def test_basic_budget_accounting():
querier = dp.DpQuerySession(db_path, privacy_budget=10 * epsilon)
assert (querier.remaining_budget == 10 *
epsilon), "Remaining budget is computed incorrectly"
querier.get_count(movie_name, rating_threshold, epsilon)
assert (querier.remaining_budget == 9 *
epsilon), "Remaining budget is computed incorrectly"
def test_multiple_queries_2():
querier = dp.DpQuerySession(db_path, privacy_budget=5 * epsilon)
values = np.zeros(5)
for i, stars in enumerate(range(1, 6)):
noisy_count = querier.get_count(movie_name,
rating_threshold=stars,
epsilon=epsilon)
values[i] = noisy_count
# Same noise should be returned for the identical queries, which are defined by the values of
# both movie_name and rating_threshold.
assert np.all(values[1:] != values[0]
), "Got the exact same response to different queries."
def test_budget_depleted():
querier = dp.DpQuerySession(db_path, privacy_budget=2 * epsilon)
querier.get_count(movie_name, 1, epsilon)
querier.get_count(movie_name, 2, epsilon)
# Budget is fully spent by now. The next query should raise BudgetDepletedError.
raised = False
try:
querier.get_count(movie_name, 3, epsilon)
except Exception as e:
raised = True
assert isinstance(
e, dp.BudgetDepletedError
), f"Expected BudgetDepletedError exception, got {type(e)}"
assert raised, "Does not raise an error when privacy budget is depleted."
def test_multiple_queries_1():
num_trials = 30
querier = dp.DpQuerySession(db_path, privacy_budget=2 * epsilon)
values = np.zeros(num_trials)
for i in range(num_trials):
noisy_count = querier.get_count(movie_name,
rating_threshold,
epsilon=epsilon)
values[i] = noisy_count
# Responses to identical queries should have the same noise. Otherwise, the attacker can
# retrieve the real value from multiple repetitions of the same query.
assert np.all(values == values[0]), (
"The attacker might be able to retrieve the exact value of a given query because the "
"noise is freshly drawn for every response.")
assert querier.remaining_budget == epsilon, (
"Remaining budget is incorrectly computed when "
"queries are repeated.")
def test_noise_distribution():
num_trials = 300
values = np.zeros(num_trials)
for i in range(num_trials):
querier = dp.DpQuerySession(db_path, privacy_budget=10 * epsilon)
noisy_count = querier.get_count(movie_name,
rating_threshold,
epsilon=epsilon)
values[i] = noisy_count
noise = values - _get_real_count(db_path, movie_name, rating_threshold)
# Check that the noise follows Laplace distribution with scale 1/epsilon
with warnings.catch_warnings():
warnings.simplefilter("ignore")
_, p_value = stats.kstest(noise, "laplace", (0, 1. / epsilon))
p_value_thresh = 0.005
assert (
not np.isnan(p_value) and p_value > p_value_thresh
), "The added noise does not seem to result in the required level of privacy."