to evaluate whether it's possible to build an automatic sensitivity calculation mechanism which could reliably produce sensitivity scores using only a forward pass through a dynamic graph. However, we want to extend the v1 experiment in several directions:
from collections import Counter
import numpy as np
class PrivateNumber():
def __init__(self, value, max_val, min_val):
self.value = value
self.max_val = max_val
self.min_val = min_val
def __add__(self, other):
# add to a private number
if(isinstance(other, PrivateNumber)):
entities = self.entities.union(other.entities)
new_val = self.value + other.value
entities = set(self.max_val.keys()).union(set(other.max_val.keys()))
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_max_val[entity] = self.max_val[entity] + other.max_val[entity]
new_min_val[entity] = self.min_val[entity] + other.min_val[entity]
return PrivateNumber(self.value + other.value,
new_max_val,
new_min_val)
entities = self.entities
# add to a public number
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_max_val[entity] = self.max_val[entity] + other
new_min_val[entity] = self.min_val[entity] + other
return PrivateNumber(self.value + other,
new_max_val,
new_min_val)
def __sub__(self, other):
return self + (-other)
def __mul__(self, other):
if(isinstance(other, PrivateNumber)):
entities = self.entities.union(other.entities)
new_self_max_val = Counter()
new_self_min_val = Counter()
for entity in entities:
# the biggest positive number this entity could contribute is when
# it is multiplied by the largest value of the same sign from other
new_self_max_val[entity] = max(self.min_val[entity] * other.xmin,
self.max_val[entity] * other.xmax)
# the smallest negative number this entity could contribute is when
# it is multiplied by the largest value of the opposite sign from other
new_self_min_val[entity] = min(self.min_val[entity] * other.xmax,
self.max_val[entity] * other.xmin)
new_other_max_val = Counter()
new_other_min_val = Counter()
for entity in entities:
# the biggest positive number this entity could contribute is when
# it is multiplied by the largest value of the same sign from other
new_other_max_val[entity] = max(other.min_val[entity] * self.xmin,
other.max_val[entity] * self.xmax)
# the smallest negative number this entity could contribute is when
# it is multiplied by the largest value of the opposite sign from other
new_other_min_val[entity] = min(other.min_val[entity] * self.xmax,
other.max_val[entity] * self.xmin)
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_max_val[entity] = max(new_self_max_val[entity], new_other_max_val[entity])
new_min_val[entity] = min(new_self_min_val[entity], new_other_min_val[entity])
return PrivateNumber(self.value * other.value,
new_max_val,
new_min_val)
entities = self.entities
new_max_val = Counter()
for entity in entities:
new_max_val[entity] = self.max_val[entity] * other
new_min_val = Counter()
for entity in entities:
new_min_val[entity] = self.min_val[entity] * other
if(other > 0):
return PrivateNumber(self.value * other,
new_max_val,
new_min_val)
else:
return PrivateNumber(self.value * other,
new_min_val,
new_max_val)
def __truediv__(self, other):
if(isinstance(other, PrivateNumber)):
raise Exception("probably best not to do this - it's gonna be inf a lot")
entities = self.entities
new_max_val = Counter()
for entity in entities:
new_max_val[entity] = self.max_val[entity] / other
new_min_val = Counter()
for entity in entities:
new_min_val[entity] = self.min_val[entity] / other
return PrivateNumber(self.value / other,
new_max_val,
new_min_val)
def __gt__(self, other):
"""BUG!: Counter() defaults to 0"""
if(isinstance(other, PrivateNumber)):
entities = self.entities.union(other.entities)
new_self_max_val = Counter()
new_self_min_val = Counter()
for entity in entities:
if not (self.min_val[entity] > other.xmax or self.max_val[entity] < other.xmin):
new_self_max_val[entity] = 1
else:
new_self_max_val[entity] = 0
new_self_min_val[entity] = 0
new_other_max_val = Counter()
new_other_min_val = Counter()
for entity in entities:
if not (other.min_val[entity] > self.xmax or other.max_val[entity] < self.xmin):
new_other_max_val[entity] = 1
else:
new_other_max_val[entity] = 0
new_other_min_val[entity] = 0
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_max_val[entity] = max(new_self_max_val[entity], new_other_max_val[entity])
new_min_val[entity] = min(new_self_min_val[entity], new_other_min_val[entity])
return PrivateNumber(int(self.value > other.value),
new_max_val,
new_min_val)
entities = self.entities
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_min_val[entity] = 0
if(other <= self.max_val[entity] and other >= self.min_val[entity]):
new_max_val[entity] = 1
else:
new_max_val[entity] = 0
return PrivateNumber(int(self.value > other),
new_max_val,
new_min_val)
def __lt__(self, other):
"""BUG!: Counter() defaults to 0"""
if(isinstance(other, PrivateNumber)):
entities = self.entities.union(other.entities)
new_self_max_val = Counter()
new_self_min_val = Counter()
for entity in entities:
if not (self.min_val[entity] > other.xmax or self.max_val[entity] < other.xmin):
new_self_max_val[entity] = 1
else:
new_self_max_val[entity] = 0
new_self_min_val[entity] = 0
new_other_max_val = Counter()
new_other_min_val = Counter()
for entity in entities:
if not (other.min_val[entity] > self.xmax or other.max_val[entity] < self.xmin):
new_other_max_val[entity] = 1
else:
new_other_max_val[entity] = 0
new_other_min_val[entity] = 0
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_max_val[entity] = max(new_self_max_val[entity], new_other_max_val[entity])
new_min_val[entity] = min(new_self_min_val[entity], new_other_min_val[entity])
return PrivateNumber(int(self.value < other.value),
new_max_val,
new_min_val)
entities = self.entities
new_max_val = Counter()
new_min_val = Counter()
for entity in entities:
new_min_val[entity] = 0
if(other <= self.max_val[entity] and other >= self.min_val[entity]):
new_max_val[entity] = 1
else:
new_max_val[entity] = 0
return PrivateNumber(int(self.value < other),
new_max_val,
new_min_val)
def __neg__(self):
return self * -1
def max(self, other):
if(isinstance(other, PrivateNumber)):
raise Exception("Not implemented yet")
entities = self.entities
new_min_val = Counter()
for entity in entities:
new_min_val[entity] = max(self.min_val[entity], other)
return PrivateNumber(max(self.value, other),
self.max_val,
new_min_val)
def min(self, other):
if(isinstance(other, PrivateNumber)):
raise Exception("Not implemented yet")
entities = self.entities
new_max_val = Counter()
for entity in entities:
new_max_val[entity] = min(self.max_val[entity], other)
return PrivateNumber(min(self.value, other),
new_max_val,
self.min_val)
def hard_sigmoid(self):
return self.min(1).max(0)
def hard_sigmoid_deriv(self):
return ((self < 1) * (self > 0)) + (self < 0) * 0.01 - (self > 1) * 0.01
def __repr__(self):
return str(self.value) + " " + str(self.max_val) + " " + str(self.min_val)
@property
def xmin(self):
return self.min_val.most_common(len(self.min_val))[-1][1]
@property
def xmax(self):
return self.max_val.most_common(1)[0][1]
@property
def entities(self):
return set(self.max_val.keys())
@property
def sensitivity(self):
sens = Counter()
for entity, value in self.max_val.items():
sens[entity] = value - self.min_val[entity]
return sens.most_common()[0][1]
x = PrivateNumber(0.5,Counter({"bob":4, "amos":3}),Counter({"bob":3, "amos":2}))
y = PrivateNumber(1,Counter({"bob":1}),Counter({"bob":-1}))
z = PrivateNumber(-0.5,Counter({"sue":2}),Counter({"sue":-1}))
a = x > y
a.sensitivity
0
a = x + y
b = a * z
b
-1.0 Counter({'sue': 8, 'bob': 8, 'amos': 6}) Counter({'amos': -3, 'sue': -6, 'bob': -6})
# class PrivacyAccountant():
# def __init__(self, default_budget = 0.1):
# self.entity2epsilon = {}
# self.entity2id = {}
# self.default_budget = default_budget
# def add_entity(self, entity_id, budget=None):
# """Add another entity to the system to be tracked.
# Args:
# entity_id: a string or other unique identifier of the entity
# budget: the epsilon level defining this user's privacy budget
# """
# if(budget is None):
# budget = self.default_budget
# self.entity2id[entity_id] = len(self.entity2id)
# self.entity2epsilon[self.entity2id[entity_id]] = budget
# accountant = PrivacyAccountant()
# class DPTensor():
# def __init__(self, data, entities, max_values=None, min_values=None):
# assert data.shape == entities.shape#[0:-1]
# self.data = data
# self.entities = entities
# if max_values is None:
# max_values = np.inf + np.zeros_like(self.data)
# assert max_values.shape == data.shape
# self.max_values = max_values
# if min_values is None:
# min_values = -np.inf + np.zeros_like(self.data)
# assert min_values.shape == data.shape
# self.min_values = min_values
# def sum(self, dim=0):
# _new_data = self.data.sum(dim)
# return _new_data
# @property
# def sensitivity(self):
# return self.max_values - self.min_values
# results, tags = grid.search("diabetes","#data", verbose=False)
# dataset = results['alice'][0][0:5][:,0:4]
# n_ent = dataset.shape[0]
# n_classes = dataset.shape[1]
# for i in range(n_ent):
# accountant.add_entity("Diabetes Patient #" + str(i))
# d2 = dataset.clone().get()
# entities = th.arange(0,n_ent).view(-1,1).expand(n_ent,n_classes)#.unsqueeze(2)
# db = DPTensor(data=d2,
# entities=entities,
# max_values=d2.max(0)[0].expand(n_ent,n_classes),
# min_values=d2.min(0)[0].expand(n_ent,n_classes))