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

harmonize analysis parameters with current DPSGD API

Browse source 
PiperOrigin-RevId: 255080643
This commit is contained in:
Nicolas Papernot 2019-06-25 17:05:41 -07:00 committed by A. Unique TensorFlower
parent 45bcb3a0e4
commit 6171474465
1 changed files with 4 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
privacy/analysis/compute_dp_sgd_privacy.py
View file

@ -61,6 +61,10 @@ flags.mark_flag_as_required('epochs')
def apply_dp_sgd_analysis(q, sigma, steps, orders, delta):
"""Compute and print results of DP-SGD analysis."""
# compute_rdp requires that sigma be the ratio of the standard deviation of
# the Gaussian noise to the l2-sensitivity of the function to which it is
# added. Hence, sigma here corresponds to the `noise_multiplier` parameter
# in the DP-SGD implementation found in privacy.optimizers.dp_optimizer
rdp = compute_rdp(q, sigma, steps, orders)
eps, _, opt_order = get_privacy_spent(orders, rdp, target_delta=delta)
@ -80,13 +84,10 @@ def main(argv):
del argv # argv is not used.
q = FLAGS.batch_size / FLAGS.N # q - the sampling ratio.
if q > 1:
raise app.UsageError('N must be larger than the batch size.')
orders = ([1.25, 1.5, 1.75, 2., 2.25, 2.5, 3., 3.5, 4., 4.5] +
list(range(5, 64)) + [128, 256, 512])
steps = int(math.ceil(FLAGS.epochs * FLAGS.N / FLAGS.batch_size))
apply_dp_sgd_analysis(q, FLAGS.noise_multiplier, steps, orders, FLAGS.delta)