Security and Privacy in Machine Learning for Health Systems: Strategies and Challenges.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Cybersecurity in Medicine

Background:

• Machine learning (ML) offers significant potential for physician decision support in healthcare.
• However, ML applications in health systems are vulnerable to security attacks and privacy breaches.
• Addressing these vulnerabilities is crucial for safe and effective ML deployment in medicine.

Purpose of the Study:

• To investigate and analyze existing research on security and privacy in machine learning for health systems.
• To identify prevalent attack vectors, defense mechanisms, and privacy-preserving strategies.
• To discuss the challenges associated with implementing these strategies and to guide future research.

Main Methods:

• A systematic literature review was conducted, involving manual searches and defined search strings.
• Papers were filtered by title, abstract, and full text, followed by an analysis of their contributions.
• Forty relevant papers focusing on attacks, defenses, and privacy in health ML were collected and discussed.

Main Results:

• Identified trends in attacks, including universal adversarial perturbations (UAPs), generative adversarial network (GAN)-based attacks, and DeepFakes.
• Highlighted defense trends such as adversarial training, GAN-based strategies, and out-of-distribution (OOD) detection for adversarial examples (AE).
• Found key privacy-preserving strategies including federated learning (FL), differential privacy, and hybrid approaches to enhance FL.

Conclusions:

• Security and privacy in ML for health are critical due to increasing risks and vulnerabilities.
• The study provides a comprehensive overview of current strategies and challenges.
• This research aims to guide future investigations into securing ML applications in healthcare systems.