def numeric_sparse(data, query_list, T, c, epsilon, delta):
if delta == 0:
epsilon_1 = 8 * epsilon / 9
epsilon_2 = 2 * epsilon / 9
def sigma(eps):
return 2 * c / eps
else:
epsilon_1 = np.sqrt(512) * epsilon / (np.sqrt(512) + 1)
epsilon_2 = 2 * epsilon / (np.sqrt(512) + 1)
def sigma(eps):
return np.sqrt(32 * c * np.log(2 / delta)) / eps
T_hat_count = T + random.laplace(scale=sigma(epsilon_1))
count = 0
answer_list = []
for query in query_list:
v_query_1 = random.laplace(scale=2 * sigma(epsilon_1))
query_response = data.evaluate(query)
if query_response + v_query_1 >= T_hat_count:
v_query_2 = random.laplace(scale=sigma(epsilon_2))
answer_list.append(query_response + v_query_2)
count = count + 1
T_hat_count = T + random.laplace(scale=sigma(epsilon_1))
else:
answer_list.append(STOP_SIGN)
if count >= c:
return answer_list
return answer_list
def find(data, domain, goal_number, failure, eps, sparse=True):
"""
Based on "Locating a Small Cluster Privately" by Kobbi Nissim, Uri Stemmer, and Salil Vadhan. PODS 2016.
Given a data set, finds the radius of an approximately minimal cluster of points with
approximately the desired amount of points
:param data: list of points in R^dimension
:param domain: tuple(absolute value of domain's end as int, minimum intervals in domain as float)
:param goal_number: the number of desired points in the resulting cluster
:param failure: 0 < float < 1. chances that the procedure will fail to return an answer
:param eps: float > 0. privacy parameter
:param sparse: 1 > float > 0. privacy parameter
:return: the radius of the resulting cluster
"""
# max(abs(np.min(data)), np.max(data))
all_distances = distances(data)
# TODO change variable name
# 'a' need to greater than - log(domain[0] / failure) / eps
a = 2 * log(domain[0] / failure) / eps
thresh = goal_number - a - log(1 / failure) / eps
# TODO verify that the noise addition is correct
if __max_average_ball__(0, all_distances, goal_number) + laplace(0, 1 / eps, 1) > thresh:
return 0
dimension = data.shape[1]
# TODO maybe a little less sparse?
if sparse:
new_domain = __sparse_domain__(domain, dimension)
else:
new_domain = __create_regular_domain__(domain, dimension)
def quality(d, r):
return min(goal_number - __max_average_ball__(r / 2, all_distances, goal_number),
__max_average_ball__(r, all_distances, goal_number) - goal_number + 2*a) / 2
return exponential_mechanism_big(data, new_domain, quality, eps / 2)
def sanitize(samples, alpha, beta, eps, delta):
"""
:param samples:
:param alpha:
:param beta:
:param eps:
:param delta:
:return:
"""
max_sample = max(samples)
dim = ceil(log2(max_sample + 1))
end_domain = 2**int(dim)
remaining_samples = set(range(end_domain))
est = dict.fromkeys(remaining_samples, 0)
new_beta = alpha * beta / 4
new_eps = eps / sqrt(32 * log(5/delta) / alpha)
new_delta = alpha * delta / 5
for i in range(int(2/alpha)):
q = __point_choice_quality__(remaining_samples)
b = choosing_mechanism_big(samples, remaining_samples, q, 1, alpha/2, new_beta, new_eps, new_delta)
if b != 'bottom':
remaining_samples.remove(b)
# remaining_samples[b] -= 1
# remaining_samples += Counter()
est[b] = concept_query(samples, point_concept(b)) + laplace(0, 1 / eps / len(samples), 1)[0]
return est
def sanitize(samples, alpha, beta, eps, delta):
"""
:param samples:
:param alpha:
:param beta:
:param eps:
:param delta:
:return:
"""
max_sample = max(samples)
dim = ceil(log2(max_sample + 1))
end_domain = 2 ** int(dim)
remaining_samples = set(range(end_domain))
est = dict.fromkeys(remaining_samples, 0)
new_beta = alpha * beta / 4
new_eps = eps / sqrt(32 * log(5 / delta) / alpha)
new_delta = alpha * delta / 5
for i in range(int(2 / alpha)):
q = __point_choice_quality__(remaining_samples)
b = choosing_mechanism_big(samples, remaining_samples, q, 1, alpha / 2, new_beta, new_eps, new_delta)
if b != "bottom":
remaining_samples.remove(b)
# remaining_samples[b] -= 1
# remaining_samples += Counter()
est[b] = concept_query(samples, point_concept(b)) + laplace(0, 1 / eps / len(samples), 1)[0]
return est
def get_noisy_distribution_of_attributes(attributes,
encoded_dataset,
epsilon=0.1):
data = encoded_dataset.copy().loc[:, attributes]
data['count'] = 1
stats = data.groupby(attributes).sum()
iterables = [
range(int(encoded_dataset[attr].max()) + 1) for attr in attributes
]
full_space = DataFrame(columns=attributes, data=list(product(*iterables)))
stats.reset_index(inplace=True)
stats = merge(full_space, stats, how='left')
stats.fillna(0, inplace=True)
if epsilon:
k = len(attributes) - 1
num_tuples, num_attributes = encoded_dataset.shape
noise_para = laplace_noise_parameter(k, num_attributes, num_tuples,
epsilon)
laplace_noises = laplace(0, scale=noise_para, size=stats.index.size)
stats['count'] += laplace_noises
stats.loc[stats['count'] < 0, 'count'] = 0
return stats
def _get_attribute_frequency_counts(self, attributes, encoded_dataset):
""" Differentially private mechanism to get attribute frequency counts"""
# Get attribute counts for category combinations present in data
counts = encoded_dataset.groupby(attributes).size()
counts.name = 'count'
counts = counts.reset_index()
# Get all possible attribute combinations
attr_combs = [
range(self.DataDescriber.attr_dict[attr].domain_size)
for attr in attributes
]
full_space = DataFrame(columns=attributes,
data=list(product(*attr_combs)))
full_counts = merge(full_space, counts, how='left')
full_counts.fillna(0, inplace=True)
# Get Laplace noise sample
noise_sample = laplace(0,
scale=self.laplace_noise_scale,
size=full_counts.index.size)
full_counts['count'] += noise_sample
full_counts.loc[full_counts['count'] < 0, 'count'] = 0
return full_counts
def noise_iid(size, noise_scale, noise_tail="normal", **kwargs):
if noise_tail == "normal":
return npr.normal(0, 1, size=size) * noise_scale
elif noise_tail == "laplace":
return npr.laplace(0, 1, size=size) * noise_scale
elif noise_tail == "cauchy":
return npr.standard_cauchy(size=size) * noise_scale
def randomize(self, oh_data, private=True):
""" Add Laplace noise. """
for row in oh_data:
for i in range(len(row)):
b = laplace(scale=1.0/self.epsilon) if private else 0
row[i] += b
return oh_data
def histograms(data, dimension, shift, side, eps, delta):
"""
Based on Theorem 2.5 from "Locating a Small Cluster Privately"
by Kobbi Nissim, Uri Stemmer, and Salil Vadhan. PODS 2016.
find parts of R^d that contain a lot of data-points
when partitioning R^d into boxes of the same size
the boxes partitioning is given by the shift and the size of the 'boxes'
:param data: list of points in R^dimension
:param dimension: the dimension of the space which the points are taken from
:param shift: the partition's shift, the i-th value represents the shift in the i-th axis
:param side: the side-length of each 'box' in the partition
:param eps: privacy parameter
:param delta: privacy parameter
:return: parts of the partition that contain a lot of data-points
"""
my_box = partial(__box_containing_point__, partition=shift, dimension=dimension, side_length=side)
# those are the parts of the partition that have at least one point
# when each box appears as many times as the numbers of points in it
boxes = [my_box(point) for point in data]
boxes_quality = Counter(boxes)
non_zero = False
for b in boxes_quality:
boxes_quality[b] += laplace(0, 2/eps, 1)[0]
if boxes_quality[b] < 2*np.log(2/delta)/eps:
boxes_quality[b] = 0
# the current boxes_quality won't be '0' so the process can return an answer
elif not non_zero:
non_zero = True
if not non_zero:
raise ValueError('No high quality box')
return max(boxes_quality)
def variance_operation(elist, elist_sq, auths, dp):
G = EcGroup(nid=conf.EC_GROUP)
#E(x^2) = (S(ri^2)/N)
plain_sum_sqs = list_sum_decryption(elist_sq, auths)
#print "plain_sum_sqs: " + str(plain_sum_sqs)
first = plain_sum_sqs/len(elist)
#E(x)^2 = (S(ri)/N)^2
esum = Classes.Ct.sum(elist)
plain_sum = collective_decryption(esum, auths)
tmp = float(plain_sum) / float(len(elist))
second = tmp * tmp
#print "second: " + str(second)
variance = first - second
if (conf.DP and dp):
#deltaf = 1/relays * 80 (80 is an arbitrary value for our.)
d = float(200)/float(len(elist))
from numpy.random import laplace
scale = float(d) / conf.EPSILON
noise = laplace(0, scale)
print "noise " + str(noise)
noisy_var = variance + noise
else:
noisy_var=variance
return str(noisy_var)
def median_operation(sk_sum, auths, dp):
from numpy.random import laplace
proto = Classes.get_median(sk_sum, min_b = 0, max_b = 1000, steps = 20) #Compute Median
plain = None
total_noise = 0
while True:
v = proto.send(plain)
if isinstance(v, int):
break
plain = collective_decryption(v, auths)
if (conf.DP and dp):
#print "sksum:" + str(sk_sum.epsilon)
if sk_sum.epsilon != 0 and sk_sum.delta != 0:
noise = 0
#print conf.DP and dp
scale = float(sk_sum.d) / float(sk_sum.epsilon)
#print sk_sum.d
#print sk_sum.epsilon
noise = int(round(laplace(0, scale)))
#print "noise: " + str(noise)
plain += noise
total_noise += noise
#print "*: " + str(plain)
#print "Estimated median: " + str(v)
#print "Total Noise Added: " + str(total_noise)
return str(v)
def slaplace(nSources, nSamples):
'''
Generates an nSources x nSamples array of sources which are Laplace distributed.
p(x) = (1/sqrt(2))*exp(-sqrt(2)*x)
'''
s = laplace(size=(nSources, nSamples))
return s / s.std(axis=1)[:, newaxis]
def slaplace(nSources,nSamples):
'''
Generates an nSources x nSamples array of sources which are Laplace distributed.
p(x) = (1/sqrt(2))*exp(-sqrt(2)*x)
'''
s = laplace(size=(nSources,nSamples))
return s/s.std(axis=1)[:,newaxis]
def get_noise(yDim, xDim, white_noise_func_val):
if white_noise_func_val == "Uniform":
return rnd.random(size=(yDim, xDim))
elif white_noise_func_val == "Beta":
return rnd.beta(2, 2, size=(yDim, xDim))
elif white_noise_func_val == "Normal":
return rnd.normal(size=(yDim, xDim))
elif white_noise_func_val == "Exponential":
return rnd.exponential(size=(yDim, xDim))
else:
return rnd.laplace(size=(yDim, xDim))
def get_cropping_params(bb_loc_laplace_b_param=0.2,
bb_size_laplace_b_param=0.06,
bbox_shrinkage_limit=0.6,
bbox_expansion_limit=1.4):
'''Get random draws for parameters involved in random crops. Output those params
Args:
----
bb_loc_laplace_b_param: Scale parameter in the laplace distribution that
determines where to center the hypothetical bounding box used to
create crop. Applied for draws of both x and y shifts.
bb_size_laplace_b_param" Scale parameter in the laplace distribution that
determines the size of the hypothetical bounding box used to create
crop. Applied for draws of both x and y scaling.
bbox_shrinkage_limit: The minimum size in each dimension of hypothetical
bounding box for cropping, as a fraction of previous bounding box
size in that dimension
bbox_expansion_limit: Maximum size in each dimension of hypothetical bounding
box for cropping, as a fraction of previous bounding box size in that
dimension.
Output:
------
x_center_shift: fraction of box width to shift center of bbox for next crop
y_center_shift: fraction of box height to shift center of bbox for next crop
x_size_shift: bbox width used to make crop, as frac of previous bbox width
y_size_shift: bbox height used to make crop, as frac of previous bbox height
'''
x_center_shift = laplace(scale=bb_loc_laplace_b_param) #delta_x in paper
y_center_shift = laplace(scale=bb_loc_laplace_b_param) #delta_y in paper
x_size_shift = (1 - laplace(scale=bb_size_laplace_b_param))
x_size_shift = np.clip(x_size_shift, bbox_shrinkage_limit,
bbox_expansion_limit)
y_size_shift = (1 - laplace(scale=bb_size_laplace_b_param))
y_size_shift = np.clip(y_size_shift, bbox_shrinkage_limit,
bbox_expansion_limit)
return x_center_shift, y_center_shift, x_size_shift, y_size_shift
def compute_average_BF(n, m, nsim, Xmean, Xstd, Ymean, Ystd, Ydist='Normal',
thetaval = np.linspace(0.001, 60, 100), rdseed = 12231):
if thetaval is None:
ntheta = 1
else:
ntheta = np.shape(thetaval)[0]
BFmat = np.zeros((nsim, ntheta))
ProbM1mat = np.zeros((nsim, ntheta))
for ll in range(nsim):
np.random.seed(rdseed)
X = np.reshape(normal(Xmean, Xstd, n), (n, 1))
Zx = normal(Xmean, Xstd, int(m / 2))
if Ydist == 'Normal':
Y = np.reshape(normal(Ymean, Ystd, n), (n, 1))
Zy = normal(Ymean, Ystd, int(m / 2))
elif Ydist == 'Laplace':
Y = np.reshape(laplace(Ymean, Ystd, n), (n, 1))
Zy = laplace(Ymean, Ystd, int(m / 2))
else:
raise NotImplementedError
Z = np.reshape(np.concatenate((Zx, Zy)), (m, 1))
if thetaval is None:
K = GaussianKernel()
XY = np.reshape(np.concatenate((X, Y)), (2 * n, 1))
median_heuristic_theta = K.get_sigma_median_heuristic(XY)
BF_val, prob_M1_val = compute_ProbM1(X, Y, Z, np.array([median_heuristic_theta]), Independent=True)
else:
BF_val, prob_M1_val = compute_ProbM1(X, Y, Z, thetaval, Independent=True)
median_heuristic_theta = None
BFmat[ll, :] = BF_val.reshape(-1)
ProbM1mat[ll,:] = prob_M1_val.reshape(-1)
rdseed += 1
return BFmat, ProbM1mat, median_heuristic_theta
def varRange(n):
for i in range(1000):
fillTable(normal(0, 1, n))
addRow("normal", n)
for i in range(1000):
fillTable(standard_cauchy(n))
addRow("cauchy", n)
for i in range(1000):
fillTable(laplace(0, 2**(-0.5), n))
addRow("laplace", n)
for i in range(1000):
fillTable(poisson(10, n))
addRow("poisson", n)
for i in range(1000):
fillTable(uniform(-1 * (3**0.5), 3**0.5, n))
addRow("uniform", n)
def find(data, goal_number, failure, eps, delta, promise=-1):
# TODO docstring
"""
:param data:
:param goal_number:
:param failure:
:param eps:
:param delta:
:param promise:
:return:
"""
domain = abs(max(np.max(data, axis=0)) - min(np.min(data, axis=0)))
if promise == -1:
promise = __promise__(data, domain, eps, delta, failure)
all_distances = distances(data)
if __max_average_ball__(0, all_distances, goal_number) + laplace(0, 4/eps, 1) >\
goal_number - 2*promise - 4/eps*log(2/failure):
return 0
extended_domain = 2**int(ceil(log2(domain)))
max_averages_by_radius = [
__max_average_ball__(r, all_distances, goal_number)
for r in arange(0, extended_domain, 0.5)
]
def quality(d, r):
try:
return min(
goal_number - max_averages_by_radius[r],
max_averages_by_radius[2 * r] - goal_number + 4 * promise) / 2
except IndexError:
raise IndexError('error while trying to qualify %f' % r)
# TODO must complete those two
def radius_interval_bounding(data_set, domain_end, j):
return max(
min(quality(data_set, i) for i in xrange(a, a + 2**j))
for a in xrange(domain_end - 2**j))
def max_radius_in_interval(data_set, i):
return max(quality(data_set, r) for r in i)
return evaluate(data, domain, quality, promise, 0.5, eps, delta,
radius_interval_bounding, max_radius_in_interval)
def sub_sampling(point, num_of_samples, l_b=-1, u_b=1, scale=0.3):
'''
:param point: one point
:param number_of_samples: # of sample generated
:param l_b: low bound of data
:param u_b: upper bound of data
:param scale: 1/2lambda*exp(-|x-\mu|/lambda)
:return:
'''
point_dimension = len(point)
data = np.zeros((num_of_samples, point_dimension))
for i in range(num_of_samples):
within_range = False
while not within_range:
cur_sample = laplace(point, scale*np.ones((point_dimension,)))
if all([e <= u_b and e >= l_b for e in cur_sample]):
within_range = True
data[i,:] = cur_sample
return data
def add_laplacian_noise(image, loc=0.0, scale=1.0):
"""Add Laplacian noise to image.
It has probability density function
f(x; mu, lambda) = 1 / (2 * lambda) * exp(-|x - mu| / lambda)
Parameters
----------
image : numpy 2D array
Input image
loc : float
The position, mu, of the distribution peak
scale: float
lambda, the exponential decay
Returns
-------
numpy 2D array
Image with Laplacian noise
"""
noise = laplace(loc, scale, image.shape)
return array(clip(image + noise, 0, 255), dtype=int)
def test_laplaces():
# laplace seems to work
dist_list = []
param_list = []
myvar1 = 1
H = RA.uniform(-10, 10, (m, n)) # H is observed
G = H + RA.laplace(0, myvar1, (m, n))
for j in range(n):
dist_list.append('laplace')
param_list.append([myvar1])
y = G @ x + RA.randn(m)
myvar2 = 1
dist_list.append('laplace')
param_list.append([myvar2])
lf = loglikelihood(H, y, dist_list, param_list)
f, g = lf.func_and_grad(x)
f1, _ = lf.func_and_grad(x + eps * e1)
f2, _ = lf.func_and_grad(x + eps * e2)
approx_grad = (1 / eps) * np.array([f1 - f, f2 - f])
print(g[0])
print(approx_grad)
def sample_and_encode(X, scale=1.):
#X = diabetes.data
#truth = numpy.zeros_like(X[0])
#y = diabetes.target
truth = rng.laplace(0, scale, size=X.shape[1])
y = numpy.dot(X, truth)
yc = y - y.mean()
print
print ' T |', '%6.3f |' % abs(truth).sum(),
print' '.join('%6.3f' % b for b in truth)
guess = numpy.zeros_like(yc)
for i, (guess, beta) in enumerate(lmj.lars.lars(yc, X)):
print '%2d |' % i, '%6.3f |' % abs(beta).sum(),
print ' '.join('%6.3f' % b for b in beta),
print '| yerr %.5f' % numpy.linalg.norm(yc - guess),
print ' xerr %.5f' % numpy.linalg.norm(truth - beta)
return yc, guess
def mean_operation(elist, auths, dp):
G = EcGroup(nid=conf.EC_GROUP)
esum = Classes.Ct.sum(elist)
plain_sum = collective_decryption(esum, auths)
mean = float(plain_sum)/float(len(elist))
if (conf.DP and dp):
#deltaf = 1/relays * 80 (80 is an arbitrary value for our data.)
d = float(100)/float(len(elist))
from numpy.random import laplace
scale = float(d) / conf.EPSILON
noise = laplace(0, scale)
noisy_mean = mean + noise
#print "noise " + str(noise)
else:
noisy_mean=mean
return str(noisy_mean)
def laplace_mechanism(x: Union[int, float, ndarray], sensitivity: float,
privacy_budget: PrivacyBudget) -> Union[float, ndarray]:
"""Differentially private Laplace mechanism. Add Laplacian noise to the value:
.. math::
x + Laplace\left(\mu=0, \sigma=\\frac{\Delta f}{\epsilon}\\right)
The result guarantees :math:`(\epsilon,0)`-differential privacy.
:param x: Sensitive input data
:param sensitivity: The global L1-sensitivity :math:`\Delta f` of `x`
:param privacy_budget: The privacy budget :math:`(\epsilon,0)` used for the outputs
:return: Input data protected by noise
"""
check_positive(privacy_budget.epsilon)
check_positive(sensitivity)
shape = None if isinstance(x, (int, float)) else x.shape
noise = laplace(loc=0.,
scale=sensitivity / privacy_budget.epsilon,
size=shape)
return x + noise
def generate_data(D_x, D_y, N_int, N_ext, beta=[], sigma=0.1, DP_eps=1.0):
if len(beta) == 0:
beta = npr.randn(D_x, D_y)
x_int = npr.randn(N_int, D_x)
x_ext = npr.randn(N_ext, D_x)
y_int = x_int.dot(beta) + sigma * npr.randn(N_int, D_y)
y_ext = x_ext.dot(beta) + sigma * npr.randn(N_ext, D_y)
xx_int = x_int.T.dot(x_int) / N_int
xx_ext = x_ext.T.dot(x_ext) / N_ext
xy_int = x_int.T.dot(y_int) / N_int
xy_ext = x_ext.T.dot(y_ext) / N_ext
DP_xx_scale = 1. / (DP_eps / 2.)
DP_xy_scale = 2. * D_x**2 / (DP_eps / 2.)
xx_ext_DP = xx_ext + sps.wishart.rvs(
df=D_x + 1, scale=np.eye(D_x) * DP_xx_scale / N_ext)
xy_ext_DP = xy_ext + npr.laplace(scale=DP_xy_scale / N_ext,
size=(D_x, D_y))
return ({
'N_ext': N_ext,
'N_int': N_int,
'D_x': D_x,
'D_y': D_y,
'DP_xx_scale': DP_xx_scale,
'DP_xy_scale': DP_xy_scale,
'xx_int': xx_int,
'xx_ext_DP': xx_ext_DP,
'xy_int': xy_int,
'xy_ext_DP': xy_ext_DP,
}, {
'x_int': x_int,
'x_ext': x_ext,
'y_int': y_int,
'y_ext': y_ext,
'xx_ext': xx_ext,
'xy_ext': xy_ext,
'beta': beta
})
def sample(self, size=None):
"""Samples from the distribution
Keyword Arguments:
size {int or tuple of ints} -- Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are drawn.
If size is None (default), a single value is returned if loc and scale are both scalars. Otherwise,
np.broadcast(loc, scale).size samples are drawn. (default: {None})
Returns:
ndarray or scalar -- Drawn samples from the parameterized distribution.
"""
if self.distribution_type == DistributionType.UNIFORM:
return random.uniform(self.low, self.high, size)
elif self.distribution_type == DistributionType.GAUSSIAN:
return random.normal(self.average, self.standard_deviation, size)
elif self.distribution_type == DistributionType.LAPLACIAN:
return random.laplace(self.average, self.standard_deviation, size)
elif self.distribution_type == DistributionType.EXPONENTIAL:
return int(math.ceil(random.exponential(self.beta)))
elif self.distribution_type == DistributionType.CONSTANT:
return self.value
else:
raise NotImplementedError("Distribution type not yet implemented")
def thresholdout(train, holdout, threshold, tolerance):
if np.abs(train - holdout) < threshold + laplace(scale=tolerance):
return train
else:
return holdout + laplace(scale=tolerance)
def __rec_sanitize__(samples, domain_range, alpha, beta, eps, delta, dimension):
# print domain_range
# print calls
global calls
global san_data
# step 1
if calls == 0:
return
calls -= 1
# step 2
# the use of partial is redundant
samples_domain_points = partial(points_in_subset, samples)
noisy_points_in_range = samples_domain_points(subset=domain_range) + laplace(0, 1/eps, 1)
sample_size = len(samples)
# step 3
if noisy_points_in_range < alpha*sample_size/8:
base_range = domain_range
san_data.extend(base_range[1] * noisy_points_in_range)
return san_data
# step 4
domain_size = domain_range[1] - domain_range[0] + 1
log_size = int(ceil(log(domain_size, 2)))
# not needed
# size_tag = 2**log_size
# step 6
def quality(data, j):
return min(point_count_intervals_bounding(data, domain_range, j)-alpha * sample_size / 32,
3 * alpha * sample_size / 32 - point_count_intervals_bounding(data, domain_range, j-1))
# not needed if using exponential_mechanism
# step 7
# promise = alpha * sample_size / 32
# step 8
new_eps = eps/3/log_star(dimension)
# new_delta = delta/3/log_star(dimension)
# note the use of exponential_mechanism instead of rec_concave
z_tag = exponential_mechanism(samples, range(log_size+1), quality, new_eps)
z = 2 ** z_tag
# step 9
if z_tag == 0:
point_counter = Counter(samples)
def special_quality(data, b):
return point_counter[b]
b = choosing_mechanism(samples, range(domain_range[0], domain_range[1] + 1), special_quality,
1, alpha/64., beta, eps, delta)
a = b
# step 10
else:
first_intervals = __build_intervals_set__(samples, 2*z, domain_range[0], domain_range[1] + 1)
second_intervals = __build_intervals_set__(samples, 2*z_tag, domain_range[0], domain_range[1] + 1, True)
intervals = [(i, i+2*z-1) for i in first_intervals+second_intervals]
a, b = choosing_mechanism(samples, intervals, points_in_subset, 2, alpha/64., beta, eps, delta)
if type(a) == str:
raise ValueError("stability problem - choosing_mechanism returned 'bottom'")
# step 11
# although not mentioned I assume the noisy value should be rounded
noisy_count_ab = int(samples_domain_points((a, b)) + laplace(0, 1/eps, 1))
san_data.extend([b] * noisy_count_ab)
# step 12
if a > domain_range[0]:
rec_range = (domain_range[0], a - 1)
__rec_sanitize__(samples, rec_range, alpha, beta, eps, delta, dimension)
if b < domain_range[1]:
rec_range = (b + 1, domain_range[1])
__rec_sanitize__(samples, rec_range, alpha, beta, eps, delta, dimension)
return san_data
def laplace(self, mean, scale):
'''Parameters:\n
mean: float. scale: float, >=0.
'''
return r.laplace(mean, scale, self.size)
return out
if __name__ == "__main__":
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
# Generate random samples for testing the algorithm
Ndim = 3
Nsamples = 1000
# Generate random points . . . some Gausian some Laplacian
mean = np.array(Ndim * [5.0,])
cov = sp.eye(Ndim) + 5.0
samples = random.multivariate_normal(mean, cov, Nsamples)
samples[:50] = random.laplace(5, scale =.0000001, size=(50, mean.size))
# Make a 3D figure
fig = plt.figure(0)
fig.clear()
ax = fig.add_subplot(1,1,1, projection='3d')
ax.scatter(samples[:,0], samples[:,1], samples[:,2], marker='.')
def no_test_DP_median():
d, w = 25, 7
print("Sketch: d=%s w=%s (Cmp. size: %s%%)" % (d, w, (float(100*d*w)/1000)))
# Setup the crypto
G = EcGroup()
sec = G.order().random()
y = sec * G.generator()
# Get some test data
narrow_vals = 1000
wide_vals = 200
from numpy.random import laplace
from collections import defaultdict
datapoints = defaultdict(list)
eps = ["Inf", 1, 0.5, 0.1, 0.05, 0.01, 0.005, 0.001]
for epsilon in eps:
print(epsilon)
for _ in range(40):
vals = [gauss(300, 25) for _ in range (narrow_vals)]
vals += [gauss(500, 200) for _ in range (wide_vals)]
vals = sorted([int(v) for v in vals])
median = vals[int(len(vals) / 2)]
# Add each sample to the sketch
cs = CountSketchCt.epsilondelta(0.05, 0.05, y) # CountSketchCt(d, w, y)
for x in vals:
cs.insert("%s" % x)
# Now use test the median function
proto = get_median(cs, min_b = 0, max_b = 1000, steps = 20)
plain = None
no_decryptions = 0
while True:
v = proto.send(plain)
if isinstance(v, int):
break
no_decryptions += 1
noise = 0
if isinstance(epsilon, float):
scale = float(d) / epsilon
noise = int(round(laplace(0, scale)))
plain = v.dec(sec) + noise
print("Estimated median: %s\t\tAbs. Err: %s" % (v, abs(v - median)))
datapoints[epsilon] += [ abs(v - median) ]
import matplotlib.pyplot as plt
from numpy import mean, std
upper_err = []
core_err = []
lower_err = []
for e in eps:
samples = sorted(datapoints[e])
core_err.append(mean(samples))
upper_err.append(mean(samples) + std(samples) / (len(samples)**0.5))
lower_err.append(mean(samples) - std(samples) / (len(samples)**0.5))
eps_lab = range(len(eps))
eps = ["Inf"] + [e * 10 for e in eps][1:]
plt.plot(eps_lab, core_err, label="Absolute Error")
plt.yscale('log')
# v = v_issue / (len(cnt_issue)**0.5)
plt.fill_between(x=eps_lab, y1=lower_err, y2=upper_err, alpha=0.2, color="b")
plt.xticks(eps_lab, eps)
plt.xlabel(r'Differential Privacy parameter (epsilon)')
plt.ylabel('Absolute Error (mean & std. of mean)')
plt.title(r'Median Estimation - Quality vs. Protection')
# plt.axis([1, 10, 0, 1700])
# plt.grid(True)
plt.savefig("Quality.pdf")
# plt.show()
plt.close()
total = 0; # 这是一个全局变量
# 可写函数说明
def sum2( arg1, arg2 ):
#返回2个参数的和."
total = arg1 + arg2; # total在这里是局部变量.
print ("函数内是局部变量 : ", total)
return total
#调用sum函数
total = sum2( 10, 20 )
print ("函数外是全局变量 : ", total)
# global 和 nonlocal关键字可以改变作用域
print(nmr.laplace(1.1))
## python3 数据结构
#列表list可以修改，tuple和字符串不行
A = [12,3,45.6]
print(A.count('str'))
A.append(17.9)
print(A)
A.sort(reverse=False)
print(A)
A.sort(reverse=True)
print(A)
queue_me = deque(['1', 'aeee', 'dugu'])
print(queue_me.popleft())
queue_me.append('ssss')
queue_me.append('sad')
def Laplace_noise(privacy_bugdet, L1_sensitivity, dim):
return rn.laplace(0, L1_sensitivity / np.float(privacy_bugdet), dim)
def gen_gaussian_linear_data(n=10,
d=100,
norm_beta=1,
beta=None,
var_eps=0.1,
s=None,
seed=1,
shift_type='None',
shift_val=0.1,
logistic=False):
'''Generate data
n : int
number of samples
d : int
dimension
norm_beta: float
norm of beta
var_eps: float
variance of epsilon
snr = norm_beta^2 / var_eps
'''
npr.seed(seed=seed)
# x
x = npr.randn(n, d)
if 'shift' in shift_type:
x += shift_val
elif 'scale' in shift_type:
S2 = np.cumsum(np.ones(d))
S2 /= np.sum(S2)
S2 = np.diag(np.sqrt(S2 * d))
x = x @ S2
elif 'spike' in shift_type:
v = np.ones(d)
v /= npl.norm(v)
S = np.eye(d) + (np.sqrt(shift_val) - 1) * np.outer(v, v)
x = x @ S
elif shift_type == 'lap':
x = npr.laplace(size=(n, d))
# beta
if s == None:
s = d
if beta is None:
beta = np.zeros(d)
beta[:s] = npr.randn(s)
beta[:s] /= npl.norm(beta[:s])
if norm_beta == 'd':
norm_beta = d
beta[:s] *= norm_beta
var_mult = 0 if var_eps == 0 else np.sqrt(var_eps)
eps = var_mult * npr.randn(n)
y = x @ beta + eps
if logistic:
# pass through an inv-logit function
pr = 1 / (1 + np.exp(-y))
# binomial distr (bernoulli response var)
# n trials, probability p
z = np.random.uniform(size=n) # random number 0-1
y = (z <= pr).astype(np.int32)
return x, y, beta
def thresholdout(train, holdout, threshold, tolerance):
if np.abs(train - holdout) < threshold + laplace(scale=tolerance):
return train
else:
return holdout + laplace(scale=tolerance)
def Lap(mu,lamb): return rand.laplace(mu,lamb);
def slaplace(nSources, nSamples):
"""
Generates an nSources x nSamples array of sources which are Laplace distributed.
"""
s = laplace(size=(nSources, nSamples))
return s / s.std(axis=1)[:, newaxis]