DataTerminology.comNEW
Also known as: RDP
A privacy accounting framework defining differential privacy via Rényi divergence, enabling tighter composition bounds for iterative algorithms.
DESCRIPTION
Rényi Differential Privacy (RDP) reframes privacy guarantees using Rényi divergence at order α. This facilitates precise composition across repeated mechanisms, improving accounting for iterative training like DP-SGD. RDP parameters can be converted to (ε, δ) for reporting and compliance. It relates to the moments accountant technique and zero-concentrated DP. RDP is widely implemented in privacy libraries and used by industry for practical budgeting. Proper calibration of noise and order selection is essential.
ALTERNATIVE DEFINITIONS
Ilya Mironov (arXiv)
We give a formulation of differential privacy based on Rényi divergence, and derive tight composition bounds for privacy loss.
Introduces RDP and emphasizes tight composition bounds.
Google Differential Privacy
Rényi differential privacy (RDP) is a relaxation of differential privacy based on Rényi divergence that simplifies privacy analysis under composition.
Focuses on simplification and composition in practical tooling.
Wikipedia
Rényi differential privacy defines privacy guarantees via the Rényi divergence of order α, enabling tighter analyses for iterative mechanisms.
Highlights the role of order α and iterative analyses.
RELATED TERMS
Privacy Loss Composition Moments Accountant Epsilon Delta Privacy Budget Gaussian Mechanism DP-SGD zCDP
COMPETING TERMS
Differential Privacy Zero-Concentrated Differential Privacy
RELATIONS
- Broader: Differential Privacy
KNOWLEDGE AREAS
Data Privacy & Compliance AI Governance Data Science AI MLOps
TARGET AUDIENCE
Data Privacy Officer Data Scientist AI Engineer MLOps Engineer Chief AI Officer
EXTERNAL SOURCES
- Wikipedia - Differential privacy
- arXiv - Original RDP paper
- Google Differential Privacy - RDP documentation and library usage
- OpenDP - Open-source DP library and resources
TOP QUESTIONS
Why use Rényi Differential Privacy for accounting?
It provides tighter composition bounds for iterative algorithms, improving accuracy of privacy budgeting over many steps.
Category: Relevancy
How does RDP relate to (ε, δ)-DP?
RDP parameters can be converted to equivalent (ε, δ) guarantees, enabling reporting and compliance within familiar DP frameworks.
Category: Misunderstanding
What is currently a hot topic around RDP?
Privacy accounting for large-scale DP-SGD with subsampling, and standardized reporting of privacy budgets in ML pipelines.
Category: Topic
What are common mistakes using RDP?
Incorrectly choosing order α or miscalibrating noise can yield misleading privacy guarantees; use vetted libraries and cross-check conversions.
Category: Mistakes
What risks exist with poor RDP accounting?
Underestimated privacy loss can violate policies or regulations; overestimation may harm utility unnecessarily.
Category: Risks
What alternatives compete with RDP?
Zero-concentrated DP (zCDP) and the moments accountant provide related approaches to privacy composition analysis.
Category: Alternatives
How is RDP applied in ML training?
In DP-SGD, RDP tracks cumulative privacy loss over iterations, guiding noise levels and stopping criteria.
Category: Management
Where to start learning RDP?
Read Mironov’s original paper, then explore Google’s DP library documentation and OpenDP resources.
Category: Example
How should RDP be reported to stakeholders?
Convert RDP to (ε, δ) values, include assumptions, sampling rates, and mechanism details to ensure transparency.
Category: Requirements