Personalized insulin dose manipulation attack and its detection using interval-based temporal patterns and machine...

Area of Science:

• Biomedical Engineering
• Cybersecurity
• Data Science

Background:

• Diabetes devices like insulin pumps improve patient quality of life but are vulnerable to cyberattacks.
• Insulin dose manipulation attacks (overdose/underdose) can lead to severe hypoglycemia or hyperglycemia, endangering patients.
• Existing detection methods may not address sophisticated, personalized attacks.

Purpose of the Study:

• To propose a sophisticated, personalized insulin dose manipulation attack on insulin pump systems.
• To develop and evaluate an automated machine learning-based system for detecting these attacks.
• To assess the impact of model granularity on attack detection accuracy.

Main Methods:

• Collected a large dataset (225,780 logs) from real insulin pumps and continuous glucose monitors (CGMs) of 47 type 1 diabetes patients.
• Developed a novel method for predicting blood glucose (BG) levels to simulate manipulation attacks.
• Employed advanced temporal pattern mining and machine learning algorithms (Logistic Regression, Random Forest, TPF, ANN) for attack detection.

Main Results:

• Achieved nearly 90% accuracy in predicting BG levels, enabling realistic attack simulation.
• Successfully detected insulin manipulation attacks using temporal patterns and various ML methods.
• Overdose attacks were detected more effectively than underdose attacks (AUC scores).
• Adult vs. pediatric models showed better overdose detection, while the general model excelled in underdose detection; pediatric attack detection was more challenging.

Conclusions:

• The proposed BG prediction method is accurate and valuable for medical applications.
• The developed ML-based system effectively detects sophisticated insulin pump cyberattacks.
• Understanding model granularity is crucial for optimizing attack detection in diverse patient populations (adults vs. children).