Disparate privacy risks from medical AI

Area of Science:

• Medical Artificial Intelligence
• Data Privacy
• Machine Learning Security

Background:

• Medical AI models improve diagnostics but use sensitive patient data.
• Existing privacy attack assessments are aggregate, not patient-level.
• Individual patient privacy risks in AI training data are poorly understood.

Purpose of the Study:

• To conduct patient-level privacy audits of AI diagnostic models.
• To evaluate the effectiveness of membership inference attacks (MIAs) on individual patient data.
• To identify disparities in privacy risks among different patient groups.

Main Methods:

• Focused on membership inference attacks (MIAs).
• Audited AI models across diverse medical datasets.
• Analyzed attack success rates at the individual patient level.
• Investigated the impact of model capacity and patient demographics on privacy risk.

Main Results:

• MIAs achieved near-perfect success rates for individual patients.
• High attack success rates increased with model capacity.
• Underrepresented groups faced disproportionately higher attack success.
• Aggregate privacy metrics significantly underestimated individual patient risk.

Conclusions:

• Patient-level privacy audits reveal substantial individual privacy risks in medical AI.
• Aggregate privacy metrics are insufficient for assessing real-world privacy exposure.
• Disparities in privacy risk highlight the need for equitable privacy protection.
• Further research is needed for advanced risk assessment and mitigation strategies.