def mean_estimation_experiment():
# generated data
data = np.clip(np.random.normal(loc=0.5, scale=0.2, size=[100000]), 0, 1)
print("this is generated data\n", data)
discretized_data = [
dp.discretization(value=value, lower=0, upper=1) for value in data
]
print("this is discretized data\n", discretized_data)
mean = np.average(data)
print("the mean_solutions of original data is: ", mean)
mean_d = np.average(discretized_data)
print("the mean_solutions of discretized data is: ", mean_d)
epsilon = 1
# dp_data = [dp.random_response_old(B=value, p=dp.eps2p(epsilon)) for value in discretized_data]
dp_data = [
dp.random_response(bits=value, p=dp.eps2p(epsilon))
for value in discretized_data
]
est_one = dp.random_response_reverse(data_list=np.asarray(dp_data),
p=dp.eps2p(epsilon))
est_mean = est_one / len(dp_data)
print("the estimated mean_solutions is: ", est_mean)
def kv_de_f2m(p_kv_list: np.ndarray, epsilon_k, set_value=0):
if not isinstance(p_kv_list, np.ndarray):
raise Exception("type error of p_kv_list: ", type(p_kv_list))
f = np.average(p_kv_list[:, 0])
p = ldplib.eps2p(epsilon=epsilon_k)
f = (p-1+f) / (2*p-1)
m_all = np.average(p_kv_list[:, 1])
m = (m_all - (1 - f) * set_value) / f
return f, m
def random_response_for_hist(user_vector, epsilon):
"""
每个用户对自己的数据进行random response操作
:param user_vector: [0,0,0,0,...,0,0,1,0,0,...]
:param epsilon: privacy budget
:return: [1,0,0,1,1,...]
"""
return ldplib.random_response(bit_array=user_vector,
p=ldplib.eps2p(epsilon=epsilon / 2))
def kv_de_bisample(p_kv_list: np.ndarray, epsilon):
pos_values = p_kv_list[p_kv_list[:, 0] == 1]
neg_values = p_kv_list[p_kv_list[:, 0] == 0]
f_pos = np.average(pos_values[:, 1])
f_neg = np.average(neg_values[:, 1])
p = ldplib.eps2p(epsilon)
f = (2 * p - 2 + f_pos + f_neg) / (2 * p - 1)
m = (f_pos - f_neg) / (f_pos + f_neg + 2 * p - 2)
return f, m
def run(data):
"""
this is a frequency estimation example
"""
frequency = np.sum(data) / len(data)
print("baseline frequency = ", frequency)
epsilon = 1
p = ldplib.eps2p(epsilon)
perturbed_data = ldplib.random_response(bit_array=data, p=p)
f = np.sum(perturbed_data) / len(perturbed_data)
estimated_frequency = (f + p - 1) / (2 * p - 1)
print("estimated frequency = ", estimated_frequency)
print("estimation error = ", np.fabs(estimated_frequency - frequency))
def bisample_decode(p_val_lst, epsilon):
val_lst = np.asarray(p_val_lst)
pos_lst = val_lst[val_lst[:, 0] == 1]
neg_lst = val_lst[val_lst[:, 0] == 0]
pos_val = pos_lst[:, 1]
neg_val = neg_lst[:, 1]
f_pos = 1.0 * sum(pos_val) / len(pos_val)
f_neg = 1.0 * sum(neg_val) / len(neg_val)
p = ldplib.eps2p(epsilon)
m = (f_pos - f_neg) / (f_pos + f_neg + 2 * p - 2)
mr = (1 - f_pos - f_neg) / (2 * p - 1)
return mr, m
def kv_en_onehot(kv, epsilon):
"""
encode a kv into [a,b,c], where:
a=1 represents if the k == 0
b represents if v == -1
c represents if v == 1
"""
k, v = int(kv[0]), kv[1]
onehot = np.zeros([3])
if k == 0:
onehot[0] = 1
else:
d_v = ldplib.discretization(v, -1, 1)
if d_v == -1:
onehot[1] = 1
else:
onehot[2] = 1
return ldplib.random_response(bit_array=onehot,
p=ldplib.eps2p(epsilon / 2))
def kv_en_f2m(kv, epsilon_k, epsilon_v, method, set_value=0):
v = kv[1] if kv[0] == 1 else set_value
p_k = ldplib.random_response_old(bits=int(kv[0]), p=ldplib.eps2p(epsilon_k))
p_v = method(v, epsilon_v)
return p_k, p_v
每个用户对自己的数据进行random response操作
:param user_vector: [0,0,0,0,...,0,0,1,0,0,...]
:param epsilon: privacy budget
:return: [1,0,0,1,1,...]
"""
return ldplib.random_response(bit_array=user_vector,
p=ldplib.eps2p(epsilon=epsilon / 2))
if __name__ == '__main__':
# 生成数据
n = 1000000
data_list = generate_data(user_number=n)
# 隐私参数
epsilon = 1
p = ldplib.eps2p(epsilon / 2)
# 得到原始数据的直方图
original_hist = np.sum(data_list, axis=0)
print("this is original hist: \n", original_hist)
# guessed_hist = original_hist * (2*p-1) + n * (1-p)
# print("this is guessed hist: \n", guessed_hist)
# aggregator收集并处理数据
perturbed_data = np.asarray(
[random_response_for_hist(user, epsilon) for user in data_list])
perturbed_hist = np.sum(perturbed_data, axis=0)
print("this is the hist by the aggregator: \n", perturbed_hist)
# aggregator校正数据
'''
# 得到原始数据的直方图
original_hist = np.sum(data_list, axis=0)
print("this is original hist: \n", original_hist)
# 隐私参数
epsilon = 1
# aggregator收集并处理数据
perturbed_data_list = np.asarray(
[random_response_for_hist(user, epsilon) for user in data_list])
print("this is the hist by the aggregator: \n",
np.sum(perturbed_data_list, axis=0))
# aggregator校正数据
# estimate_hist = [dp.random_response_adjust(rr_sum, len(data_list), epsilon / 2) for rr_sum in rr_sums]
# print(np.sum(estimate_hist))
estimate_hist = dp.random_response_reverse(data_list=perturbed_data_list,
p=dp.eps2p(epsilon=epsilon / 2))
# 展示原始数据的直方图
print("this is estimated hist: \n", estimate_hist)
# 画图
fig = plt.figure(figsize=[12, 5])
ax1 = fig.add_subplot(121) # 2*2的图形 在第一个位置
ax1.bar(range(len(original_hist)), original_hist)
ax2 = fig.add_subplot(122)
ax2.bar(range(len(estimate_hist)), estimate_hist)
plt.show()
