Incorporating Machine Learning into Medical Device Security

Abstract

Internet of Things healthcare devices such as pacemakers and glucose delivery systems have disrupted the medical industry by paving the way for personalized and remote systems. These devices, however, do not come without risk. Implantable and wearable medical devices (IWMDs) demand precision and reliability, which, when combined with small-scale implementation, leaves little room for complex and comprehensive security measures. The lack of robust security in remote medical devices can leave patients vulnerable to privacy breaches or physical attacks. The need for availability and accessibility of IWMDs in the event of an emergency eliminates the possibility of traditional security measures, such as encryption. Previous research has developed the use of a complementary medical monitor (MedMon) to detect and prevent malicious adversaries from attacking remote healthcare systems. This research incorporates machine learning techniques into MedMon in order to differentiate between normal and abnormal user behavior of an insulin delivery system. Prior to this investigation, researchers had yet to explore medical security through machine learning. We believe that by incorporating machine learning into medical device security, we can create a personalized protection plan to minimize vulnerabilities to wireless attacks. In order to collect data, we use HackRF coupled with GNU Radio Companion to observe and analyze the wireless communication between an insulin pump and its controller. We collect physical and behavioral features such as frequency and power of the signal, time differentials, and bolus injection levels to classify the data accordingly. Various machine learning techniques including decision tress, support vector machines, and ensemble methods aid in this classification. The results of this research show a promising future for this technique, with accuracies up to 91% for the Naïve Bayes model.