The Threat of Adversarial Attack on a COVID-19 CT Image-Based Deep Learning System

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cybersecurity in Healthcare

Background:

• The COVID-19 pandemic spurred AI adoption in medical diagnostics, particularly for CT image analysis.
• Existing AI research for COVID-19 primarily focuses on performance, neglecting crucial security and reliability aspects.
• Adversarial attacks pose a significant threat to the integrity of AI-driven medical diagnostic systems.

Purpose of the Study:

• To investigate the vulnerability of a deep learning system for COVID-19 CT image classification to adversarial attacks.
• To quantify the impact of adversarial attacks on the diagnostic accuracy of AI models.
• To propose strategies for enhancing the security and reliability of AI systems in medical imaging.

Main Methods:

• Development of a deep learning model for classifying COVID-19 and non-COVID-19 CT images.
• Implementation of the Fast Gradient Sign Method (FGSM) adversarial attack algorithm.
• Evaluation of model performance before and after the adversarial attack.

Main Results:

• The initial deep learning system achieved an average accuracy of 76.27% in identifying COVID-19 CT images.
• Adversarial attacks significantly degraded classification performance, reducing accuracy for non-COVID-19 images from 80% to 0%.
• Demonstrated the susceptibility of AI models to manipulation, compromising diagnostic reliability.

Conclusions:

• Deep learning systems for COVID-19 CT image analysis are vulnerable to adversarial attacks.
• There is a critical need to address security and reliability in medical AI development.
• Future research should focus on creating robust and secure AI models for trustworthy medical diagnostics.