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

address schien comments

Browse source
This commit is contained in:
Matthew Jagielski 2021-01-19 13:16:55 -05:00
parent 3bf78f46fe
commit e468af41dd
3 changed files with 15 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
tensorflow_privacy/privacy/analysis/compute_noise_from_budget.py
View file

@ -18,12 +18,13 @@ The script applies the RDP accountant to estimate privacy budget of an iterated
Sampled Gaussian Mechanism. The mechanism's parameters are controlled by flags. Sampled Gaussian Mechanism. The mechanism's parameters are controlled by flags.
Example: Example:
compute_dp_sgd_privacy compute_noise_from_budget
--N=60000 \ --N=60000 \
--batch_size=256 \ --batch_size=256 \
--epsilon=2.92 \ --epsilon=2.92 \
--epochs=60 \ --epochs=60 \
--delta=1e-5 --delta=1e-5 \
--min_noise=1e-6
The output states that DP-SGD with these parameters should The output states that DP-SGD with these parameters should
use a noise multiplier of 1.12. use a noise multiplier of 1.12.
@ -50,6 +51,7 @@ flags.DEFINE_integer('batch_size', None, 'Batch size')
flags.DEFINE_float('epsilon', None, 'Target epsilon for DP-SGD') flags.DEFINE_float('epsilon', None, 'Target epsilon for DP-SGD')
flags.DEFINE_float('epochs', None, 'Number of epochs (may be fractional)') flags.DEFINE_float('epochs', None, 'Number of epochs (may be fractional)')
flags.DEFINE_float('delta', 1e-6, 'Target delta') flags.DEFINE_float('delta', 1e-6, 'Target delta')
flags.DEFINE_float('min_noise', 1e-5, 'Minimum noise level for search.')
def main(argv): def main(argv):
@ -60,7 +62,7 @@ def main(argv):
assert FLAGS.epsilon is not None, 'Flag epsilon is missing.' assert FLAGS.epsilon is not None, 'Flag epsilon is missing.'
assert FLAGS.epochs is not None, 'Flag epochs is missing.' assert FLAGS.epochs is not None, 'Flag epochs is missing.'
compute_noise(FLAGS.N, FLAGS.batch_size, FLAGS.epsilon, compute_noise(FLAGS.N, FLAGS.batch_size, FLAGS.epsilon,
FLAGS.epochs, FLAGS.delta) FLAGS.epochs, FLAGS.delta, FLAGS.min_noise)
if __name__ == '__main__': if __name__ == '__main__':

7
tensorflow_privacy/privacy/analysis/compute_noise_from_budget_lib.py
View file

@ -45,7 +45,7 @@ def apply_dp_sgd_analysis(q, sigma, steps, orders, delta):
return eps, opt_order return eps, opt_order
def compute_noise(n, batch_size, target_epsilon, epochs, delta): def compute_noise(n, batch_size, target_epsilon, epochs, delta, noise_lbd):
"""Compute noise based on the given hyperparameters.""" """Compute noise based on the given hyperparameters."""
q = batch_size / n # q - the sampling ratio. q = batch_size / n # q - the sampling ratio.
if q > 1: if q > 1:
@ -54,10 +54,11 @@ def compute_noise(n, batch_size, target_epsilon, epochs, delta):
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 = 1e-5 # minimum possible noise init_noise = noise_lbd # minimum possible noise
init_epsilon, _ = apply_dp_sgd_analysis(q, init_noise, steps, orders, delta) init_epsilon, _ = apply_dp_sgd_analysis(q, init_noise, steps, orders, delta)
if init_epsilon < target_epsilon: # 1e-5 was an overestimate if init_epsilon < target_epsilon: # noise_lbd was an overestimate
print("min_noise too large for target epsilon.")
return 0 return 0
cur_epsilon = init_epsilon cur_epsilon = init_epsilon

11
tensorflow_privacy/privacy/analysis/compute_noise_from_budget_test.py
View file

@ -26,14 +26,15 @@ from tensorflow_privacy.privacy.analysis import compute_noise_from_budget_lib
class ComputeNoiseFromBudgetTest(parameterized.TestCase): class ComputeNoiseFromBudgetTest(parameterized.TestCase):
@parameterized.named_parameters( @parameterized.named_parameters(
('Test0', 60000, 150, 0.941870567, 15, 1e-5, 1.3), ('Test0', 60000, 150, 0.941870567, 15, 1e-5, 1e-5, 1.3),
('Test1', 100000, 100, 1.70928734, 30, 1e-7, 1.0), ('Test1', 100000, 100, 1.70928734, 30, 1e-7, 1e-6, 1.0),
('Test2', 100000000, 1024, 5907984.81339406, 10, 1e-7, 0.1), ('Test2', 100000000, 1024, 5907984.81339406, 10, 1e-7, 1e-5, 0.1),
('Test3', 100000000, 1024, 5907984.81339406, 10, 1e-7, 1, 0),
) )
def test_compute_noise(self, n, batch_size, target_epsilon, epochs, def test_compute_noise(self, n, batch_size, target_epsilon, epochs,
delta, expected_noise): delta, min_noise, expected_noise):
target_noise = compute_noise_from_budget_lib.compute_noise( target_noise = compute_noise_from_budget_lib.compute_noise(
n, batch_size, target_epsilon, epochs, delta) n, batch_size, target_epsilon, epochs, delta, min_noise)
self.assertAlmostEqual(target_noise, expected_noise) self.assertAlmostEqual(target_noise, expected_noise)
if __name__ == '__main__': if __name__ == '__main__':