pprl-core 0.1.3

Facilities for performing privacy-preserving record linkage with Bloom filters.

This package enables core facilities for performing PPRL based on Bloom filters in Python. It is mostly backed by the bitarray package which implements memory-efficient arrays of bits in Python. This package is composed of several submodules which implement different aspects of performing PPRL. It is used by the PPRL service package under the hood to power its PPRL capabilities.

Bitarray primitives

pprl_core.bits contains functions for setting bits in a bitarray. It implements the double hash, enhanced double hash, triple hash and random hash schemes for setting bits based on a set number of initial hash values in a bitarray. It also offers other utility functions for working on PPRL with Bloom filters.

from bitarray import bitarray
from pprl_core import bits

ba = bitarray(20)

# These are all equivalent and result in the bit with the index 5 to be set.
bits.set_bit(ba, 5)
bits.set_bit(ba, 25)
bits.set_bit(ba, -6)

# These are also equivalent and will return True.
bits.test_bit(ba, 5)
bits.test_bit(ba, 25)
bits.test_bit(ba, -6)

# pprl_core.bits implements the double hash, enhanced double hash, random hash and triple hash schemes.
# Depending on chosen scheme, the corresponding functions require different initial hash values.
h0, h1, h2 = 13, 37, 42
k = 5

ba_double = bitarray(32)
bits.double_hash(ba_double, k, h0, h1)
print(ba_double)
# => bitarray('01000010000000000010000100001000')

ba_enhanced_double = bitarray(32)
bits.enhanced_double_hash(ba_enhanced_double, k, h0, h1)
print(ba_enhanced_double)
# => bitarray('10000000000100000010000010100000')

ba_triple = bitarray(32)
bits.triple_hash(ba_triple, k, h0, h1, h2)
print(ba_triple)
# => bitarray('01000000001000000010000000100100')

from random import Random

ba_random = bitarray(32)
bits.random_hash(ba_random, k, Random(h0))
print(ba_random)
# => bitarray('00000000010100101010000000000000')

# Compute the size of a bitarray such that a certain percentage of its bits are set after
# a number of bits are picked at random and set. In this example, the percentage is set to 50% 
# and the amount of random bit sets is 100.
print(bits.optimal_size(.5, 100))
# => 145

# Serialize and deserialize a bitarray into a Base64-encoded form. The size of a deserialized
# bitarray may not always be the same size of the bitarray that generated the Base64 representation.
# This is because the deserialization will always return a bitarray whose length is a multiple of 8.
ba = bitarray("0010101110101001001010110101011101010010100000011101010100111100")
ba_b64_str = bits.to_base64(ba)
print(ba_b64_str)
# => "K6krV1KB1Tw="
ba_from_b64 = bits.from_base64(ba_b64_str)
print(ba == ba_from_b64)
# => True

Hardening

pprl_core.harden contains factory functions for creating hardeners that can be applied to bitarrays. These functions are guaranteed to always return a modified copy of the bitarrays they are supposed to harden. They will never modify the input bitarrays.

from pprl_core import harden
from random import Random
from bitarray import bitarray

# Create a new random bitarray.
rng = Random(727)
ba = bitarray([rng.random() < 0.5 for _ in range(64)])
print(ba)
# => bitarray('0000010100000100110010111001010101001000111110011011100100101000')

# Harden a bitarray by balancing its bits. With an original bitarray size of 64, the bitarray
# is expanded to a size of 128 in which 50% of its bits should be set. So the resulting bit
# count should be 64.
harden_balance = harden.balance()
ba_balance = harden_balance(ba.copy())
print(ba_balance)
# => bitarray('00000101000001001100101110010101010010001111100110111001001010001111101011111011001101000110101010110111000001100100011011010111')
print(ba.count(), ba_balance.count())
# => 27 64

# Harden a bitarray by performing XOR-folding. The resulting bitarray size should be half of the
# original bitarray.
harden_xor = harden.xor_fold()
ba_xor = harden_xor(ba.copy())
print(ba_xor)
# => bitarray('01001101111111010111001010111101')
print(len(ba), len(ba_xor))
# => 64 32

# Harden a bitarray by flipping bits using "randomized response". Performing an XOR of the resulting
# bitarray with the original bitarray will reveal the bits that have been modified as a result of this
# operation. This hardener requires a function that returns a random number generator and a probability
# with which a bit may be modified.
harden_rand_resp = harden.randomized_response(lambda: Random(727 * 2), .5)
ba_rand_resp = harden_rand_resp(ba.copy())
print(ba_rand_resp)
# => bitarray('0000010110000010110011001101110001001000111111011011101100001000')
print(ba ^ ba_rand_resp)
# => bitarray('0000000010000110000001110100100100000000000001000000001000100000')

# Harden a bitarray by randomly permuting its bits. This hardener requires a function that returns a
# random number generator for selecting bits to swap.
harden_permute = harden.permute(lambda: Random(727 * 3))
ba_permute = harden_permute(ba.copy())
print(ba_permute)
# => bitarray('0010110010110010101011110001010111000110001001001110000100001000')

# Harden a bitarray by having all bits be the result of an XOR of its left and right neighbors.
harden_rule_90 = harden.rule_90()
ba_rule_90 = harden_rule_90(ba.copy())
print(ba_rule_90)
# => bitarray('0000100010001011111100101110000000110101100011111010111011000100')

# Harden a bitarray by moving a sliding window over the bitarray which is used to instantiate a 
# random number generator to draw random bits to mutate. In this example, the sliding window has 
# a size of 8 bits and moves forward 4 bits after 2 random bits have been mutated.
harden_rehash = harden.rehash(8, 4, 2)
ba_rehash = harden_rehash(ba.copy())
print(ba_rehash)
# => bitarray('0000110101011110110110111001011111111010111111011011110100111000')

Bitarray similarity

pprl_core.similarity contains functions for computing the similarity of bitarrays. It implements the Dice coefficient, Cosine similarity and the Jaccard index.

from pprl_core import similarity
from random import Random
from bitarray import bitarray

# Create new random bitarrays.
rng = Random(727)
ba_1 = bitarray([rng.random() < 0.5 for _ in range(32)])
ba_2 = bitarray([rng.random() < 0.5 for _ in range(32)])
print(ba_1)
# => bitarray('00000101000001001100101110010101')
print(ba_2)
# => bitarray('01001000111110011011100100101000')

# For all similarity functions, let n1 and n2 be the amount of set bits in ba_1 and ba_2 respectively,
# and let n12 be the amount of set bits in the intersection of ba_1 and ba_2.

# In ba_1 and ba_2, there are only 3 positions where bits are set in both bitarrays. Each similarity
# function will treat this a bit differently.
print((ba_1 & ba_2).count())
# => 3

# Dice coefficient (2 * n12 / (n1 + n2))
print(similarity.dice(ba_1, ba_2))
# => 0.2222222222222222

# Cosine similarity (n12 / sqrt(n1 * n2))
print(similarity.cosine(ba_1, ba_2))
# => 0.22360679774997896

# Jaccard index (n12 / (n1 + n2 - n12))
print(similarity.jaccard(ba_1, ba_2))
# => 0.125

String transformation

pprl_core.transform contains factory functions for performing preprocessing on strings.

from pprl_core import transform

# String normalization performs several steps. All non-ASCII characters are replaced with their
# closest ASCII variants. Unicode normalization in the NFKD form is performed. Non-ASCII characters
# are removed. All characters are converted to their lowercase counterparts and consecutive whitespaces 
# are reduced to a single one.
normalize = transform.normalize()
print(normalize("Müller-Ludenscheidt"))
# => "muller-ludenscheidt"

# Character filtering allows for the definition of a sequence of characters which must be removed
# from a string.
character_filter = transform.character_filter("äöüß")
print(character_filter("Müller-Ludenscheidt"))
# => "Mller-Ludenscheidt"

# Number formatting takes in any numeric string and reduces or expands it to a set amount of decimal places.
number_zero_digits = transform.number(0)
number_six_digits = transform.number(6)
print(number_zero_digits("12.34"))
# => "12"
print(number_six_digits("12.34"))
# => "12.340000"

# Date time formatting takes in date and time strings in a set format and outputs in another.
date_time = transform.date_time("%Y-%m-%d", "%d.%m.%Y")
print(date_time("2024-06-29"))
# => "29.06.2024"

# Phonetic code transformation applies a phonetic code on an input string. It uses the pyphonetics
# library under the hood.
from pyphonetics import Soundex

phonetic_code = transform.phonetic_code(Soundex())
print(phonetic_code("Müller-Ludenscheidt"))
# => "M464"

# Mapping transformation allows for single characters or entire character sequences to be
# replaced with another. 
mapping = transform.mapping({
    "male": "m",
    "female": "f"
})

print(mapping("male"))
# => "m"

# If no default value is set and no mapping is present, this will throw an error.
print(mapping("unknown"))
# => ValueError: value `unknown` has no mapping, or no default value is present

mapping_with_default = transform.mapping({
    "male": "m",
    "female": "f"
}, default_val="u")

# Setting a default value will prevent this error.
print(mapping("unknown"))
# => "u"

# By default, only entire strings are mapped. For inline transformations, set the corresponding 
# parameter to True.
mapping_inline = transform.mapping({
    "ä": "ae",
    "ö": "oe",
    "ü": "ue",
    "ß": "ss"
}, inline=True)

print(mapping_inline("Müller-Ludenscheidt"))
# => "Mueller-Ludenscheidt"

Additional phonetic codes

pprl_core.phonetics_extra contains additional phonetic code implementations that are compatible with pyphonetics. At the time, only the "Kölner Phonetik" is implemented, which is a phonetic code that is specialized for the German language.

from pprl_core import phonetics_extra

cologne = phonetics_extra.ColognePhonetics()
print(cologne.phonetics("Müller-Ludenscheidt"))
# => "65752682"

License

MIT.