Towards evaluating the robustness of deep diagnostic models by adversarial attack

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Cybersecurity in Healthcare

Background:

• Deep learning models, particularly neural networks, are crucial for medical image analysis and disease diagnosis.
• However, these models exhibit vulnerabilities to adversarial examples, posing significant security risks in clinical settings.
• Adversarial examples are imperceptible perturbations that can lead to high-confidence erroneous predictions by deep diagnostic models.

Purpose of the Study:

• To evaluate the robustness of deep diagnostic models against adversarial attacks in medical image classification tasks.
• To investigate the mechanisms by which adversarial examples compromise model performance.
• To develop and validate novel defense strategies to mitigate the impact of adversarial examples.

Main Methods:

• Conducted two types of adversarial attacks on three deep diagnostic models across single-label and multi-label classification tasks.
• Analyzed model performance, intermediate features, feature discriminability, and label correlations for both clean and adversarial images.
• Designed and implemented two novel defense methods: Multi-Perturbations Adversarial Training (MPAdvT) and Misclassification-Aware Adversarial Training (MAAdvT).

Main Results:

• Deep diagnostic models demonstrated significant unreliability when subjected to adversarial attacks.
• Analysis revealed how adversarial examples affect model outputs and internal feature representations.
• The proposed defense methods, MPAdvT and MAAdvT, substantially improved model robustness against adversarial examples.

Conclusions:

• Deep diagnostic models are susceptible to adversarial attacks, compromising their reliability in clinical applications.
• MPAdvT and MAAdvT effectively enhance the resilience of deep diagnostic models against adversarial perturbations.
• Implementing these defense mechanisms is crucial for secure and trustworthy AI in medical diagnostics.