FIS-AI 2026 Abstracts

Area 1 - FIS-AI
Full Papers Short Papers

Area 1 - FIS-AI

Full Papers
Paper Nr: 16
Title:

A Unified Deep Learning-Based Framework for Anticipatory Driver Drowsiness Detection Leveraging the Composite Drowsiness Score
Authors:

Vikul J. Pawar, Kailash D. Kharat, Suraj R. Pardeshi, Vaibhav M. Mokale and Vijayshri A. Injamuri
Abstract: One of the leading causes of highway accidents all over the world is driver drowsiness. To resolve this problem, this paper will present a framework of proactive drowsiness detection and avoidance of accidents using deep learning technology. In the model, the spatial, temporal, and contextual features are integrated by a hybrid architecture that is built on Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Vision Transformers (ViTs), as learning frameworks. CNNs identify important spatial features of each frame and LSTMs learn to represent fatigue changing with time. ViTs improve performance through attention mechanisms used to highlight important behavioral indicators such as long duration of eye closure, frequency of yawning, and movement of the head. These complementary features are combined to produce a dynamic drowsiness measurement of the level of alertness of the driver. The framework is tested on the NTHU Drowsy Driver Detection Dataset and real-life driving videos in different light and environmental conditions. The proposed CNN-LSTM-ViT model outperforms standalone CNN and LSTM models with precision of 98.7 and recall of 98.9. The framework has low latency rate which makes it appealing in real-time Advanced Driver Assistance Systems.
Paper Nr: 18
Title:

From Natural Language to Interpretable Code: Automated Code Generation for Healthcare with Large Language Models - A Comparative Analysis
Authors:

Yuexi Chen, Gauri Vaidya, Alison N. O'Connor and Meghana Kshirsagar
Abstract: This article presents a comparative evaluation of three large language models (LLMs), namely GPT-4o, Gemini 2.0 Flash 2.0 Flash, and Claude 3.5 Sonnet, examining their ability to automate key healthcare workflows while adhering to algorithmic constraints and supporting interpretability and fairness. The models were evaluated using Python, JavaScript, and Go under varying levels of prompt completeness across four healthcare tasks of increasing complexity: bed allocation, dynamic patient bed reallocation, ambulance dispatch, and patient triage. We introduce a multidimensional evaluation framework that captures model performance across task complexity, prompt completeness, and programming language, with an emphasis on generating functionally correct, transparent, and reliable code. This framework enables a systematic analysis of how effectively LLMs translate natural language specifications into executable logic under realistic, constraint rich healthcare scenarios. Experimental results show that all three models generate constraint compliant solutions for simpler tasks such as bed management. However, as task complexity increases and multiple constraints must be balanced, clear performance differences emerge. Claude 3.5 Sonnet consistently outperforms GPT-4o and Gemini 2.0 Flash 2.0 Flash by producing more robust, interpretable, and reliable code. These findings highlight Claude 3.5 Sonnet's stronger potential for transparent and dependable automation of critical healthcare services using LLM based code generation. The code is publicly available at: https://github.com/gauriivaidya/alter-automated-healthcare-tasks.
Paper Nr: 19
Title:

Semantic Consistency in Hierarchical Classification via Probabilistic Logic Constraints
Authors:

Filippos Gouidis, Antonis Argyros and Dimitris Plexousakis
Abstract: Deep neural networks trained for hierarchical classification often produce semantically inconsistent predictions, assigning higher confidence to fine-grained classes than to their parent categories. Such violations undermine interpretability and trust, particularly in safety-critical applications. In this work, we propose a lightweight neurosymbolic framework for hierarchical classification that enforces explicit semantic consistency at the probability level during training. Our approach introduces a differentiable logic-based loss that constrains parent-child class probabilities to respect known taxonomic relationships, together with a consensus loss that aligns coarse-grained predictions with aggregated fine-grained evidence. Unlike post-hoc explanation methods, interpretability in our framework emerges intrinsically from the logical structure imposed on the model’s output space. Through controlled ablations and cross-backbone evaluations on three benchmark datasets we demonstrate that the proposed semantic-aware training substantially reduces logical violations without degrading classification accuracy. Furthermore, semantic auditing using representational similarity and error depth metrics shows that consistency-enforced models exhibit more rational failure modes, confusing semantically related classes rather than producing implausible errors. These results indicate that simple, explicit semantic constraints can significantly improve the interpretability and trustworthiness of hierarchical classifiers while remaining compatible with standard deep learning architectures.
Short Papers
Paper Nr: 5
Title:

Generative AI and LLMs in Cybersecurity:~Literature Review and Future Research Directions
Authors:

Abibat A. Lasisi, Santiago Rocha and Ramoni O. Lasisi
Abstract: Generative Artificial Intelligence (GenAI) and Large Language Models (LLMs) are advanced technologies that enable multimodal capabilities, including texts, audios, videos, sounds, images, voice, and expressions.~GenAI and LLMs are emerging technologies with limited research connecting them to cybersecurity.~GenAI provides opportunities to automate threat detection, response, and predictive analysis.~The use of LLM-based agents in cybersecurity presents promising new directions to combat digital threats.~This research provides a comprehensive literature review of the application of GenAI and LLMs in cybersecurity.~Guided by the well-known cybersecurity framework of the National Institute of Standards and Technology, we classify the selected articles using a thematic analysis approach.~Our review then identifies potential future research areas in emerging cybersecurity applications for the successful adoption of GenAI and LLMs.
Paper Nr: 10
Title:

Origins of Toxicity in Human-LLM Conversations
Authors:

Johannes Schneider, Arianna Casanova Flores and Anne-Cathérine Kranz
Abstract: Toxic content generated by AI is a critical issue for regulation, safety and practical use. Large language models (LLMs) often strictly refuse to respond to user prompts flagged as moderately toxic, which raises important questions about censorship and limits their applications. However, the origins of AI toxicity remain unclear: Is it user-driven, or does it emerge autonomously? This study analyzes real-world human interactions with LLMs using state-of-the-art industry-toxicity detectors. Our findings suggest that while LLMs are often blamed for generating toxic content, such toxic outputs are frequently provoked, encouraged or explicitly requested by users. Furthermore, while some interactions are intentionally harmful, others arise from curiosity, humor, or emotional expression. These more subtle cases challenge conventional toxicity classifications, suggesting the need for a more nuanced and flexible approach. To address these issues, we propose strategies to enhance toxicity detection and mitigation in LLMs, promoting more effective and contextually aware AI system design.
Paper Nr: 12
Title:

A Structured Knowledge Model Linking Healthcare Privacy Risks to Privacy-Enhancing Technologies for AI-Enabled Health Systems
Authors:

Fatima Khalique, Ibrahim Tariq Javed and Katie Crowley
Abstract: Increased applications of AI in healthcare has simultaneously raised concerns for privacy, security and regulatory compliance. Health data exchange due to its sensitive nature in context of patient privacy poses challenges that need to be addressed. Therefore, many privacy preserving and privacy enhancing techniques have been proposed and applied to address the various risks associated with patient privacy. these include Homomorphic Encryption, Trusted Execution Environments (TEEs), Differential Privacy, Federated Learning, Zero-Knowledge Proofs, Secure Multi-Party Computation (SMPC). However, while these techniques offer strong guarantees and mature methodologies, there exists a gap in guiding software practitioners and architects for applying appropriate Privacy Enhancing Techniques (PET) to their specific healthcare privacy risk scenario at a given AI workflow stage. This paper presents a structured knowledge model that maps healthcare privacy risk scenarios to appropriate privacy enhancing technologies for AI enabled systems. We present a taxonomy of privacy risks derived from existing literature and analyse the capabilities and limitations of available Privacy Enhancing techniques and integrate them into a structured conceptual framework. We then synthesize these findings into a set of governance-aware architectural patterns that systematically integrate PETs into AI pipelines to support privacy-by-design, accountability, and regulatory compliance. The proposed synthesis bridges the gap between theoretical PET research and practical AI system design by providing reusable architectural guidance rather than isolated technical solutions. The proposed framework enables design-stage privacy requirement assessment for AI enabled healthcare systems. This conceptual framework aims to guide future work on addressing privacy concerns when using AI enabled healthcare technologies.
Paper Nr: 17
Title:

A Data-Centric Approach to Mitigating Class Imbalance in Low-Resource Telugu Natural Language Processing
Authors:

Kedhar Eashwar Seetammagari and Nikola Nikolov
Abstract: Machine Learning and Natural Language Processing (NLP) have made significant progress in recent years; however, many widely spoken languages remain low-resource due to limited annotated data and severe class imbalance. Telugu is a prominent example, where minority classes in text classification tasks are substantially underrepresented, leading to biased performance and unreliable evaluation. This work addresses class imbalance in Telugu text classification through targeted minority-class oversampling, supported by de-duplication and rigorous evaluation controls. The framework is evaluated on a large Telugu benchmark across four tasks: Sentiment Analysis, Emotion Identification, Hate Speech Detection, and Sarcasm Detection, using fine-tuned Telugu BERT and RoBERTa models. The proposed framework yields relative improvements in Macro F1-score of 2.7%, 15.0%, 4.4%, and 3.6% across the four tasks, alongside improved minority-class recall, including gains for the severely underrepresented fear class. Overall, the article demonstrates a reproducible approach for mitigating class imbalance and improving evaluation reliability in low-resource language NLP.
Paper Nr: 8
Title:

An Interpretable Insider Threat Detection Framework: Correlating Digital Communications and Behavioral Metadata
Authors:

Nikitha Thammaiah, Neha Girish, Naru Meghana, Shivaprakash G. and S. Pradeep Kumar Kenny
Abstract: Insider threats represent a critical security challenge as malicious actors exploit legitimate access to compromise organizational systems and exfiltrate sensitive data. This paper presents a correlation-driven insider threat detection framework that integrates Natural Language Processing (NLP) analysis of digital communications with User Behavioral Analytics (UBA). The framework employs a hybrid ensemble of Isolation Forest, One-Class SVM, and Autoencoder models to detect anomalous communication and interaction patterns. To address the ``black-box'' nature of AI-based security systems, we introduce an interpretability layer that explains user risk scores through semantic and behavioral context, aligning with Security Operations Center (SOC) analyst workflows. Evaluation on the Enron email dataset demonstrates that the proposed multi-modal approach significantly outperforms traditional unimodal baselines, yielding robust detection rates with minimal false positives. The results confirm that correlating linguistic intent with behavioral metadata provides superior anomaly separation compared to isolation-based methods.