Adversarial Attack and Defense in Breast Cancer Deep Learning Systems

Area of Science:

• Artificial Intelligence in Medicine
• Medical Image Analysis
• Computational Pathology

Background:

• Deep learning significantly enhances medical diagnosis, particularly for breast cancer pathology images, improving accuracy and efficiency.
• Existing deep learning models, however, are susceptible to adversarial attacks, posing security risks for critical medical applications.
• Vulnerabilities in deep learning systems can lead to misclassification of breast cancer pathology images, impacting patient care.

Purpose of the Study:

• To investigate the vulnerability of deep learning models used in breast cancer pathology image recognition to adversarial attacks.
• To develop and evaluate a robust deep learning system with enhanced defense mechanisms against such attacks.
• To improve the security and reliability of AI-driven diagnostic tools for medical imaging.

Main Methods:

• Utilized the Fast Gradient Sign Method (FGSM) adversarial attack algorithm to test the susceptibility of breast cancer deep learning systems.
• Developed a novel deep learning system specifically designed for breast cancer pathology image recognition with improved security features.
• Evaluated the defense performance of the new system against adversarial perturbations.

Main Results:

• Demonstrated that current deep learning systems for breast cancer pathology image analysis are vulnerable to FGSM adversarial attacks, causing misclassifications.
• The newly developed deep learning system exhibited superior defense performance against adversarial attacks compared to existing models.
• The enhanced system maintained high accuracy in classifying breast cancer pathology images even under attack.

Conclusions:

• Deep learning models for medical image diagnosis require robust security measures to prevent adversarial manipulation.
• The developed deep learning system offers a more secure and reliable approach for breast cancer pathology image recognition.
• Enhancing the security of medical deep learning systems is crucial for their safe and effective clinical deployment.