Evaluating Differential Privacy on Correlated Datasets Using Pointwise Maximal Leakage

2025-02-08

Annual Privacy Forum

AI Summary

This paper addresses the weakening of differential privacy (DP) guarantees in settings with correlated data, where conventional privacy budgets fail to accurately quantify actual information leakage. To resolve this, it introduces Pointwise Maximal Leakage (PML) into the DP analytical framework for the first time. Leveraging information-theoretic modeling, the method characterizes how data correlations amplify privacy leakage—thereby relaxing the standard i.i.d. assumption—and establishes a PML-based quantification framework for privacy loss under correlation. Theoretical analysis derives tight bounds on privacy loss, while empirical evaluation across diverse correlated datasets demonstrates that PML more precisely reflects actual leakage than the classical ε-parameter, significantly enhancing the verifiability and practicality of privacy mechanism design. The core contribution is the first theoretically grounded and empirically validated DP quantification paradigm explicitly tailored to correlated data.

Technology Category

• Machine Learning: Privacy
• Data Mining & Knowledge Management: Representing, Reasoning, and Using Provenance, Trust
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Problem

• Assessing differential privacy on correlated datasets.
• Quantifying information leakage using pointwise maximal leakage.
• Highlighting limitations of DP in preserving privacy.

Innovation

• Leverages Pointwise Maximal Leakage
• Assesses Differential Privacy weaknesses
• Highlights need for tailored mechanisms