Phantom Anonymization: Adversarial testing for membership inference risks in anonymized health data

Area of Science:

• Health Informatics
• Data Privacy
• Cybersecurity

Background:

• Medical research datasets containing sensitive individual information pose significant privacy risks.
• Anonymization techniques are vital for protecting health data but quantifying residual risks remains challenging.

Purpose of the Study:

• To introduce a novel framework for quantifying residual membership inference risks in anonymized tabular data.
• To adapt and apply techniques from synthetic data assessment for evaluating anonymization effectiveness.

Main Methods:

• Developed a framework utilizing a classifier trained to detect target records in anonymized datasets.
• Employed data anonymized using the same methods as the target dataset for classifier training.
• Conducted experiments across diverse anonymization strategies and adversarial conditions.

Main Results:

• The proposed framework effectively identifies residual privacy risks in anonymized datasets.
• Anonymization method effectiveness is contingent on the chosen privacy model and data modification strategies.
• Significant variations in risk were observed based on how data met predefined risk thresholds.

Conclusions:

• The framework offers an empirical method for assessing membership inference risks across various anonymization techniques.
• Enables direct comparison of residual risks between anonymized and synthetic datasets due to shared methodology.