626_privacy/tensorflow_privacy
2
0
Fork 1
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

Use calibrate_dp_mechanism from differential_privacy library instead of custom binary search.

Browse source 
PiperOrigin-RevId: 466798182
This commit is contained in:
Galen Andrew 2022-08-10 15:19:16 -07:00 committed by A. Unique TensorFlower
parent 857fe8f482
commit ca077a5b12
1 changed files with 15 additions and 28 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

43
tensorflow_privacy/privacy/analysis/compute_noise_from_budget_lib.py
View file

@ -18,19 +18,6 @@ import math
from absl import app from absl import app
import dp_accounting import dp_accounting
from scipy import optimize
def apply_dp_sgd_analysis(q, sigma, steps, orders, delta):
"""Compute and print results of DP-SGD analysis."""
accountant = dp_accounting.rdp.RdpAccountant(orders)
event = dp_accounting.SelfComposedDpEvent(
dp_accounting.PoissonSampledDpEvent(q,
dp_accounting.GaussianDpEvent(sigma)),
steps)
accountant.compose(event)
return accountant.get_epsilon_and_optimal_order(delta)
def compute_noise(n, batch_size, target_epsilon, epochs, delta, noise_lbd): def compute_noise(n, batch_size, target_epsilon, epochs, delta, noise_lbd):
@ -42,26 +29,26 @@ def compute_noise(n, batch_size, target_epsilon, epochs, delta, noise_lbd):
list(range(5, 64)) + [128, 256, 512]) list(range(5, 64)) + [128, 256, 512])
steps = int(math.ceil(epochs * n / batch_size)) steps = int(math.ceil(epochs * n / batch_size))
init_noise = noise_lbd # minimum possible noise def make_event_from_noise(sigma):
init_epsilon, _ = apply_dp_sgd_analysis(q, init_noise, steps, orders, delta) return dp_accounting.SelfComposedDpEvent(
dp_accounting.PoissonSampledDpEvent(
q, dp_accounting.GaussianDpEvent(sigma)), steps)
def make_accountant():
return dp_accounting.rdp.RdpAccountant(orders)
accountant = make_accountant()
accountant.compose(make_event_from_noise(noise_lbd))
init_epsilon = accountant.get_epsilon(delta)
if init_epsilon < target_epsilon: # noise_lbd was an overestimate if init_epsilon < target_epsilon: # noise_lbd was an overestimate
print('min_noise too large for target epsilon.') print('noise_lbd too large for target epsilon.')
return 0 return 0
cur_epsilon = init_epsilon target_noise = dp_accounting.calibrate_dp_mechanism(
max_noise, min_noise = init_noise, 0 make_accountant, make_event_from_noise, target_epsilon, delta,
dp_accounting.LowerEndpointAndGuess(noise_lbd, noise_lbd * 2))
# doubling to find the right range
while cur_epsilon > target_epsilon: # until noise is large enough
max_noise, min_noise = max_noise * 2, max_noise
cur_epsilon, _ = apply_dp_sgd_analysis(q, max_noise, steps, orders, delta)
def epsilon_fn(noise): # should return 0 if guess_epsilon==target_epsilon
guess_epsilon = apply_dp_sgd_analysis(q, noise, steps, orders, delta)[0]
return guess_epsilon - target_epsilon
target_noise = optimize.bisect(epsilon_fn, min_noise, max_noise)
print( print(
'DP-SGD with sampling rate = {:.3g}% and noise_multiplier = {} iterated' 'DP-SGD with sampling rate = {:.3g}% and noise_multiplier = {} iterated'
' over {} steps satisfies'.format(100 * q, target_noise, steps), ' over {} steps satisfies'.format(100 * q, target_noise, steps),