Dirichlet-Mechanism/samplers.py at main · donlapark/Dirichlet-Mechanism · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
import numpy as np
import pandas as pd

from utils import (epsilon2r,
                   epsilon2scale,
                   adp2epsilon_optim,
                   r2epsilon)


class DirichletMechanism:

    def __init__(self, epsilon, prior=1, lambda_=2, Delta_2sq=2, Delta_inf=1):
        self.epsilon = epsilon
        self.lambda_ = lambda_
        self.Delta_2sq = Delta_2sq
        self.Delta_inf = Delta_inf
        self.prior = prior
        self.r = epsilon2r(self.epsilon,
                           self.prior,
                           self.lambda_,
                           self.Delta_2sq,
                           self.Delta_inf)
        self.alpha = self.prior + 4*(self.lambda_-1)*self.r*self.Delta_inf

    def sample(self, x, seed=None):
        r = np.random.default_rng(seed).standard_gamma(self.r*np.array(x)+self.alpha)
        return r / r.sum(-1, keepdims=True)

    def sample_series(self, x, seed=None):
        r = np.random.default_rng(seed).standard_gamma(self.r*np.array(x)+self.alpha)
        r = r / r.sum(-1, keepdims=True)
        return pd.Series(r, index=x.index)

    def set_epsilon(self, new_epsilon):
        self.epsilon = new_epsilon
        self.r = epsilon2r(self.epsilon,
                           self.prior,
                           self.lambda_,
                           self.Delta_2sq,
                           self.Delta_inf)
        self.alpha = self.prior + (self.lambda_-1)*self.r*self.Delta_inf

    def set_lambda(self, new_lambda):
        self.lambda_ = new_lambda
        self.r = epsilon2r(self.epsilon,
                           self.prior,
                           self.lambda_,
                           self.Delta_2sq,
                           self.Delta_inf)
        self.alpha = self.prior + (self.lambda_-1)*self.r*self.Delta_inf

    def set_r(self, r):
        self.r = r
        self.epsilon = r2epsilon(self.r,
                                 self.prior,
                                 self.lambda_,
                                 self.Delta_2sq,
                                 self.Delta_inf)
        self.alpha = self.prior + 4*(self.lambda_-1)*self.r*self.Delta_inf

    def get_alpha(self):
        return self.alpha

    def set_dp_epsilon(self, eps_hat, delta):
        self.epsilon = adp2epsilon_optim(eps_hat, delta)
        self.lambda_ = 1 / self.epsilon + 1
        self.r = epsilon2r(self.epsilon,
                           self.prior,
                           self.lambda_,
                           self.Delta_2sq,
                           self.Delta_inf)
        self.alpha = self.prior + (self.lambda_-1)*self.r*self.Delta_inf


class GaussianMechanism:

    def __init__(self, epsilon, lambda_=2, Delta_2sq=2):
        self.epsilon = epsilon
        self.lambda_ = lambda_
        self.Delta_2sq = Delta_2sq
        self.sigma = np.sqrt(self.lambda_*self.Delta_2sq/(2*self.epsilon))

    def sample(self, x, tol=1e-6, seed=None):
        p = np.clip(x+np.random.default_rng(seed).normal(0, self.sigma, np.shape(x)), a_min=tol, a_max=None)
        p = p/p.sum(-1, keepdims=True)
        return p

    def sample_series(self, x, tol=1e-6, seed=None):
        p = np.clip(x+np.random.default_rng(seed).normal(0, self.sigma, np.shape(x)), a_min=tol, a_max=None)
        p = p/p.sum()
        return p

    def set_epsilon(self, new_epsilon):
        self.epsilon = new_epsilon
        self.sigma = np.sqrt(self.lambda_*self.Delta_2sq/(2*self.epsilon))

    def get_sigma(self):
        return self.sigma

    def set_dp_epsilon(self, eps_hat, delta):
        self.epsilon = None
        self.lambda_ = None
        self.sigma = np.sqrt(2 * np.log(1.25/delta) * self.Delta_2sq) / eps_hat


class LaplaceMechanism:

    def __init__(self, epsilon, lambda_=2, Delta_1=2, d=None):
        self.epsilon = epsilon
        self.lambda_ = lambda_
        self.Delta_1 = Delta_1
        self.d = d
        self.scale = epsilon2scale(self.epsilon, self.lambda_, self.Delta_1)

    def sample(self, x, tol=1e-6, seed=None):
        p = np.clip(x+np.random.default_rng(seed).laplace(0, self.scale, np.shape(x)), a_min=tol, a_max=None)
        p = p/p.sum(-1, keepdims=True)
        return p

    def sample_series(self, x, tol=1e-6, seed=None):
        p = np.clip(x+np.random.default_rng(seed).laplace(0, self.scale, np.shape(x)), a_min=tol, a_max=None)
        p = p/p.sum()
        return p

    def set_rho(self, new_epsilon):
        self.epsilon = new_epsilon
        self.scale = self.Delta_1*np.sqrt(self.lambda_/(2*self.epsilon))

    def get_scale(self):
        return self.scale

    def set_dp_epsilon(self, eps_hat, delta):
        self.epsilon = None
        self.lambda_ = None
        self.scale = self.Delta_1 / eps_hat
        self.scale *= np.sqrt(np.log(0.5 / delta) / np.log(1 / delta))


class MLECalculator:

    def __init__(self, prior):
        self.prior = prior

    def sample(self, x, seed=None):
        p = x + self.prior
        p = p/p.sum(-1, keepdims=True)
        return p

    def sample_series(self, x, seed=None):
        p = x + self.prior
        p = p/p.sum()
        return p