Reproduce “Differential Privacy as a Mutual Information Constraint” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Paul Cuff, Lanqing Yu

Researchers from Princeton University established Mutual-Information Differential Privacy (MI-DP) as an information-theoretic equivalent to differential privacy, demonstrating that an ǫ-DP mechanism implies ǫ-MI-DP, which in turn implies an (ǫ', δ')-DP mechanism. This reformulation clarifies DP's strong adversary assumption and simplifies the derivation of its properties.

Key insights

  • Expressing differential privacy as conditional mutual information (I(X_i; Y | X_{-i})) explicitly highlights the 'strong adversary assumption' and clarifies DP's guarantees.
  • The MI-DP framework enables simplified proofs for core differential privacy properties, such as composition theorems, using fundamental information theory rules.
  • Differential privacy, through MI-DP, is shown to restrict information leakage only when conditioned on the remainder of the database, revealing its limitations in bounding unconditional leakage with correlated data.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.