methodology, a comprehensive comparative analysis of modern LLMs for offensive security automation, and an extensible multi-agent PT framework that advances the reliability, efficiency, and ethical governance of AI-driven cybersecurity operations.
155: Augmenting Large Language Models with Causal Risk Patterns for AI Deployment Risk Assessment presented by Gareth Mcconomy (Ulster University)*
Large language models (LLMs) achieve impressive benchmark performance (e.g., MMLU, HumanEval) yet exhibit systematic gaps in real-world deployment contexts. This is particularly concerning for AI governance applications where LLMs might identify risk categories but fail to trace the causal mechanisms through which harms actually unfold. We propose that structuring documented AI risks as explicit causal patterns, including sequential mechanisms (i.e., the chain of events from trigger to harm), feedback loops (i.e., cycles where outputs
become inputs, amplifying or dampening effects), and quantified effects from empirical research (i.e., magnitudes documented in peer-reviewed studies), can scaffold more reliable LLM reasoning through human-AI collaboration. We test this through a controlled study comparing pattern-augmented LLMs against vanilla LLMs using GPT-4o and Grok 4 to analyse five real estate AI deployment cases. Results show substantial improvements: GPT4o showed large improvement in causal depth (i.e., the number of sequential causal steps traced in the analysis) with Cohen’s d of 1.20, while Grok showed even larger effects (d = 1.93) alongside improved evidence grounding (d = 2.10). Validation using Gemini 2.5 Pro and DeepSeek as judges confirmed findings with moderate to high inter-rater agreement (r = 0.45 to 0.68). Qualitative analysis reveals that augmented outputs exhibit explicit mechanistic transfer (i.e., recognising that a causal pattern from one domain applies to a structurally similar situation in another), mapping patterns from domains such as aviation automation and reinforcement learning pricing to novel real estate contexts. These findings suggest that structured causal knowledge can enable a new modality of human-AI collaboration for AI risk assessment.
165: VAST-Blockchain: Hybrid Ledger Anchoring for Verifiable Governance and Auditability of Aligned AI presented by Soraya Partow (Georgia Southwestern State University)*; Satyaki Nan (Georgia Southwestern State University)
Centralized alignment deployments remain vulnerable to silent rule tampering and retrospective log manipulation. We introduce VAST-Blockchain, a hybrid architecture that strengthens deployment integrity for aligned AI by binding fast off-chain moral reasoning to governance-controlled on-chain state and tamper-evident commitments. We extend Value-Aligned Structured Tracking (VAST) by encoding ethical constraints as governed smart contracts and anchoring decision evidence using transparency-log style Merkle commitments via signed tree heads (STHs). We formalize a threat model spanning insider edits, hub rollback/equivocation, Byzantine governance, and Sybil
capture and specify an optimistic anchoring mechanism with slashable accountability. A prototype evaluation across three domains shows locked-rule immutability and externally verifiable log integrity under adversarial attempts with modest end-to-end anchoring overhead (2.3–2.95 s per decision) under modeled layer-2 finality delay (1.2–1.8 s)
Room 135 abstracts and final papers
125: Adaptive CTGAN: Data Augmentation for Imbalanced Cybersecurity Datasets presented by Devcharan Krishna Naik (University of Massachusetts Dartmouth); Yuchou Chang (University of Massachusetts Dartmouth)*; Ashokkumar Patel (University of Massachusetts Dartmouth)
Machine learning-based Network Intrusion Detection Systems (NIDS) are critical for identifying threats in large-scale networks but are severely compromised by the class imbalance inherent in real-world traffic. While Conditional Tabular GANs (CTGANs) provide a generative approach to data augmentation, existing frameworks often fail to effectively capture complex class relationships and minority patterns. This paper introduces Adaptive CTGAN, an enhanced framework improving upon the conditioning mechanism and training process of conventional CTGANs. The proposed model integrates a learnable class embedding layer to encode semantic relationships among attack categories and employs a dynamic conditional sampling strategy to adaptively prioritize difficult samples. Evaluated on the CIC-IDS-2017 benchmark using the Train-on-Synthetic, Test-on-Real (TSTR) paradigm, Adaptive CTGAN demonstrates superior synthetic data fidelity and diversity. Experimental results confirm significant improvements in minority-class detection and downstream classifier F1-scores, validating the method for robust, privacy-preserving NIDS data augmentation.
46: An Explainable Machine Learning Approach of Detecting Denial of Service Attacks in 5G Networks presented by João Victor Jales Ramos ( Electrical Engineering Department, Federal University of Campina Grande); Matheus Vilarim P. dos Santos (Electrical Engineering Department, Federal University of Campina Grande); Fernando Luiz F. Barros (Electrical Engineering Department, Federal University of Campina Grande); Edmar C. Gurjão (Electrical Engineering Department, Federal University of Campina Grande); Paola Pimentel Furlanetto (Independent Researcher)*
This paper presents an end-to-end forecast-to-replenishment pipeline that integrates machine learning (ML) demand forecasting with mixed integer linear programming (MILP) for retail inventory control. Using the Walmart weekly sales dataset, we benchmark linear and tree-based forecasters (including Random Forest and gradient boosted trees (XGBoost)) and a stacked ensemble that combines complementary learners via out-of-fold stacking. Across rolling origin evaluation with a 4 week horizon, the stacked ensemble achieves the best Root mean squared error (RMSE) and coefficient of determination (R2) on average, with lower variance across folds while Random Forest attains the
lowest symmetric mean absolute percentage error (sMAPE), indicating robust scale normalized error across heterogeneous store department series. We therefore use the stacked ensemble demand signal as the default input to the MILP replenishment model, which minimizes holding, ordering, and stockout costs subject to service level constraints. The resulting ML-MILP system demonstrates that improving forecast skill and coupling it to optimization yields measurable gains in both cost and service performance.
69: Reservoir Computing Based Anomaly Detection in Industrial Control Systems presented by Andre Slonopas (APUS)*; Edward Olbrych (Virginia Polytechnic Institute and State University); Michael Corley (Army National Guard); Jacob Strahan (Huntington Ingalls Industries)
Traditional enterprise payroll systems suffer from monolithic architectures that tightly couple business functions, creating significant challenges in scalability, maintainability, and extensibility. Users must navigate complex technical interfaces, and adding new capabilities requires invasive code modifications. This paper presents an A2A-MCP multi-agent architecture for enterprise payroll management systems utilizing Google’s Agent to-Agent (A2A) protocol combined with Large Language Model (LLM) driven intelligent routing and Model Context Protocol (MCP) integration. The proposed system addresses these critical limitations by implementing a modular, scalable agent-based
architecture where specialized agents handle distinct payroll functions including payment processing, employee management, payslip generation, and reporting. Beyond query-response interactions, our system supports full CRUD (Create, Read, Update, Delete) operations and autonomous task triggering, enabling agents to execute business actions rather than merely providing information. Our approach introduces three key innovations: (1) an LLM-powered routing mechanism that analyzes natural language commands and directs them to
appropriate specialized agents with confidence scoring, (2) dual LLM integration where both the orchestrator and individual agents leverage language models for intent understanding and response generation respectively, and (3) MCP server integration enabling standardized tool access for database operations including transactional writes. The architecture demonstrates significant advantages in modularity, extensibility, and user experience through natural language interaction, while enabling plug-and-play agent deployment without system modifications.
139: MRI-AgentNet: A Vision Language Models-Based Multi-Agent AI System for Solving Inverse Problems in MRI presented by Gulfam Ahmed Saju (University of Massachusetts Dartmouth); Marjan Akhi (University of Massachusetts Dartmouth); Yuchou Chang (University of Massachusetts Dartmouth)*
Magnetic Resonance Imaging (MRI) reconstruction and motion correction are fundamental inverse problems in medical imaging, traditionally addressed using deep learning models trained for specific corruption types. However, these models often lack generalizability, requiring manual intervention to determine the appropriate correction strategy. Furthermore, single-agent automated AI systems are inherently prone to bias and overfitting, as they rely on a single inference pathway, which may fail in cases of ambiguous or mixed corruption types. To address these limitations, we introduce MRI-AgentNet, a Vision-Language Model (VLM)-powered multi-agent AI system that autonomously detects MRI corruption types and selects the optimal correction model. MRI-AgentNet employs a hierarchical multi-agent framework where multiple AI agents independently assess corruption types. The system refines corruption predictions through structured cross-validation, which ensures consensus before selecting the optimal correction model. This multi-agent system reduces bias by leveraging diverse inference pathways and structured hierarchical validation. To ensure optimal correction, a MetaModel learns a structured mapping between agentinferred corruption labels and the appropriate correction model. Experimental results demonstrate that MRI-AgentNet achieves expert-level accuracy in both corruption detection and correction, which reduces the need for manual intervention by radiologists. By automating the entire MRI corruption analysis and restoration pipeline, MRI-AgentNet enables a fully autonomous, AI-driven correction process. This can enhance the efficiency of MRI reconstruction while reducing operational costs. This work introduces a scalable, multi-agent AI framework that integrates vision-language reasoning, inverse problem-solving, and MRI artifact restoration. It establishes a foundation for fully automated AI-driven medical imaging.
140: Quantization and its Effects about Reinforcement Learning on Quantum Neural Network Wrapped in Classical Encoder presented by Jacob Fronzaglia (University of Massachusetts Dartmouth); Yuchou Chang (University of Massachusetts Dartmouth)*; Christopher Hixenbaugh (Naval Undersea Warfare Center Division Newport)
This paper provides an analysis of a comparison between a quantized quantum neural network and a non-quantized quantum neural network to see if quantization has any possible benefits to quantum neural networks in reinforcement learning. The REINFORCE algorithm, Softmax-Parameterized Quantum Circuits wrapped in a classical encoder, fake quantization, and the CartPole-v1 dataset were also used for these experiments. The findings include a faster convergence to higher mean average reward (from 10 runs) for the quantized quantum neural network, standard deviation and policy entropy steadily decreasing for a longer number of episodes indicating more stable learning, and correlations found between entropy and average rewards and its derivatives. We also found some possible downsides for quantized quantum neural networks, including longer runtimes and less stable evaluation rewards. The experiments were run on Google Colab on the CPU device.
148: Multi-Dimensional Evaluation Framework for Generative AI Applications with Dynamic Data Sources presented by Vishwanath Pattar (Hewlett Packard Enterprise)*; Vishnuram Venkataswamy Sampathkumar (Hewlett Packard Enterprise); Lohith Reddy Kalluru (Hewlett Packard Enterprise)
Embedding Large Language Models (LLMs) into enterprise Generative AI applications that use dynamic data streams introduces distinctive validation challenges. Existing deterministic quality assurance approaches are inadequate for non- deterministic, context-sensitive outputs that depend on changing live data sources and user question phrasing. This paper presents a Multi-Dimensional Evaluation Framework that comprehensively assesses the correctness, reliability, and operational behavior of generative AI systems. The framework consists of schema-driven test data generation from various data sources integrated into the system to obtain answers for user questions and expand them into semantically diverse paraphrased variants using LLM pipelines. To overcome the challenge of validating against changing data, the framework implements dynamic ground-truth generation techniques. Security guardrails are validated, and LLM-based techniques assess analytical accuracy, conceptual alignment, hallucination risk, and contextual relevance. The framework also incorporates a comprehensive observability layer, including structured logging, test segregation, hierarchical reporting, and end-to-end traceability. Together, these capabilities offer a scalable, repeatable framework that ensures the reliability of enterprise Generative AI application deployments.
171: Multiplier-Free LLM Linear Layers via Weights-Only Power-of-Two QAT presented by Ikteder Akhand Udoy (Boise State University)*; Omiya Hassan (Boise State University)
Large language models (LLMs) are commonly instantiated as Transformer architectures, where deployment cost is dominated by multiply-accumulate (MAC) operations, especially in the attention and MLP linear projections. We present a weights-only power-of-two quantization-aware training (Po2-QAT) method that constrains these linear-layer weights to compact signed power-of-two sums, enabling a multiplier-free shift–add realization (with an appropriate backend); we report operation-count/energy proxies. Po2-QAT leverages an on-the-fly CSD-k projection in the forward pass with a straight-through estimator (STE) and stabilizes optimization by progressively tightening the constraint: during training we blend floating-point weights with their Po2-projected counterparts, increasing the Po2 contribution linearly from 0 to 1 over a fixed number of steps. We evaluate Po2-QAT on WikiText-2 language modeling for DistilGPT-2 and Llama-3.2-1B, reporting perplexity, throughput (tokens/s), and peak GPU memory. On DistilGPT-2, the baseline achieves 36.22 test PPL versus 37.67 with Po2-QAT (k=2), with throughput decreasing from 111,950 to 74,448 tokens/s and peak memory increasing from 1.05 to 2.20 GB. On Llama-3.2-1B, the baseline achieves 13.60 test PPL versus 14.00 with Po2-QAT, with throughput decreasing from 520 to 460 tokens/s and peak memory increasing from 8.80 to 9.90 GB. Using a standard per-operation cost model, an arithmetic energy proxy for constrained linear layers indicates ≈ 0.043k relative arithmetic energy, resulting ≈ 0.086× for k=2. Across both models, Po2-QAT eliminates multipliers in the constrained linear layers by replacing each MAC contribution with a bounded k-term shift-add sequence under matched data, tokenization, and optimization settings.
72: Historical Project and Dependent Task Status Verification-Aware Integration of DEVOPS with AI Using QF-MT presented by Vineeth Billa Kanti (Independent Researcher)*; Likitha Guthikonda (Independent Researcher)
The development, deployment, and operations of software are automated by the integration of Artificial Intelligence (AI) with the DevOps lifecycle. Yet, the existing studies didn’t aggregate the historical project regarding the dependent source linked with other dependencies and testing based on dependency task status. Thus, this paper presents a Quotient Filter-based Merkle-Tree (QF-MT)-based historical project and dependent task status verification-aware integration of DevOps with AI. Firstly, the Digital Signature (DS) is created in the plan stage using Cyclic Quadratic Probing-based Digital Signature Algorithm (CQP-DSA). Then, the DS is verified in the code stage. Similarly, this DS
creation and verification are performed at each and every stage. The code development and integration are done in the code stage for the requirements. Here, pre-processing methods like variable extraction, control flow graph generation, and Merkle tree construction by QF-MT are done. Afterward, the developed code is compiled and run in the build stage. Further, for testing, the compiled code is given. Here, by comparing the code with the QF-MT, a dependent task is identified. Afterward, by using Transfer Learning-Enhanced Rational Contrastive Gated Recurrent Unit (TL-ERCGRU), the defect analysis is done. Next, the tested code is released, deployed, operated, and monitored,
and the feedback is further provided. As per the outcome, the proposed model took a less merkle tree creation time of 4747ms.
Keywords: Dependency task status, Silhouette Density-Based Spatial Clustering of Davies Cohen's Bouldin Applications with Noise (SDBSCDCBAN), Development and Operations (DevOps), Artificial Intelligence (AI), Control flow graph generation, Defect analysis, and Transfer Learning (TL).
181: DeepDetect: Learning All-in-One Dense Keypoints presented by Shaharyar Ahmed Khan Tareen (University of Houston)*; Filza Khan Tareen (National University of Sciences and Technology); Xiaojing Yuan (University of Houston)
Keypoint detection is the foundation of many computer vision tasks, including image registration, structure-from-motion, 3D reconstruction, visual odometry, and SLAM. Traditional detectors (SIFT, ORB, BRISK, FAST, etc.) and learning based methods (SuperPoint, R2D2, QuadNet, LIFT, etc.) have shown strong performance gains yet suffer from key limitations: sensitivity to photometric changes, low keypoint density and repeatability, limited adaptability to challenging scenes, and lack of semantic understanding, often failing to prioritize visually important regions. We present DeepDetect, an intelligent, all-in-one, dense detector that unifies the strengths of classical detectors using deep learning. Firstly, we create ground-truth masks by fusing outputs of 7 keypoint and 2 edge detectors, extracting diverse visual cues from corners and blobs to prominent edges and textures in the images. Afterwards, a lightweight and efficient model: “ESPNet”, is trained using fused masks as labels, enabling DeepDetect to focus semantically on images while producing highly dense keypoints, that are adaptable to diverse and visually degraded conditions. Evaluations on Oxford, HPatches, and Middlebury datasets demonstrate that DeepDetect surpasses other detectors achieving maximum values of 0.5143 (average keypoint density), 0.9582 (average repeatability), 338,118 (correct matches), and 842,045 (voxels in stereo 3D reconstruction).
Created and maintained by Ballos Associates