def _compute_rdp(votes, baseline, threshold, sigma1, sigma2, delta, orders,
data_ind):
"""Computes the (data-dependent) RDP curve for Confident GNMax."""
rdp_cum = np.zeros(len(orders))
rdp_sqrd_cum = np.zeros(len(orders))
answered = 0
for i, v in enumerate(votes):
if threshold is None:
logq_step1 = 0 # No thresholding, always proceed to step 2.
rdp_step1 = np.zeros(len(orders))
else:
logq_step1 = pate.compute_logpr_answered(threshold, sigma1,
v - baseline[i, ])
if data_ind:
rdp_step1 = pate.compute_rdp_data_independent_threshold(
sigma1, orders)
else:
rdp_step1 = pate.compute_rdp_threshold(logq_step1, sigma1,
orders)
if data_ind:
rdp_step2 = pate.rdp_data_independent_gaussian(sigma2, orders)
else:
logq_step2 = pate.compute_logq_gaussian(v, sigma2)
rdp_step2 = pate.rdp_gaussian(logq_step2, sigma2, orders)
q_step1 = np.exp(logq_step1)
rdp = rdp_step1 + rdp_step2 * q_step1
# The expression below evaluates
# E[(cost_of_step_1 + Bernoulli(pr_of_step_2) * cost_of_step_2)^2]
rdp_sqrd = (rdp_step1**2 + 2 * rdp_step1 * q_step1 * rdp_step2 +
q_step1 * rdp_step2**2)
rdp_sqrd_cum += rdp_sqrd
rdp_cum += rdp
answered += q_step1
if ((i + 1) % 1000 == 0) or (i == votes.shape[0] - 1):
rdp_var = rdp_sqrd_cum / i - (rdp_cum /
i)**2 # Ignore Bessel's correction.
eps_total, order_opt = pate.compute_eps_from_delta(
orders, rdp_cum, delta)
order_opt_idx = np.searchsorted(orders, order_opt)
eps_std = ((i + 1) * rdp_var[order_opt_idx])**.5 # Std of the sum.
print(
'queries = {}, E[answered] = {:.2f}, E[eps] = {:.3f} (std = {:.5f}) '
'at order = {:.2f} (contribution from delta = {:.3f})'.format(
i + 1, answered, eps_total, eps_std, order_opt,
-math.log(delta) / (order_opt - 1)))
sys.stdout.flush()
_, order_opt = pate.compute_eps_from_delta(orders, rdp_cum, delta)
return order_opt
def _compute_rdp_list_threshold(num_teachers, threshold, sigma, order):
key = (num_teachers, threshold, sigma, order)
if key in _rdp_thresholds:
return _rdp_thresholds[key]
res = np.zeros(num_teachers + 1)
for v in range(0, num_teachers + 1):
logp = scipy.stats.norm.logsf(threshold - v, scale=sigma)
res[v] = pate.compute_rdp_threshold(logp, sigma, order)
_rdp_thresholds[key] = res
return res
def _compute_rdp_list_threshold(num_teachers, threshold, sigma, order):
key = (num_teachers, threshold, sigma, order)
if key in _rdp_thresholds:
return _rdp_thresholds[key]
res = np.zeros(num_teachers + 1)
for v in range(0, num_teachers + 1):
logp = scipy.stats.norm.logsf(threshold - v, scale=sigma)
res[v] = pate.compute_rdp_threshold(logp, sigma, order)
_rdp_thresholds[key] = res
return res
def _compute_rdp(votes, baseline, threshold, sigma1, sigma2, delta, orders,
data_ind):
"""Computes the (data-dependent) RDP curve for Confident GNMax."""
rdp_cum = np.zeros(len(orders))
rdp_sqrd_cum = np.zeros(len(orders))
answered = 0
for i, v in enumerate(votes):
if threshold is None:
logq_step1 = 0 # No thresholding, always proceed to step 2.
rdp_step1 = np.zeros(len(orders))
else:
logq_step1 = pate.compute_logpr_answered(threshold, sigma1,
v - baseline[i,])
if data_ind:
rdp_step1 = pate.compute_rdp_data_independent_threshold(sigma1, orders)
else:
rdp_step1 = pate.compute_rdp_threshold(logq_step1, sigma1, orders)
if data_ind:
rdp_step2 = pate.rdp_data_independent_gaussian(sigma2, orders)
else:
logq_step2 = pate.compute_logq_gaussian(v, sigma2)
rdp_step2 = pate.rdp_gaussian(logq_step2, sigma2, orders)
q_step1 = np.exp(logq_step1)
rdp = rdp_step1 + rdp_step2 * q_step1
# The expression below evaluates
# E[(cost_of_step_1 + Bernoulli(pr_of_step_2) * cost_of_step_2)^2]
rdp_sqrd = (
rdp_step1**2 + 2 * rdp_step1 * q_step1 * rdp_step2 +
q_step1 * rdp_step2**2)
rdp_sqrd_cum += rdp_sqrd
rdp_cum += rdp
answered += q_step1
if ((i + 1) % 1000 == 0) or (i == votes.shape[0] - 1):
rdp_var = rdp_sqrd_cum / i - (
rdp_cum / i)**2 # Ignore Bessel's correction.
eps_total, order_opt = pate.compute_eps_from_delta(orders, rdp_cum, delta)
order_opt_idx = np.searchsorted(orders, order_opt)
eps_std = ((i + 1) * rdp_var[order_opt_idx])**.5 # Std of the sum.
print(
'queries = {}, E[answered] = {:.2f}, E[eps] = {:.3f} (std = {:.5f}) '
'at order = {:.2f} (contribution from delta = {:.3f})'.format(
i + 1, answered, eps_total, eps_std, order_opt,
-math.log(delta) / (order_opt - 1)))
sys.stdout.flush()
_, order_opt = pate.compute_eps_from_delta(orders, rdp_cum, delta)
return order_opt
def analyze_gnmax_conf_data_dep(votes, threshold, sigma1, sigma2, delta):
# Short list of orders.
# orders = np.round(np.logspace(np.log10(20), np.log10(200), num=20))
# Long list of orders.
orders = np.concatenate((np.arange(20, 40, .2), np.arange(40, 75, .5),
np.logspace(np.log10(75), np.log10(200), num=20)))
n = votes.shape[0]
num_classes = votes.shape[1]
num_teachers = int(sum(votes[0, ]))
if threshold is not None and sigma1 is not None:
is_data_ind_step1 = pate.is_data_independent_always_opt_gaussian(
num_teachers, num_classes, sigma1, orders)
else:
is_data_ind_step1 = [True] * len(orders)
is_data_ind_step2 = pate.is_data_independent_always_opt_gaussian(
num_teachers, num_classes, sigma2, orders)
eps_partitioned = np.full(n, None, dtype=Partition)
order_opt = np.full(n, None, dtype=float)
ss_std_opt = np.full(n, None, dtype=float)
answered = np.zeros(n)
rdp_step1_total = np.zeros(len(orders))
rdp_step2_total = np.zeros(len(orders))
ls_total = np.zeros((len(orders), num_teachers))
answered_total = 0
for i in range(n):
v = votes[i, ]
if threshold is not None and sigma1 is not None:
logq_step1 = pate.compute_logpr_answered(threshold, sigma1, v)
rdp_step1_total += pate.compute_rdp_threshold(
logq_step1, sigma1, orders)
else:
logq_step1 = 0. # always answer
pr_answered = np.exp(logq_step1)
logq_step2 = pate.compute_logq_gaussian(v, sigma2)
rdp_step2_total += pr_answered * pate.rdp_gaussian(
logq_step2, sigma2, orders)
answered_total += pr_answered
rdp_ss = np.zeros(len(orders))
ss_std = np.zeros(len(orders))
for j, order in enumerate(orders):
if not is_data_ind_step1[j]:
ls_step1 = pate_ss.compute_local_sensitivity_bounds_threshold(
v, num_teachers, threshold, sigma1, order)
else:
ls_step1 = np.full(num_teachers, 0, dtype=float)
if not is_data_ind_step2[j]:
ls_step2 = pate_ss.compute_local_sensitivity_bounds_gnmax(
v, num_teachers, sigma2, order)
else:
ls_step2 = np.full(num_teachers, 0, dtype=float)
ls_total[j, ] += ls_step1 + pr_answered * ls_step2
beta_ss = .49 / order
ss = pate_ss.compute_discounted_max(beta_ss, ls_total[j, ])
sigma_ss = ((order * math.exp(2 * beta_ss)) / ss)**(1 / 3)
rdp_ss[j] = pate_ss.compute_rdp_of_smooth_sensitivity_gaussian(
beta_ss, sigma_ss, order)
ss_std[j] = ss * sigma_ss
rdp_total = rdp_step1_total + rdp_step2_total + rdp_ss
answered[i] = answered_total
_, order_opt[i] = pate.compute_eps_from_delta(orders, rdp_total, delta)
order_idx = np.searchsorted(orders, order_opt[i])
# Since optimal orders are always non-increasing, shrink orders array
# and all cumulative arrays to speed up computation.
if order_idx < len(orders):
orders = orders[:order_idx + 1]
rdp_step1_total = rdp_step1_total[:order_idx + 1]
rdp_step2_total = rdp_step2_total[:order_idx + 1]
eps_partitioned[i] = Partition(step1=rdp_step1_total[order_idx],
step2=rdp_step2_total[order_idx],
ss=rdp_ss[order_idx],
delta=-math.log(delta) /
(order_opt[i] - 1))
ss_std_opt[i] = ss_std[order_idx]
if i > 0 and (i + 1) % 1 == 0:
print(
'queries = {}, E[answered] = {:.2f}, E[eps] = {:.3f} +/- {:.3f} '
'at order = {:.2f}. Contributions: delta = {:.3f}, step1 = {:.3f}, '
'step2 = {:.3f}, ss = {:.3f}'.format(
i + 1, answered[i], sum(eps_partitioned[i]), ss_std_opt[i],
order_opt[i], eps_partitioned[i].delta,
eps_partitioned[i].step1, eps_partitioned[i].step2,
eps_partitioned[i].ss))
sys.stdout.flush()
return eps_partitioned, answered, ss_std_opt, order_opt
def analyze_gnmax_conf_data_dep(votes, threshold, sigma1, sigma2, delta):
# Short list of orders.
# orders = np.round(np.logspace(np.log10(20), np.log10(200), num=20))
# Long list of orders.
orders = np.concatenate((np.arange(20, 40, .2),
np.arange(40, 75, .5),
np.logspace(np.log10(75), np.log10(200), num=20)))
n = votes.shape[0]
num_classes = votes.shape[1]
num_teachers = int(sum(votes[0,]))
if threshold is not None and sigma1 is not None:
is_data_ind_step1 = pate.is_data_independent_always_opt_gaussian(
num_teachers, num_classes, sigma1, orders)
else:
is_data_ind_step1 = [True] * len(orders)
is_data_ind_step2 = pate.is_data_independent_always_opt_gaussian(
num_teachers, num_classes, sigma2, orders)
eps_partitioned = np.full(n, None, dtype=Partition)
order_opt = np.full(n, None, dtype=float)
ss_std_opt = np.full(n, None, dtype=float)
answered = np.zeros(n)
rdp_step1_total = np.zeros(len(orders))
rdp_step2_total = np.zeros(len(orders))
ls_total = np.zeros((len(orders), num_teachers))
answered_total = 0
for i in range(n):
v = votes[i,]
if threshold is not None and sigma1 is not None:
logq_step1 = pate.compute_logpr_answered(threshold, sigma1, v)
rdp_step1_total += pate.compute_rdp_threshold(logq_step1, sigma1, orders)
else:
logq_step1 = 0. # always answer
pr_answered = np.exp(logq_step1)
logq_step2 = pate.compute_logq_gaussian(v, sigma2)
rdp_step2_total += pr_answered * pate.rdp_gaussian(logq_step2, sigma2,
orders)
answered_total += pr_answered
rdp_ss = np.zeros(len(orders))
ss_std = np.zeros(len(orders))
for j, order in enumerate(orders):
if not is_data_ind_step1[j]:
ls_step1 = pate_ss.compute_local_sensitivity_bounds_threshold(v,
num_teachers, threshold, sigma1, order)
else:
ls_step1 = np.full(num_teachers, 0, dtype=float)
if not is_data_ind_step2[j]:
ls_step2 = pate_ss.compute_local_sensitivity_bounds_gnmax(
v, num_teachers, sigma2, order)
else:
ls_step2 = np.full(num_teachers, 0, dtype=float)
ls_total[j,] += ls_step1 + pr_answered * ls_step2
beta_ss = .49 / order
ss = pate_ss.compute_discounted_max(beta_ss, ls_total[j,])
sigma_ss = ((order * math.exp(2 * beta_ss)) / ss) ** (1 / 3)
rdp_ss[j] = pate_ss.compute_rdp_of_smooth_sensitivity_gaussian(
beta_ss, sigma_ss, order)
ss_std[j] = ss * sigma_ss
rdp_total = rdp_step1_total + rdp_step2_total + rdp_ss
answered[i] = answered_total
_, order_opt[i] = pate.compute_eps_from_delta(orders, rdp_total, delta)
order_idx = np.searchsorted(orders, order_opt[i])
# Since optimal orders are always non-increasing, shrink orders array
# and all cumulative arrays to speed up computation.
if order_idx < len(orders):
orders = orders[:order_idx + 1]
rdp_step1_total = rdp_step1_total[:order_idx + 1]
rdp_step2_total = rdp_step2_total[:order_idx + 1]
eps_partitioned[i] = Partition(step1=rdp_step1_total[order_idx],
step2=rdp_step2_total[order_idx],
ss=rdp_ss[order_idx],
delta=-math.log(delta) / (order_opt[i] - 1))
ss_std_opt[i] = ss_std[order_idx]
if i > 0 and (i + 1) % 1 == 0:
print('queries = {}, E[answered] = {:.2f}, E[eps] = {:.3f} +/- {:.3f} '
'at order = {:.2f}. Contributions: delta = {:.3f}, step1 = {:.3f}, '
'step2 = {:.3f}, ss = {:.3f}'.format(
i + 1,
answered[i],
sum(eps_partitioned[i]),
ss_std_opt[i],
order_opt[i],
eps_partitioned[i].delta,
eps_partitioned[i].step1,
eps_partitioned[i].step2,
eps_partitioned[i].ss))
sys.stdout.flush()
return eps_partitioned, answered, ss_std_opt, order_opt
def _find_optimal_smooth_sensitivity_parameters(votes, baseline, num_teachers,
threshold, sigma1, sigma2,
delta, ind_step1, ind_step2,
order):
"""Optimizes smooth sensitivity parameters by minimizing a cost function.
The cost function is
exact_eps + cost of GNSS + two stds of noise,
which captures that upper bound of the confidence interval of the sanitized
privacy budget.
Since optimization is done with full view of sensitive data, the results
cannot be released.
"""
rdp_cum = 0
answered_cum = 0
ls_cum = 0
# Define a plausible range for the beta values.
betas = np.arange(.3 / order, .495 / order, .01 / order)
cost_delta = math.log(1 / delta) / (order - 1)
for i, v in enumerate(votes):
if threshold is None:
log_pr_answered = 0
rdp1 = 0
ls_step1 = np.zeros(num_teachers)
else:
log_pr_answered = pate.compute_logpr_answered(
threshold, sigma1, v - baseline[i, ])
if ind_step1: # apply data-independent bound for step 1 (thresholding).
rdp1 = pate.compute_rdp_data_independent_threshold(
sigma1, order)
ls_step1 = np.zeros(num_teachers)
else:
rdp1 = pate.compute_rdp_threshold(log_pr_answered, sigma1,
order)
ls_step1 = pate_ss.compute_local_sensitivity_bounds_threshold(
v - baseline[i, ], num_teachers, threshold, sigma1, order)
pr_answered = math.exp(log_pr_answered)
answered_cum += pr_answered
if ind_step2: # apply data-independent bound for step 2 (GNMax).
rdp2 = pate.rdp_data_independent_gaussian(sigma2, order)
ls_step2 = np.zeros(num_teachers)
else:
logq_step2 = pate.compute_logq_gaussian(v, sigma2)
rdp2 = pate.rdp_gaussian(logq_step2, sigma2, order)
# Compute smooth sensitivity.
ls_step2 = pate_ss.compute_local_sensitivity_bounds_gnmax(
v, num_teachers, sigma2, order)
rdp_cum += rdp1 + pr_answered * rdp2
ls_cum += ls_step1 + pr_answered * ls_step2 # Expected local sensitivity.
if ind_step1 and ind_step2:
# Data-independent bounds.
cost_opt, beta_opt, ss_opt, sigma_ss_opt = None, 0., 0., np.inf
else:
# Data-dependent bounds.
cost_opt, beta_opt, ss_opt, sigma_ss_opt = np.inf, None, None, None
for beta in betas:
ss = pate_ss.compute_discounted_max(beta, ls_cum)
# Solution to the minimization problem:
# min_sigma {order * exp(2 * beta)/ sigma^2 + 2 * ss * sigma}
sigma_ss = ((order * math.exp(2 * beta)) / ss)**(1 / 3)
cost_ss = pate_ss.compute_rdp_of_smooth_sensitivity_gaussian(
beta, sigma_ss, order)
# Cost captures exact_eps + cost of releasing SS + two stds of noise.
cost = rdp_cum + cost_ss + 2 * ss * sigma_ss
if cost < cost_opt:
cost_opt, beta_opt, ss_opt, sigma_ss_opt = cost, beta, ss, sigma_ss
if ((i + 1) % 100 == 0) or (i == votes.shape[0] - 1):
eps_before_ss = rdp_cum + cost_delta
eps_with_ss = (eps_before_ss +
pate_ss.compute_rdp_of_smooth_sensitivity_gaussian(
beta_opt, sigma_ss_opt, order))
print(
'{}: E[answered queries] = {:.1f}, RDP at {} goes from {:.3f} to '
'{:.3f} +/- {:.3f} (ss = {:.4}, beta = {:.4f}, sigma_ss = {:.3f})'
.format(i + 1, answered_cum, order, eps_before_ss, eps_with_ss,
ss_opt * sigma_ss_opt, ss_opt, beta_opt, sigma_ss_opt))
sys.stdout.flush()
# Return optimal parameters for the last iteration.
return beta_opt, ss_opt, sigma_ss_opt
def _find_optimal_smooth_sensitivity_parameters(
votes, baseline, num_teachers, threshold, sigma1, sigma2, delta, ind_step1,
ind_step2, order):
"""Optimizes smooth sensitivity parameters by minimizing a cost function.
The cost function is
exact_eps + cost of GNSS + two stds of noise,
which captures that upper bound of the confidence interval of the sanitized
privacy budget.
Since optimization is done with full view of sensitive data, the results
cannot be released.
"""
rdp_cum = 0
answered_cum = 0
ls_cum = 0
# Define a plausible range for the beta values.
betas = np.arange(.3 / order, .495 / order, .01 / order)
cost_delta = math.log(1 / delta) / (order - 1)
for i, v in enumerate(votes):
if threshold is None:
log_pr_answered = 0
rdp1 = 0
ls_step1 = np.zeros(num_teachers)
else:
log_pr_answered = pate.compute_logpr_answered(threshold, sigma1,
v - baseline[i,])
if ind_step1: # apply data-independent bound for step 1 (thresholding).
rdp1 = pate.compute_rdp_data_independent_threshold(sigma1, order)
ls_step1 = np.zeros(num_teachers)
else:
rdp1 = pate.compute_rdp_threshold(log_pr_answered, sigma1, order)
ls_step1 = pate_ss.compute_local_sensitivity_bounds_threshold(
v - baseline[i,], num_teachers, threshold, sigma1, order)
pr_answered = math.exp(log_pr_answered)
answered_cum += pr_answered
if ind_step2: # apply data-independent bound for step 2 (GNMax).
rdp2 = pate.rdp_data_independent_gaussian(sigma2, order)
ls_step2 = np.zeros(num_teachers)
else:
logq_step2 = pate.compute_logq_gaussian(v, sigma2)
rdp2 = pate.rdp_gaussian(logq_step2, sigma2, order)
# Compute smooth sensitivity.
ls_step2 = pate_ss.compute_local_sensitivity_bounds_gnmax(
v, num_teachers, sigma2, order)
rdp_cum += rdp1 + pr_answered * rdp2
ls_cum += ls_step1 + pr_answered * ls_step2 # Expected local sensitivity.
if ind_step1 and ind_step2:
# Data-independent bounds.
cost_opt, beta_opt, ss_opt, sigma_ss_opt = None, 0., 0., np.inf
else:
# Data-dependent bounds.
cost_opt, beta_opt, ss_opt, sigma_ss_opt = np.inf, None, None, None
for beta in betas:
ss = pate_ss.compute_discounted_max(beta, ls_cum)
# Solution to the minimization problem:
# min_sigma {order * exp(2 * beta)/ sigma^2 + 2 * ss * sigma}
sigma_ss = ((order * math.exp(2 * beta)) / ss)**(1 / 3)
cost_ss = pate_ss.compute_rdp_of_smooth_sensitivity_gaussian(
beta, sigma_ss, order)
# Cost captures exact_eps + cost of releasing SS + two stds of noise.
cost = rdp_cum + cost_ss + 2 * ss * sigma_ss
if cost < cost_opt:
cost_opt, beta_opt, ss_opt, sigma_ss_opt = cost, beta, ss, sigma_ss
if ((i + 1) % 100 == 0) or (i == votes.shape[0] - 1):
eps_before_ss = rdp_cum + cost_delta
eps_with_ss = (
eps_before_ss + pate_ss.compute_rdp_of_smooth_sensitivity_gaussian(
beta_opt, sigma_ss_opt, order))
print('{}: E[answered queries] = {:.1f}, RDP at {} goes from {:.3f} to '
'{:.3f} +/- {:.3f} (ss = {:.4}, beta = {:.4f}, sigma_ss = {:.3f})'.
format(i + 1, answered_cum, order, eps_before_ss, eps_with_ss,
ss_opt * sigma_ss_opt, ss_opt, beta_opt, sigma_ss_opt))
sys.stdout.flush()
# Return optimal parameters for the last iteration.
return beta_opt, ss_opt, sigma_ss_opt