Adversarial Examples-Security Threats to COVID-19 Deep Learning Systems in Medical IoT Devices.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Cybersecurity

Background:

• Medical Internet of Things (IoT) devices are crucial for managing pandemics like COVID-19.
• Deep learning (DL) algorithms analyze data from medical IoT for COVID-19 detection (e.g., radiological images, thermal data, facial recognition).
• Existing DL models show vulnerabilities to adversarial perturbations, posing security risks.

Purpose of the Study:

• To evaluate the security of DL-based COVID-19 diagnostic methods against adversarial examples (AEs).
• To demonstrate the vulnerability of DL models lacking defenses against adversarial attacks.
• To raise awareness about adversarial threats in healthcare AI systems.

Main Methods:

• Testing various COVID-19 diagnostic DL methods using crafted adversarial examples (AEs).
• Analyzing the impact of adversarial perturbations on model performance.
• Detailing the AE generation process and attack model implementation.

Main Results:

• DL models for COVID-19 detection without adversarial defenses are vulnerable to attacks.
• Adversarial perturbations can compromise the reliability of AI-driven diagnostic tools.
• The study successfully generated and applied AEs to existing DL-based COVID-19 applications.

Conclusions:

• Healthcare systems relying on DL for pandemic management must address adversarial vulnerabilities.
• Implementing defensive strategies against adversarial attacks is essential for securing medical IoT and AI.
• Further research and development are needed to create robust and secure AI diagnostic tools for public health.