Federated Learning Privacy Budget Calculator
Combine cohort size, per-round sampling, and Gaussian noise settings to approximate the accumulated privacy loss for federated learning with DP-SGD.
Educational approximation. Validate with a full privacy accountant before publishing compliance attestations.
Examples
- 50,000 total clients, 500 sampled each round, 200 rounds, σ = 1.2 (δ defaults to 1/50,000) ⇒ ε ≈ 0.56 unitless at δ = 2.00e-5 with a 1.00% sample rate.
- 10,000 clients, 200 sampled, 100 rounds, σ = 1.0, δ = 1e-6 ⇒ ε ≈ 1.09 unitless and the composition term contributes 0.40 of the loss.
FAQ
What definition of ε does this calculator use?
It applies the analytical moments accountant approximation for DP-SGD, combining the square-root term and the quadratic composition term published in the Abadi et al. 2016 framework.
How should I choose δ?
Set δ lower than 1 divided by the total participant pool to keep the probability of a privacy breach negligible. Regulatory guidance often accepts δ ≤ 1e-5 for consumer deployments.
Does this work for adaptive clipping?
Yes, provided the adaptive clipping enforces a deterministic bound per update. If clipping thresholds vary by round, use the lowest bound observed when interpreting ε as a worst-case estimate.
Additional Information
- Result unit: privacy loss ε reported as a dimensionless quantity with the matched δ value.
- Defaults: δ falls back to 1 ÷ total clients when the optional field is blank to mirror common regulatory practice.
- Sampling ratio is capped between 0% and 100% to reflect without-replacement participation each round.