def get_preprocessor(self):
"""Preprocessor for DP Sum (float)"""
if self.has_error():
return None
# Have we already already assembled it?
#
if self.preprocessor is not None:
# Yes!
return self.preprocessor
# Build the preprocessor!
#
preprocessor = (
make_select_column(self.col_index, TOA=str) >> make_cast(TIA=str,
TOA=float)
>> make_impute_constant(constant=self.fixed_value) >>
make_clamp(bounds=self.get_bounds()) >> make_bounded_resize(
size=self.dataset_size,
bounds=self.get_bounds(),
constant=self.fixed_value) >> make_sized_bounded_sum(
size=self.dataset_size, bounds=self.get_bounds()))
self.scale = binary_search(
lambda s: self.check_scale(s, preprocessor, 1, self.epsilon),
bounds=(0.0, 1000.0))
preprocessor = preprocessor >> make_base_laplace(self.scale)
# keep a pointer to the preprocessor to re-use for .run_chain(...)
self.preprocessor = preprocessor
return preprocessor
def _compute_noise_scale(self):
if self.scale is not None:
return
lower = self.lower
upper = self.upper
max_contrib = self.max_contrib
# should probably just check and throw if not int
bounds = (int(lower), int(upper))
enable_features('floating-point', 'contrib')
bounded_sum = (
make_clamp(bounds=bounds) >> make_bounded_sum(bounds=bounds))
discovered_scale = binary_search_param(
lambda s: bounded_sum >> make_base_geometric(scale=s),
d_in=max_contrib,
d_out=(self.epsilon))
self.scale = discovered_scale
def _compute_noise_scale(self):
if self.scale is not None:
return
lower = self.lower
upper = self.upper
max_contrib = self.max_contrib
bounds = (float(lower), float(upper))
sensitivity = upper - lower
if sensitivity > 1000:
self.scale = (float(sensitivity) * max_contrib) / self.epsilon
else:
enable_features('floating-point', 'contrib')
bounded_sum = (
make_clamp(bounds=bounds) >> make_bounded_sum(bounds=bounds))
discovered_scale = binary_search_param(
lambda s: bounded_sum >> make_base_laplace(scale=s),
d_in=max_contrib,
d_out=(self.epsilon))
self.scale = discovered_scale
def __init__(self,
data,
output_info,
log_frequency,
*ignore,
per_column_epsilon=None,
discrete_column_category_prob=None,
**kwargs):
self._data = data
self._per_column_epsilon = per_column_epsilon
if per_column_epsilon:
if per_column_epsilon <= 0.0:
raise ValueError("per_column_epsilon must be positive")
bounds = (0, 1)
max_contrib = 1
enable_features('contrib')
bounded_sum = (
make_clamp(bounds=bounds) >> make_bounded_sum(bounds=bounds))
discovered_scale = binary_search_param(
lambda s: bounded_sum >> make_base_geometric(scale=s),
d_in=max_contrib,
d_out=(self._per_column_epsilon))
self._per_column_scale = discovered_scale
else:
self._per_column_scale = None
if discrete_column_category_prob is None:
warnings.warn(
"per_column_epsilon is not set, and no cached probabilites have been provided. "
"Sampler will not privatize frequencies, which may cause privacy leaks"
)
def is_discrete_column(column_info):
return (len(column_info) == 1
and column_info[0].activation_fn == "softmax")
n_discrete_columns = sum([
1 for column_info in output_info if is_discrete_column(column_info)
])
self._discrete_column_matrix_st = np.zeros(n_discrete_columns,
dtype="int32")
# Store the row id for each category in each discrete column.
# For example _rid_by_cat_cols[a][b] is a list of all rows with the
# a-th discrete column equal value b.
self._rid_by_cat_cols = []
# Compute _rid_by_cat_cols
st = 0
for column_info in output_info:
if is_discrete_column(column_info):
span_info = column_info[0]
ed = st + span_info.dim
rid_by_cat = []
for j in range(span_info.dim):
rid_by_cat.append(np.nonzero(data[:, st + j])[0])
self._rid_by_cat_cols.append(rid_by_cat)
st = ed
else:
st += sum([span_info.dim for span_info in column_info])
assert st == data.shape[1]
# Prepare an interval matrix for efficiently sample conditional vector
max_category = max([
column_info[0].dim
for column_info in output_info if is_discrete_column(column_info)
],
default=0)
self._discrete_column_cond_st = np.zeros(n_discrete_columns,
dtype='int32')
self._discrete_column_n_category = np.zeros(n_discrete_columns,
dtype='int32')
self._discrete_column_category_prob = np.zeros(
(n_discrete_columns, max_category))
self._n_discrete_columns = n_discrete_columns
self._n_categories = sum([
column_info[0].dim for column_info in output_info
if is_discrete_column(column_info)
])
eps_tot = 0.0
st = 0
current_id = 0
current_cond_st = 0
for column_info in output_info:
if is_discrete_column(column_info):
span_info = column_info[0]
ed = st + span_info.dim
category_freq = np.sum(data[:, st:ed], axis=0)
# insert privacy here
if self._per_column_scale:
geom = make_base_geometric(self._per_column_scale)
category_freq = [geom(int(v)) for v in category_freq]
eps_tot += self._per_column_epsilon
category_freq = [1 if v < 1 else v for v in category_freq]
if np.sum(category_freq) < 100:
# not enough data; use uniform distribution
category_freq = [1 for _ in category_freq]
category_freq = np.array(category_freq, dtype='float64')
if log_frequency:
category_freq = np.log(category_freq + 1)
category_prob = category_freq / np.sum(category_freq)
self._discrete_column_category_prob[
current_id, :span_info.dim] = (category_prob)
self._discrete_column_cond_st[current_id] = current_cond_st
self._discrete_column_n_category[current_id] = span_info.dim
current_cond_st += span_info.dim
current_id += 1
st = ed
else:
st += sum([span_info.dim for span_info in column_info])
self.total_spent = eps_tot
if discrete_column_category_prob is not None:
assert len(discrete_column_category_prob) == n_discrete_columns
for i in range(n_discrete_columns):
self._discrete_column_category_prob[
i, :] = discrete_column_category_prob[i]
self.total_spent = 0.0 # don't have to pay for cached noise