123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324 |
- """
- Health Information Privacy Lab
- Brad. Malin, Weiyi Xia, Steve L. Nyemba
- This framework computes re-identification risk of a dataset assuming the data being shared can be loaded into a dataframe (pandas)
- The framework will compute the following risk measures:
- - marketer
- - prosecutor
- - pitman
- References :
- https://www.scb.se/contentassets/ff271eeeca694f47ae99b942de61df83/applying-pitmans-sampling-formula-to-microdata-disclosure-risk-assessment.pdf
- This framework integrates pandas (for now) as an extension and can be used in two modes :
- Experimental mode
- Here the assumption is that we are not sure of the attributes to be disclosed, the framework will explore a variety of combinations and associate risk measures every random combinations
- Evaluation mode
- The evaluation mode assumes the set of attributes given are known and thus will evaluate risk for a subset of attributes.
- features :
- - determine viable fields (quantifiable in terms of uniqueness). This is a way to identify fields that can act as identifiers.
- - explore and evaluate risk of a sample dataset against a known population dataset
- - explore and evaluate risk on a sample dataset
- Usage:
- from pandas_risk import *
- mydataframe = pd.DataFrame('/myfile.csv')
- resp = mydataframe.risk.evaluate(id=<name of patient field>,num_runs=<number of runs>,cols=[])
- resp = mydataframe.risk.explore(id=<name of patient field>,num_runs=<number of runs>,cols=[])
- @TODO:
- - Provide a selected number of fields and risk will be computed for those fields.
- - include journalist risk
- """
- import pandas as pd
- import numpy as np
- import logging
- import json
- from datetime import datetime
- import sys
- from itertools import combinations
- @pd.api.extensions.register_dataframe_accessor("risk")
- class deid :
- """
- This class is a deidentification class that will compute risk (marketer, prosecutor) given a pandas dataframe
- """
- def __init__(self,df):
- self._df = df.fillna(' ')
- #
- # Let's get the distribution of the values so we know what how unique the fields are
- #
- values = df.apply(lambda col: col.unique().size / df.shape[0])
- self._dinfo = dict(zip(df.columns.tolist(),values))
- def explore(self,**args):
- """
- This function will perform experimentation by performing a random policies (combinations of attributes)
- This function is intended to explore a variety of policies and evaluate their associated risk.
- :pop|sample data-frame with population or sample reference
- :field_count number of fields to randomly select
- :strict if set the field_count is exact otherwise field_count is range from 2-field_count
- :num_runs number of runs (by default 5)
- """
-
- pop= args['pop'] if 'pop' in args else None
-
- if 'pop_size' in args :
- pop_size = np.float64(args['pop_size'])
- else:
- pop_size = -1
-
-
- #
- # Policies will be generated with a number of runs
- #
- RUNS = args['num_runs'] if 'num_runs' in args else 5
-
- sample = args['sample'] if 'sample' in args else pd.DataFrame(self._df)
-
- k = sample.columns.size if 'field_count' not in args else int(args['field_count']) +1
- #
- # remove fields that are unique, they function as identifiers
- #
- if 'id' in args :
- id = args['id']
- columns = list(set(sample.columns.tolist()) - set([id]))
- else:
- columns = sample.columns.tolist()
-
- # If columns are not specified we can derive them from self._dinfo
- # given the distribution all fields that are < 1 will be accounted for
- # columns = args['cols'] if 'cols' in args else [key for key in self._dinfo if self._dinfo[key] < 1]
-
- o = pd.DataFrame()
- columns = [key for key in self._dinfo if self._dinfo[key] < 1]
- _policy_count = 2 if 'policy_count' not in args else int(args['policy_count'])
-
- _policies = []
- _index = 0
- for size in np.arange(2,len(columns)) :
- p = list(combinations(columns,size))
- p = (np.array(p)[ np.random.choice( len(p), _policy_count)].tolist())
- flag = 'Policy_'+str(_index)
- _index += 1
- for cols in p :
- r = self.evaluate(sample=sample,cols=cols,flag = flag)
- p = pd.DataFrame(1*sample.columns.isin(cols)).T
- p.columns = sample.columns
- o = pd.concat([o,r.join(p)])
-
-
- # for i in np.arange(RUNS):
- # if 'strict' not in args or ('strict' in args and args['strict'] is False):
- # n = np.random.randint(2,k)
- # else:
- # n = args['field_count']
- # cols = np.random.choice(columns,n,replace=False).tolist()
- # params = {'sample':sample,'cols':cols}
- # if pop is not None :
- # params['pop'] = pop
- # if pop_size > 0 :
- # params['pop_size'] = pop_size
- # r = self.evaluate(**params)
- # #
- # # let's put the policy in place
- # p = pd.DataFrame(1*sample.columns.isin(cols)).T
- # p.columns = sample.columns
- # # o = o.append(r.join(p))
- # o = pd.concat([o,r.join(p)])
-
- o.index = np.arange(o.shape[0]).astype(np.int64)
- return o
- def evaluate(self, **args):
- """
- This function has the ability to evaluate risk associated with either a population or a sample dataset
- :sample sample dataset
- :pop population dataset
- :cols list of columns of interest or policies
- :flag user provided flag for the context of the evaluation
- """
- if 'sample' in args :
- sample = pd.DataFrame(args['sample'])
- else:
- sample = pd.DataFrame(self._df)
- if not args or 'cols' not in args:
- # cols = sample.columns.tolist()
- cols = [key for key in self._dinfo if self._dinfo[key] < 1]
- elif args and 'cols' in args:
- cols = args['cols']
- #
- #
-
- flag = 'UNFLAGGED' if 'flag' not in args else args['flag']
- #
- # @TODO: auto select the columns i.e removing the columns that will have the effect of an identifier
- #
- # if 'population' in args :
- # pop = pd.DataFrame(args['population'])
- r = {"flag":flag}
- # if sample :
-
- handle_sample = Sample()
- xi = sample.groupby(cols,as_index=False).count().values
-
- handle_sample.set('groups',xi)
- if 'pop_size' in args :
- pop_size = np.float64(args['pop_size'])
- else:
- pop_size = -1
- #
- #-- The following conditional line is to address the labels that will be returned
- # @TODO: Find a more elegant way of doing this.
- #
- if 'pop' in args :
- label_market = 'sample marketer'
- label_prosec = 'sample prosecutor'
- label_groupN = 'sample group count'
- label_unique = 'sample journalist' #'sample unique ratio'
- # r['sample marketer'] = handle_sample.marketer()
- # r['sample prosecutor'] = handle_sample.prosecutor()
- # r['sample unique ratio'] = handle_sample.unique_ratio()
- # r['sample group count'] = xi.size
- # r['sample group count'] = len(xi)
- else:
- label_market = 'marketer'
- label_prosec = 'prosecutor'
- label_groupN = 'group count'
- label_unique = 'journalist' #'unique ratio'
- # r['marketer'] = handle_sample.marketer()
- # r['prosecutor'] = handle_sample.prosecutor()
- # r['unique ratio'] = handle_sample.unique_ratio()
- # r['group count'] = xi.size
- # r['group count'] = len(xi)
- if pop_size > 0 :
- handle_sample.set('pop_size',pop_size)
- r['pitman risk'] = handle_sample.pitman()
- r[label_market] = handle_sample.marketer()
- r[label_unique] = handle_sample.unique_ratio()
- r[label_prosec] = handle_sample.prosecutor()
- r[label_groupN] = len(xi)
-
- if 'pop' in args :
- xi = pd.DataFrame({"sample_group_size":sample.groupby(cols,as_index=False).count()}).reset_index()
- yi = pd.DataFrame({"population_group_size":args['pop'].groupby(cols,as_index=False).size()}).reset_index()
- merged_groups = pd.merge(xi,yi,on=cols,how='inner')
- handle_population= Population()
- handle_population.set('merged_groups',merged_groups)
-
- r['pop. marketer'] = handle_population.marketer()
- r['pitman risk'] = handle_population.pitman()
- r['pop. group size'] = np.unique(yi.population_group_size).size
- #
- # At this point we have both columns for either sample,population or both
- #
- r['field count'] = len(cols)
- return pd.DataFrame([r])
- class Risk :
- """
- This class is an abstraction of how we chose to structure risk computation i.e in 2 sub classes:
- - Sample computes risk associated with a sample dataset only
- - Population computes risk associated with a population
- """
- def __init__(self):
- self.cache = {}
- def set(self,key,value):
- if id not in self.cache :
- self.cache[id] = {}
- self.cache[key] = value
- class Sample(Risk):
- """
- This class will compute risk for the sample dataset: the marketer and prosecutor risk are computed by default.
- This class can optionally add pitman risk if the population size is known.
- """
- def __init__(self):
- Risk.__init__(self)
- def marketer(self):
- """
- computing marketer risk for sample dataset
- """
-
-
- groups = self.cache['groups']
- # group_count = groups.size
- # row_count = groups.sum()
- group_count = len(groups)
- row_count = np.sum([_g[-1] for _g in groups])
- return group_count / np.float64(row_count)
- def prosecutor(self):
- """
- The prosecutor risk consists in determining 1 over the smallest group size
- It identifies if there is at least one record that is unique
- """
- groups = self.cache['groups']
- _min = np.min([_g[-1] for _g in groups])
- # return 1 / np.float64(groups.min())
- return 1/ np.float64(_min)
- def unique_ratio(self):
- groups = self.cache['groups']
- # row_count = groups.sum()
- row_count = np.sum([_g[-1] for _g in groups])
- # return groups[groups == 1].sum() / np.float64(row_count)
- values = [_g[-1] for _g in groups if _g[-1] == 1]
-
- return np.sum(values) / np.float64(row_count)
- def pitman(self):
- """
- This function will approximate pitman de-identification risk based on pitman sampling
- """
- groups = self.cache['groups']
- si = groups[groups == 1].size
- # u = groups.size
- u = len(groups)
- alpha = np.divide(si , np.float64(u) )
- row_count = np.sum([_g[-1] for _g in groups])
- # f = np.divide(groups.sum(), np.float64(self.cache['pop_size']))
- f = np.divide(row_count, np.float64(self.cache['pop_size']))
- return np.power(f,1-alpha)
- class Population(Sample):
- """
- This class will compute risk for datasets that have population information or datasets associated with them.
- This computation includes pitman risk (it requires minimal information about population)
- """
- def __init__(self,**args):
- Sample.__init__(self)
- def set(self,key,value):
- Sample.set(self,key,value)
- if key == 'merged_groups' :
-
- Sample.set(self,'pop_size',np.float64(value.population_group_size.sum()) )
- Sample.set(self,'groups',value.sample_group_size)
- """
- This class will measure risk and account for the existance of a population
- :merged_groups {sample_group_size, population_group_size} is a merged dataset with group sizes of both population and sample
- """
- def marketer(self):
- """
- This function requires
- """
- r = self.cache['merged_groups']
- sample_row_count = r.sample_group_size.sum()
- #
- # @TODO : make sure the above line is size (not sum)
- # sample_row_count = r.sample_group_size.size
- return r.apply(lambda row: (row.sample_group_size / np.float64(row.population_group_size)) /np.float64(sample_row_count) ,axis=1).sum()
|