Adversarial Attacks on Medical Image Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Increasing volume of medical images from diverse radiological techniques.
• Computer-aided diagnosis (CAD) systems enhance clinical applications.
• Pixel inaccuracies in imaging facilities can lead to misclassification and incorrect clinical decisions.

Purpose of the Study:

• Investigate the impact of one-pixel and multi-pixel adversarial attacks on Deep Neural Network (DNN) models.
• Evaluate the effect of pixel manipulation on classification performance and robustness.
• Assess the vulnerability of medical image classification to pixel-level attacks.

Main Methods:

• Conducted one-pixel and multi-pixel level attacks on DNN models.
• Utilized common multiclass and multi-label medical image datasets.
• Performed experiments varying the number of affected pixels to analyze impact on DNNs.

Main Results:

• Medical images demonstrated significant vulnerability to pixel attacks.
• Classification performance and robustness of DNN models were negatively affected by pixel manipulations.
• Even minor pixel alterations compromised image integrity for diagnostic purposes.

Conclusions:

• Pixel attacks pose a substantial threat to the accuracy of medical image classification.
• The robustness of DNN models against adversarial attacks is critical for reliable computer-aided diagnosis.
• Further research is needed to develop defense mechanisms for medical imaging AI.