ruby/tensorflow_privacy
1
0
Fork 0
forked from 626_privacy/tensorflow_privacy
Code Pull requests Activity

Move doc str below functions

Browse source
This commit is contained in:
woodyx218 2020-02-21 09:30:30 -05:00 committed by GitHub
parent fe82de2cfe
commit b9b2e8670f
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 24 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
tensorflow_privacy/privacy/analysis/gdp_accountant.py
View file

@ -23,14 +23,14 @@ import numpy as np
from scipy.stats import norm from scipy.stats import norm
from scipy import optimize from scipy import optimize
# Total number of examples:N
# batch size:batch_size
# Noise multiplier for DP-SGD/DP-Adam:noise_multiplier
# current epoch:epoch
# Target delta:delta
def compute_mu_uniform(epoch, noise_multi, N, batch_size): def compute_mu_uniform(epoch, noise_multi, N, batch_size):
'''Compute mu from uniform subsampling.''' '''Compute mu from uniform subsampling.'''
# Total number of examples:N
# batch size:batch_size
# Noise multiplier for DP-SGD/DP-Adam:noise_multi
# current epoch:epoch
T = epoch*N/batch_size T = epoch*N/batch_size
c = batch_size*np.sqrt(T)/N c = batch_size*np.sqrt(T)/N
return np.sqrt(2)*c*np.sqrt(np.exp(noise_multi**(-2))*\ return np.sqrt(2)*c*np.sqrt(np.exp(noise_multi**(-2))*\
@ -38,6 +38,11 @@ def compute_mu_uniform(epoch, noise_multi, N, batch_size):
def compute_mu_poisson(epoch, noise_multi, N, batch_size): def compute_mu_poisson(epoch, noise_multi, N, batch_size):
'''Compute mu from Poisson subsampling.''' '''Compute mu from Poisson subsampling.'''
# Total number of examples:N
# batch size:batch_size
# Noise multiplier for DP-SGD/DP-Adam:noise_multi
# current epoch:epoch
T = epoch*N/batch_size T = epoch*N/batch_size
return np.sqrt(np.exp(noise_multi**(-2))-1)*np.sqrt(T)*batch_size/N return np.sqrt(np.exp(noise_multi**(-2))-1)*np.sqrt(T)*batch_size/N
@ -47,6 +52,8 @@ def delta_eps_mu(eps, mu):
def eps_from_mu(mu, delta): def eps_from_mu(mu, delta):
'''Compute epsilon from mu given delta via inverse dual.''' '''Compute epsilon from mu given delta via inverse dual.'''
# Target delta:delta
def f(x): def f(x):
'''Reversely solve dual by matching delta.''' '''Reversely solve dual by matching delta.'''
return delta_eps_mu(x, mu) - delta return delta_eps_mu(x, mu) - delta
@ -54,8 +61,20 @@ def eps_from_mu(mu, delta):
def compute_eps_uniform(epoch, noise_multi, N, batch_size, delta): def compute_eps_uniform(epoch, noise_multi, N, batch_size, delta):
'''Compute epsilon given delta from inverse dual of uniform subsampling.''' '''Compute epsilon given delta from inverse dual of uniform subsampling.'''
# Total number of examples:N
# batch size:batch_size
# Noise multiplier for DP-SGD/DP-Adam:noise_multi
# current epoch:epoch
# Target delta:delta
return eps_from_mu(compute_mu_uniform(epoch, noise_multi, N, batch_size), delta) return eps_from_mu(compute_mu_uniform(epoch, noise_multi, N, batch_size), delta)
def compute_eps_poisson(epoch, noise_multi, N, batch_size, delta): def compute_eps_poisson(epoch, noise_multi, N, batch_size, delta):
'''Compute epsilon given delta from inverse dual of Poisson subsampling.''' '''Compute epsilon given delta from inverse dual of Poisson subsampling.'''
# Total number of examples:N
# batch size:batch_size
# Noise multiplier for DP-SGD/DP-Adam:noise_multi
# current epoch:epoch
# Target delta:delta
return eps_from_mu(compute_mu_poisson(epoch, noise_multi, N, batch_size), delta) return eps_from_mu(compute_mu_poisson(epoch, noise_multi, N, batch_size), delta)