Home

Project Title

Non-detectable security exploits in healthcare

Abstract

Deployment of embedded systems for military, healthcare, and many commercial applications has seen a considerable surge in recent years. This has opened the door for adversarial parties to exploit security vulnerabilities of these systems to inject malicious processes with the purpose of degrading their performance, modifying their functionality, or extracting sensitive data. The current project explores various possible attacks targetting a human fall detection system. The target system utilizes sensory data from a smartwatch to detect the action of human falling using machine learning algorithms. We will explore:

1. poisoning attacks on the training process of the recurrent neural network (RNN) used for detecting falling,
2. evasion attack on a trained network to hide the act of falling from the network,
3. and timing attack to sabotage the input data to the network.

The outcome of this project is used by developers to design more secure and tamper-resistant fall detection systems.