A hybrid deep learning model for intrusion detection in cloud-based implantable medical devices

Abstract

The rapidly evolving technologies in the healthcare sector, such as implantable medical devices (IMDs), require advanced security solutions that leverage the intelligence capabilities of these technologies while ensuring optimal safety and reliability. The IMD technology redefines healthcare service delivery by offering timely interventions, minimally invasive treatment options, and continuous patient condition monitoring to improve quality of life. Despite these achievements, IMDs face unauthorised access, data manipulation, and denial-of-service attacks, which conventional security solutions are limited in handling due to resource constraints within IMD ecosystems. As a result, different machine learning and deep learning frameworks have been proposed for real‐time threat detection. However, they still suffer from overfitting, slow inference, and excessive resource demands, hindering their effective integration into the IMD ecosystem. The study's primary goal was to design and develop a hybrid of deep autoencoders, convolutional neural networks, and long short-term memory (LSTM) strategies to provide a comprehensive detection model that reduces inference time for deployed models while enhancing performance. Autoencoders provide the fundamental architecture of the detection model, while convolutional neural networks are used in the encoder and decoder for simplicity and to capture nonlinear data effectively. The Long Short-Term Memory captures temporal dependencies in the model, enhancing overall detection capabilities. The study adopted an experimental approach, developing a hybrid deep autoencoder model to test its performance against convolutional neural networks, long short-term memory, and other conventional machine learning techniques. The results demonstrate that the hybrid model outperformed standalone models, achieving high accuracy scores across the datasets. The best model in the ICU dataset achieved 100% accuracy, precision, recall, and F1 score, and a false positive rate of 0.00%. The WUSTL had an accuracy of 79.32%, a recall of 79.92%, a precision of 79.41%, a specificity of 79.24%, and a false positive rate of 20.59%. The Edge IIoT dataset had a recall, F1, and accuracy of 96.87%, a precision of 96.94%, a specificity of 96.88%, and a false-positive rate of 3.12%. The model’s inference time was substantially reduced compared to the standard deep autoencoder model across the datasets, providing a lightweight detection environment for the intrusion detection system.