def end(self, session):
orders = [1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64))
samples = session.run(self._samples)
queries = session.run(self._queries)
formatted_ledger = privacy_ledger.format_ledger(samples, queries)
rdp = compute_rdp_from_ledger(formatted_ledger, orders)
eps = get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
print('For delta=1e-5, the current epsilon is: %.2f' % eps)
def end(self, session):
orders = [1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64))
samples = session.run(self._samples)
queries = session.run(self._queries)
formatted_ledger = privacy_ledger.format_ledger(samples, queries)
rdp = compute_rdp_from_ledger(formatted_ledger, orders)
eps = get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
sys.stdout.write(',%s' % eps)
sys.stdout.flush()
def test_compute_rdp_from_ledger(self):
orders = range(2, 33)
q = 0.1
n = 1000
l2_norm_clip = 3.14159
noise_stddev = 2.71828
steps = 3
query_entry = privacy_ledger.GaussianSumQueryEntry(
l2_norm_clip, noise_stddev)
ledger = [privacy_ledger.SampleEntry(n, q, [query_entry])] * steps
z = noise_stddev / l2_norm_clip
rdp = rdp_accountant.compute_rdp(q, z, steps, orders)
rdp_from_ledger = rdp_accountant.compute_rdp_from_ledger(ledger, orders)
self.assertSequenceAlmostEqual(rdp, rdp_from_ledger)
def _get_rdp_and_orders(self):
orders = [1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64))
if self.ledger is None:
steps = self.epochs * self.train_size // self.batch_size
sampling_probability = self.batch_size / self.train_size
rdp = compute_rdp(
q=sampling_probability,
noise_multiplier=self.noise_multiplier,
steps=steps,
orders=orders,
)
else:
# calculate with ledger
print("Formatted Ledger")
formatted_ledger = self.ledger.get_formatted_ledger(get_session())
rdp = compute_rdp_from_ledger(formatted_ledger, orders)
return rdp, orders
def end(self, session):
# Any RDP order (for order > 1) corresponds to one epsilon value. We
# enumerate through a few orders and pick the one that gives lowest epsilon.
# The variable orders may be extended for different use cases. Usually, the
# search is set to be finer-grained for small orders and coarser-grained for
# larger orders.
orders = [1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64))
samples = session.run(self._samples)
queries = session.run(self._queries)
formatted_ledger = privacy_ledger.format_ledger(samples, queries)
rdp = compute_rdp_from_ledger(formatted_ledger, orders)
# It is recommended that delta is o(1/dataset_size). In the case of MNIST,
# dataset_size is 60000, so we set delta to be 1e-5. For larger datasets,
# delta should be set smaller.
eps = get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
print('For delta=1e-5, the current epsilon is: %.2f' % eps)