Formula: alpha(lam) <= T q^2 lam(lam+1)/((1-q) sigma^2); delta = min_lam
exp(alpha(lam) - lam eps)
Usage
Dpacct(sigma, sample_rate, steps, delta = 1e-05, max_order = 64)
Arguments
- sigma
See Usage.
- sample_rate
See Usage.
- steps
See Usage.
- delta
See Usage.
- max_order
See Usage.
Value
List with epsilon, order, logmgf, delta, sigma,
sample_rate, steps.
References
Abadi, Chu, Goodfellow, McMahan, Mironov, Talwar and Zhang (2016), Deep
Learning with Differential Privacy, CCS'16, arXiv:1607.00133. Lemma 3 for the
sampled-Gaussian log moment, Theorem 2 for composability and the tail bound. Verified
against the paper.
Examples
Dpacct(sigma = 1.1, sample_rate = 0.01, steps = 1000)$epsilon
#> [1] 2.044643
