Abstract:
Leveraging the information-rich and large volume of Electronic Health Records (EHR), deep learning systems have shown great promise in assisting medical diagnosis and regulatory decisions. Although deep learning models have advantages over the traditional machine learning approaches in the medical domain, the discovery of adversarial examples has exposed great threats to the state-of-art deep learning medical systems. While most of the existing studies are focused on the impact of adversarial perturbation on medical images, few works have studied adversarial examples and potential defenses on temporal EHR data. In this work, we propose RADAR, a Recurrent Autoencoder based Detector for Adversarial examples on temporal EHR data, which is the first effort to defend adversarial examples on temporal EHR data. We evaluate RADAR on a mortality classifier using the MIMIC-III dataset. Experiments show that RADAR can filter out more than 90% of adversarial examples and improve the target model accuracy by more than \(90\%\) and F1 score by 60%. Besides, we also propose an enhanced attack by introducing the distribution divergence into the loss function such that the adversarial examples are more realistic and difficult to detect.
