Defending against adversarial attacks on Covid-19 classifier: A denoiser-based approach

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cybersecurity

Background:

• Deep Neural Networks (DNNs) show high accuracy in COVID-19 detection from chest X-rays.
• DNNs are susceptible to adversarial attacks, compromising diagnostic reliability.
• Existing defenses like adversarial training require model replacement and retraining.

Purpose of the Study:

• To evaluate the adversarial robustness of DNN-based COVID-19 classifiers.
• To assess the effectiveness of the High-Level Representation Guided Denoiser (HGD) as a defense mechanism for medical image analysis.

Main Methods:

• Adversarial attacks, including Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), were employed to test model robustness.
• The High-Level Representation Guided Denoiser (HGD) architecture was evaluated as a defense technique.
• Experiments were conducted in both white-box and black-box settings.

Main Results:

• Adversarial attacks significantly decreased the accuracy of COVID-19 classifiers.
• HGD demonstrated an accuracy increase of up to 82% in the white-box setting.
• HGD failed to defend against adversarial samples in the black-box setting.

Conclusions:

• DNN-based COVID-19 detection models are vulnerable to adversarial attacks.
• HGD shows potential as a transferable defense for medical imaging, particularly in white-box scenarios.
• Further research is needed to enhance HGD's effectiveness in black-box settings.