A hybrid blockchain based deep learning model for multivector attack detection in internet of things enabled healthcare systems

Abstract Novel advances in healthcare-related Internet of Things (IoT) systems have recently had significant impacts on clinical decision-support systems (CDSS) and patient health monitoring. Securing networks using conventional cybersecurity models becomes increasingly challenging as the regularity of open-access networks increases, exposing critical attack regions. The research presented here recommends a hybrid model combining Blockchain (BC) and Intrusion Detection Systems (IDS) built on Deep Learning (DL) (Hybrid BC + DL Model) to address such problems. This method integrates Artificial Intelligence (AI) and distributed trust management (DTM) for providing healthcare-specific security across the entire system. The hypothetical model focuses primarily on a Deep Sparse Autoencoder (DSAE) that helps standardize and reduce all the different traffic generated by medical IoT devices into a small, discrete graphical representation. These embedded technologies were securely encrypted by applying multiple layers of authentication. The primary layer is a standard Bidirectional Long Short-Term Memory (BiLSTM) network that captures temporal dependencies within healthcare data. The next layer is a set of high-powered sensor networks that can detect Distributed Denial of Service (DDoS), Man-in-The-Middle (MiTM), and Brute-Force Attacks (BFA). The simulation test result is subsequently validated using a Bayesian Product-of-Experts (BPoE) method that incorporates contextual medical challenges into the analysis, applies temperature scaling during testing, and improves clinical implementation accuracy. Networks that integrate BC technology have fixed audit logs, Smart Contracts (SC) that automate access control, and Practical Byzantine Fault Tolerance (PBFT) consensus protocols, which permit the secure communication of attack data across the healthcare industry. The proposed model improves conventional benchmarks by 7.39–20.42% and SOTA (State-of-the-Art) by 1.00–7.19%, attaining accuracy scores of 96.73% and 93.58% on the IoT–Flock and the Canadian Institute for Cybersecurity Internet of Things 2023 (CICIoT2023) datasets. In both cases, the detection latency is less than 16 ms, demonstrating real-time feasibility in controlled experimental settings. When training and testing on distinct datasets, the average score ranges from 11.52 to 13.55%, indicating moderate generalization capability as measured by cross-dataset testing. DSAE-based Feature Extraction (FE) generated a 7.28% boost to accuracy, while the Bayesian Fusion Mechanism (BFM) resulted in around a 5.06% boost in accuracy. The outcome results of this study indicate that applying trained models to collected data resulted in a significant 9.39% improvement in accuracy. There was a 7.28% boost in accuracy when DSAE-based Feature Extraction (FE) was applied, 9.39% when algorithms that had been trained were used for collected data, and 5.06% when the Bayesian Fusion Mechanism (BFM) was implemented. The research findings have confirmed all these improvements. The analysis shows that the BC function has a Network Throughput (NT) of more than 698 Times Per Second (TPS), consensus delays of less than 468 ms, and validation success rates of more than 99.4%. Accuracy is maintained above 99.6%, and SC-based security measures are fully operational within 245 ms. The proposed model aims to secure the healthcare system (HCS) and prevent data loss from digital attacks.