Enabling realistic health data re-identification risk assessment through adversarial modeling

Area of Science:

• Biomedical Informatics
• Data Privacy
• Cybersecurity

Background:

• Current re-identification risk methods for biomedical data often use worst-case attacker models.
• These models can overestimate risk, leading to excessive data editing and hindering data sharing.

Purpose of the Study:

• To introduce a framework for assessing re-identification risk that considers an attacker's actual resources and capabilities.
• To evaluate how pragmatic attacker models impact risk assessment compared to worst-case scenarios.

Main Methods:

• Integrated three risk measures: prosecutor, journalist, and marketer risks.
• Computed re-identification probabilities based on attacker capabilities and subject disclosure.
• Utilized case studies with over 1,000,000 patient records from Vanderbilt University Medical Center.
• Simulated pragmatic attacks using voter registration lists and social media posts.

Main Results:

• The developed framework showed substantially lower re-identification risk in pragmatic scenarios versus worst-case.
• Median worst-case risk was 0.987, but pragmatic scenarios reduced this by 90.1% (voter lists) and 100% (social media).
• These findings were consistent across various adversarial capabilities.

Conclusions:

• Re-identification risk is situationally dependent, not absolute.
• Appropriate adversarial modeling, considering pragmatic attackers, can enable wider biomedical data sharing.
• This approach balances data utility with robust privacy safeguards.