The article discusses the curse of full-rank covariance matrices in Gaussian mechanisms for differential privacy. Existing mechanisms suffer from high expected accuracy losses due to this curse. To address this issue, the Rank-1 Singular Multivariate Gaussian (R1SMG) mechanism is introduced, achieving (ε,δ)-DP with reduced noise magnitude. The R1SMG perturbs query results using a singular multivariate Gaussian distribution with a randomly generated rank-1 positive semi-definite matrix as its covariance matrix. By leveraging a clue from previous work by Dwork et al., the R1SMG mechanism provides better utility stability and privacy guarantees compared to classic Gaussian mechanisms.
toiselle kielelle
lähdeaineistosta
arxiv.org
Tärkeimmät oivallukset
by Tianxi Ji,Pa... klo arxiv.org 03-15-2024
https://arxiv.org/pdf/2306.02256.pdfSyvällisempiä Kysymyksiä