A Review of Adversarial Deep Learning Models in Neuroscience Research and Clinical Practice

Khaled Sh. Gaber^1,*, Ehsan khodadadi²

¹Computer Science and Intelligent Systems Research Center, Blacksburg 24060, Virginia, USA

²Department of Chemistry and Biochemistry, University of Arkansas, Fayetteville, AR 72701, USA

Emails: khsherif@jcsis.org; Ehsank@uark.edu

Abstract

Adversarial deep learning has, therefore, been tabled as one of the key research focus areas in neurosciences, and both the opportunities and drawbacks for the operation of deep learning models on neuroimaging and diagnostic jobs have been unveiled. This review examines these models' weaknesses from adversarial attacks, which can severely affect diagnosis and patient care. For example, it has been shown that slight disturbances in the level of EEG signals can confuse more profound learning algorithms employed for the identification of epilepsy, which can lead to severe diagnostic mistakes. In addition, GANs have the dual role of generating realistic neuroimaging data that can improve diagnostic processes while at the same time using adversarial images that expose the deficits of current models. This duality highlights the need to securely defend models against such risks and employ adversarial training and bio-mimic-based resilient neural network techniques. The consequence of these discoveries should not be underestimated because they reveal the necessity of showing further safety in using deep learning techniques in clinical practices. In addressing these weaknesses, the principle goal of this research is not only to help improve the diagnostic systems but also to expand the knowledge on how adversarial deep learning might affect the health, well-being and safety of patients in neuroscience.

Keywords: Adversarial deep learning; Neuroimaging; Diagnostic accuracy; Generative adversarial networks; Model robustness; Patient safety