核心概念
全ての既存のガウス機構は、完全ランク共分散行列の呪いに苦しんでおり、新しいRank-1特異多変量ガウス(R1SMG)メカニズムはこの呪いを解消します。
要約
Abstract:
Differential privacy (DP) via output perturbation is standard.
Existing Gaussian mechanisms suffer from full-rank covariance matrices.
R1SMG mechanism proposed to overcome this curse.
Introduction:
DP fundamental for privacy-preserving database operations.
Classic Gaussian mechanism adds i.i.d. noise to query results.
The Curse of Full-Rank Covariance Matrices:
Expected accuracy loss equals trace of covariance matrix.
Curse identified in various types of Gaussian mechanisms.
Lifting the Curse: Main Contributions:
R1SMG achieves (ε,δ)-DP with lower expected accuracy loss.
Stability and utility improvements over existing mechanisms.
An Ignored Clue from Dwork et al.:
Clue suggests using rank-1 covariance matrix for privacy guarantee.
R1SMG: Multivariate Gaussian Noise with A Random Rank-1 Covariance Matrix:
R1SMG mechanism introduced to achieve DP with reduced noise magnitude.
Data Comparison Table:
Mechanisms
Expected Accuracy Loss
Stability
Classic Gaussian
≥ CC(∆2 f)2
Unstable
Analytic Gaussian
≥ CA(∆2 f)2
Unstable
MVG Mechanism
≥ CM(∆2 f)2
Unstable
R1SMG Mechanism
≥ CR(∆2 f)2
Stable
統計
特定のガウス機構が期待される精度損失を下回ることができる条件:σ∗ ≥ 2(∆2 f)2/εψ
引用
"Less is more in the sense that noise of a much lower order of magnitude is needed compared with that of existing Gaussian mechanisms."