Edgeworth Accountant: An Analytical Approach to Differential Privacy Composition

Hua Wang authored a paper introducing the Edgeworth Accountant, an analytical method for composing differential privacy guarantees. The method leverages the f-differential privacy framework to track privacy loss under composition, providing non-asymptotic (ε,δ)-differential privacy bounds. The associated files, last updated on 2026-05-18, are available on figshare under a CC-BY-4.0 license.

Use Cases

• Composing privacy guarantees for multi-step private algorithms based on the described privacy-loss log-likelihood ratios (PLLRs)
• Estimating tight upper and lower bounds on (ε,δ)-differential privacy for training private models based on the Edgeworth expansion method
• Applying simplified privacy accounting to any noise-addition mechanism based on the described distribution simplification technique

Strengths

• The method provides non-asymptotic privacy bounds, a feature highlighted as a main advantage
• The approach does not significantly increase computational cost with the number of composed mechanisms, as stated in the description
• Files are available in multiple formats (PDF, ZIP, DOCX) totaling 1.4 MB

Limitations

• The dataset's row count and column structure are unknown, limiting suitability assessment
• Column-level documentation is absent; field semantics must be inferred after download

Provenance

Source

figshare

Freshness

Last updated 2026-05-18 13:51:53