def draw_mix_uniforms(pbs, pmix, n, seed):
"""
Draw n samples from a mixture of uniform distributions whose bounds
are specified in pbs, and mixing proportion is specified in pmix.
Return a one-dimensional numpy array of n samples.
"""
assert pbs.shape[0] == len(pmix)
c = len(pmix)
sam_list = []
scale = pbs[:, 1] - pbs[:, 0]
with util.NumpySeedContext(seed=seed):
# counts for each mixture component
counts = np.random.multinomial(n, pmix, size=1)
# counts is a 2d array
counts = counts[0]
# For each component, draw from its corresponding uniform
# distribution.
for i, nc in enumerate(counts):
#print i
sam_i = stats.uniform.rvs(loc=pbs[i, 0], scale=scale[i], size=nc)
#print 'pbs[{0},0]: {1}'.format(i, pbs[i, 0])
#print 'sam_i: {0}'.format(sam_i)
sam_list.append(sam_i)
sample = np.hstack(sam_list)
np.random.shuffle(sample)
return sample
def sample(self, n, seed):
with util.NumpySeedContext(seed=seed):
X = SSMixUnif1D.draw_mix_uniforms(self.pbs, self.pmix, n, seed)
X = X[:, np.newaxis]
Y = SSMixUnif1D.draw_mix_uniforms(self.qbs, self.qmix, n, seed+2)
Y = Y[:, np.newaxis]
return TSTData(X, Y, label='mix_unif1d')
def gauss_mech(quantity, sensitivity, epsilon=1.0, delta=1e-4, return_sigma=False, seed=23):
with util.NumpySeedContext(seed=seed+1000):
shape = quantity.shape
sigma = sensitivity * sqrt(2.0 * log(1.25 / delta)) / epsilon
print('Normal')
print('sigma:{}'.format(sigma))
private_quantity = quantity + np.random.normal(loc=0.0, scale=sigma, size=shape)
if return_sigma:
return private_quantity, sigma
else:
return private_quantity
def improve_gauss_mech(quantity, sensitivity, epsilon=1.0,
delta=1e-4, return_sigma=False, seed=23):
with util.NumpySeedContext(seed=seed+1000):
shape = quantity.shape
sigma = calibrateAnalyticGaussianMechanism(epsilon, delta, sensitivity)
print('Improve')
print('sigma:{}'.format(sigma))
private_quantity = quantity + np.random.normal(loc=0.0, scale=sigma, size=shape)
if return_sigma:
return private_quantity, sigma
else:
return private_quantity
def analyse_gauss_mech(quantity, sensitivity, epsilon=0.5,
delta=1e-4, gauss_noise='Normal', seed=23):
with util.NumpySeedContext(seed=seed+3000):
shape = quantity.shape
if shape[0] != shape[1]:
raiseValueError('Shape of input must be square')
J = shape[0]
if gauss_noise == 'Normal':
sigma = sensitivity * sqrt(2.0 * log(1.25 / delta)) / epsilon
elif gauss_noise == 'Improved':
sigma = calibrateAnalyticGaussianMechanism(epsilon, delta, sensitivity)
# construct symmetric matrix
print('analyse_gauss_mech sigma {}'.format(sigma))
eta = np.random.normal( loc=0.0, scale=sigma, size=(J,J))
diag_upper = np.triu(eta)
noise = diag_upper + np.tril(diag_upper.T, k=-1)
private_quantity = quantity + noise
# make it psd
private_PSD = util.PSD(private_quantity, reg=1e-4)
return private_PSD