Detecting data manipulation attacks on physiological sensor measurements in wearable medical systems

• Published in: EURASIP Journal on Information Security 2018-09, Vol.2018 (1), p.1-21, Article 13
• Main Authors: Cai, Hang, Venkatasubramanian, Krishna K.
• Format: Article
• Language: English
• Subjects: Communications Engineering; Data manipulation attacks; Data-centric attack detection; Engineering; Networks; Security Science and Technology; Signal,Image and Speech Processing; Systems and Data Security; Wearable medical systems
• ISSN: 2510-523X; 2510-523X
• DOI: 10.1186/s13635-018-0082-y

Recent years have seen the emergence of wearable medical systems (WMS) that have demonstrated great promise for improved health monitoring and overall well-being. Ensuring that these WMS accurately monitor a user’s current health state is crucial. This is especially true in the presence of adversaries who want to mount data manipulation attacks on the WMS. The goal of data manipulation attacks is to alter the measurements made by the sensors in the WMS with fictitious data that is plausible but not accurate. Such attacks force clinicians or any decision support system AI, analyzing the WMS data, to make incorrect diagnosis and treatment decisions about the patient’s health. In this paper, we present an approach to detect data manipulation attacks based on the idea that multiple physiological signals based on the same underlying physiological process (e.g., cardiac process) are inherently related to each other. We capture the commonalities between a “target” sensor measurement and another “reference” sensor measurement (which is trustworthy), by building an image reconstruction-based classifier and using this classifier to identify any unilateral changes in the target sensor measurements. This classifier is user-specific and needs to be created for every user on whom the WMS is deployed. In order to showcase our idea, we present a case study where we detect data manipulation attacks on electrocardiogram (ECG) sensor measurements in a WMS using blood pressure measurement as reference. We chose ECG and blood pressure—in arterial blood pressure (ABP) form—because both are some of the most commonly measured physiological signals in a WMS environment. Our approach demonstrates promising results with above 98% accuracy in detecting even subtle ECG alterations for both healthy subjects and those with different cardiac ailments. Finally, we show that the approach is general in that it can be used to build a model for detecting data manipulation attacks that alter ABP sensor measurements using the ECG sensor as reference.