ProbDeepNet: A Cognitively Aligned Severity Intelligence Framework for Latent Bug Semantics in Medical IoT Firmware

Abstract

Healthcare systems are increasingly vulnerable to cyber-attacks, with over 93 million patient records breached in 2023 and an annual increase of 22% in vulnerability disclosures affecting hospital networks. Manual approaches to bug severity detection are error-prone and inefficient, often failing to manage the growing volume of unstructured medical security reports. To address this, we propose a Natural Language Processing (NLP) framework that leverages a Medical NLP dataset comprising incident reports, advisories, and vulnerability disclosures. The methodology begins with NLP preprocessing and Exploratory Data Analysis (EDA) to clean, normalize, and visualize the dataset. Following this, Lightweight Robustly Optimized bidirectional encoded representation of transformers (LRoBERT) is used with word embeddings for semantic representation of text, ensuring reduced computational cost compared to heavy transformer models. Unlike existing systems based on Stochastic Gradient Descent (SGD) and Natural Gradient Boosting (NGB) Classifier, the proposed pipeline integrates Deep Neural Network (DNN) based feature selection with Natural Gradient Boosting (NGB) model also known as ProbFusionNet to improve predictive performance. The system classifies bug severity into two categories: Normal and Severity, providing actionable insights for prioritizing critical vulnerabilities. By combining contextual embeddings with advanced feature selection, the proposed approach enhances accuracy, reduces misclassification, and ensures faster real-time response. This innovation contributes to strengthening the cybersecurity posture of the healthcare ecosystem by delivering an efficient and scalable solution for automated bug severity detection.