Estimating the success of re-identifications in incomplete datasets using generative models

Area of Science:

• Data privacy
• Statistical modeling
• Computational social science

Background:

• Rich datasets are crucial for data-driven research but pose privacy risks.
• De-identification and sampling are standard methods to anonymize data.
• Existing anonymization techniques may not meet current privacy standards.

Purpose of the Study:

• To develop a generative copula-based method for estimating individual re-identification likelihood.
• To assess the effectiveness of de-identification in protecting privacy in large datasets.
• To evaluate whether current anonymization practices align with GDPR standards.

Main Methods:

• Utilized a generative copula-based approach to model data distributions.
• Estimated the probability of correct re-identification for individuals within datasets.
• Validated the method on 210 diverse populations, assessing predictive accuracy using AUC scores.

Main Results:

• The proposed method achieved high AUC scores (0.84-0.97) for predicting individual uniqueness.
• A low false-discovery rate was observed, indicating reliable re-identification risk estimation.
• Analysis revealed that 99.98% of Americans could be re-identified using 15 demographic attributes.

Conclusions:

• Current de-identification practices, even with heavy sampling, are insufficient for modern privacy standards like GDPR.
• The 'release-and-forget' model of data anonymization is technically and legally inadequate.
• There is a critical need for more robust privacy-preserving techniques in data sharing.