Attack Detection Using Federated Learning in Medical Cyber-Physical Systems

Medical Cyber-Physical Systems (MCPS) are networked systems of medical devices that provide seamless integration of physical and computation components in healthcare environments to deliver high quality care by enabling continuous monitoring and treatment. As MCPS store sensitive medical data and personal health data, security breaches and unauthorized access to this information can lead to severe repercussions for both the patient and hospital in the form of loss of privacy, abuse, physical harm and liability. The heterogeneity of devices involved in these systems (such as body sensor nodes and mobile devices) introduce large attack surfaces and hence necessitate the design of effective security solutions for these environments. In this paper, we design and implement a massively distributed, machine-learning-based intrusion detection solution for MCPS. Specifically, we explore the concept of Federated Learning to minimize the communication and computation costs involved in traditional machine learning based solutions. We evaluate our design with real patient data and against security attacks such as Denial of Service, data modification, and data injection. Experimental results illustrate that our system achieves high detection accuracy of 99.0% and a False Positive Rate of 1.0% along with a reduced network communication overhead. Lastly, we show that the system can cope with unevenly distributed data and is a scalable solution that leverages the computing resources of many mobile devices.