Patent Application
20190294819
A business organization (a retail business, a professional corporation, a financial institution, and so forth) may collect, process and/or store data that represents sensitive or confidential information about individuals or business organizations. For example, the data may be personal data that may represent names, residence addresses, medical information, salaries, banking information, and so forth. The data may be initially collected or acquired in “plaintext form,” and as such may be referred to as “plaintext data.” Plaintext data refers to ordinarily readable data. As examples, plaintext data may be a sequence of character codes, which represent the residence address of an individual in a particular language; or the plaintext data may be a number that that conveys, for example, a blood pressure reading.
For purposes of controlling access to sensitive information (e.g., information relating to confidential or sensitive information about one or more business enterprises and/or individuals) plaintext data items, which represent the sensitive information, may be converted, through a process called “pseudonymization,” into corresponding pseudonyms, or pseudonym values. In this context, a “plaintext data item” (also referred to as “plaintext,” or a “plaintext value” herein) refers to a unit of data (data representing a string, an integer, a real number, a date, and so forth) that represents ordinarily readable content. As examples, a plaintext data item, or value, may be a date, a number, a blood pressure measurement, a salary, and so forth. A pseudonym value ideally conveys no information about the entity that is associated with the corresponding plaintext value. The pseudonymization process may or may not be reversible, in that reversible pseudonym ization processes allow plaintext values to be recovered from pseudonym values, whereas irreversible pseudonym ization processes do not.
One way to convert a first dataset representing a collection of plaintext values into a corresponding second dataset representing a corresponding collection of pseudonym values is to apply an encryption cipher to the first dataset. As an example, the encryption may be format preserving encryption (FPE), which, as its name implies, preserves the format of the plaintext values. For example, the encryption of a sixteen digit credit card number results in a sixteen digit pseudonym.
The ciphertext resulting from the encryption of the plaintext values may be used as the pseudonym values. However, the encryption may alter, if not entirely obfuscate, properties of the plaintext values. For example, the encryption process may destroy information such as dataset averages and distributions, and the individual pseudonym values may have no statistical correlation with the original plaintext values. More specifically, the ciphertext may have a uniform random distribution, although the corresponding plaintext values may have a different statistical distribution, such as a normal, or Gaussian distribution.
The pseudonymization process may serve various purposes, such as regulating access to sensitive information (represented by the plaintext values) and allowing the sensitive information to be analyzed by third parties. For example, the sensitive data may be personal data, which represents personal information about the public, private and/or professional lives of individuals. In some cases, it may be useful to process pseudonymized data to gather statistical information about the underlying personal information. For example, it may be beneficial to statistically analyze pseudonymized health records (i.e., health records in which sensitive plaintext values have been replaced with corresponding pseudonym values), for purposes of gathering statistical information about certain characteristics (weights, blood pressures, diseases or conditions, diagnoses, and so forth) of particular sectors, or demographics, of the population. The pseudonymization process, such as the above-described process of generating pseudonym values by way of encryption, may, however, potentially alter, if not destroy, statistical properties of the personal information. In other words, a collection of plaintext values may have certain statistical properties that are represented by various statistical measures (means, variances, ranges, distributions, expected and so forth). These statistical properties may not be reflected in the corresponding set of pseudonym values, and accordingly, useful statistical information about the personal information may not be determined from the pseudonymized data.
In accordance with example implementations that are described herein, a pseudonym ization process converts plaintext values into corresponding pseudonym values in a process that preserves one or multiple statistical properties of the plaintext values. More specifically, in accordance with example implementations, an encryption-based pseudonymization process is used that, for each plaintext value, encrypts two data items: the plaintext value resulting in a first encrypted value; and ancillary data that is associated with the plaintext value, resulting in a second encrypted value. Thus, the pseudonymization process provides a set of first encrypted values and a set of second encrypted values. The set of first encrypted values represents a random variable, and the set of second encrypted values represents a random variable. As discussed herein, these two random variables, in turn, may be used to produce a pseudonymized value dataset, which has statistical properties similar to the plaintext value dataset.
In the context of this application, “ancillary data” for a given plaintext value refers to additional or auxiliary data that is associated with a given plaintext value. For example, in accordance with some implementations, the ancillary data may be an index of the plaintext value. For example, a collection of plaintext values may be associated with data gathered from individuals, and the collection of plaintext values may have an associated sample size. Each index, in turn, may correspond to one of the individuals of the dataset. Thus, the sample size may be, for example, 1000 (corresponding to 1000 individuals, for example), and a given index, or index number, of 53 may correspond to, for example, a particular individual in the sample.
In accordance with further example implementations, the ancillary data may be data other than an index, or index number, which is associated with the plaintext value. For example, in accordance with some implementations, a particular plaintext value may be associated with such ancillary data as a name, a zipcode, an address, and so forth.
In accordance with some implementations, for a given plaintext value, the encrypted ancillary data may be used to impart a variance to the corresponding pseudonym value. In this regard, the encrypted ancillary data may be used to align a variance that is added to or subtracted from the encrypted plaintext value. Such variances, in turn, may cause the average and/or established distribution of the set of pseudonyms to approximate the average and/or distribution of the set of plaintext values. In accordance with further example implementations, the encrypted plaintext value and the associated encrypted ancillary data may form a random value pair; and the random value pairs (i.e., one random value pair for each plaintext value) may be applied as inputs to a statistical transformation for purposes of transforming the encrypted data into a set of pseudonymized values having one or multiple predetermined statistical properties that are similar to the set of plaintext values. In accordance with yet further example implementations, the encrypted plaintext value and its associated encrypted ancillary data may, for each plaintext value, form a variance input and a sinusoid input so that the variances and sinusoid inputs may be provided to a statistical transformation to transform the encrypted data into a set of pseudonyms having one or multiple predetermined statistical properties that are similar to the set of plaintext values.
Referring to
Regardless of its particular form, in accordance with some implementations, the computer system 100 may include one or multiple processing nodes 110; and each processing node 110 may include one or multiple personal computers, workstations, servers, rack-mounted computers, special purpose computers, and so forth. Depending on the particular implementations, the processing nodes 110 may be located at the same geographical location or may be located at multiple geographical locations. Moreover, in accordance with some implementations, multiple processing nodes 110 may be rack-mounted computers, such that sets of the processing nodes 110 may be installed in the same rack. In accordance with further example implementations, the processing nodes 110 may be associated with one or multiple virtual machines that are hosted by one or multiple physical machines.
In accordance with some implementations, the processing nodes 110 may be coupled to a storage 160 of the computer system 100 through network fabric (not depicted in
The storage 160 may include one or multiple physical storage devices that store data using one or multiple storage technologies, such as semiconductor device-based storage, phase change memory-based storage, magnetic material-based storage, memristor-based storage, and so forth. Depending on the particular implementation, the storage devices of the storage 160 may be located at the same geographical location or may be located at multiple geographical locations. Regardless of its particular form, the storage 160 may store pseudonymized data records 164, i.e., data records in which certain plaintext data items have been replaced with pseudonyms.
In accordance with some implementations, a given processing node 110 may contain a pseudonymization engine 122, which is constructed to, for each plaintext value, apply an FPE cipher to the plaintext value and also apply the FPE cipher to ancillary data associated with the plaintext value to generate a corresponding pair of encrypted values. In accordance with example implementations, the encrypted value pairs are used to generate a set of pseudonym values in a manner such that the set of pseudonym values has one or multiple statistical properties that approximate, if not exactly match, statistical properties of the plaintext values.
In accordance with example implementations, the processing node 110 may include one or multiple physical hardware processors 134, such as one or multiple central processing units (CPUs), one or multiple CPU cores, and so forth. Moreover, the processing node 110 may include a local memory 138. In general, the local memory 138 is a non-transitory memory that may be formed from, as examples, semiconductor storage devices, phase change storage devices, magnetic storage devices, memristor-based devices, a combination of storage devices associated with multiple storage technologies, and so forth.
Regardless of its particular form, the memory 138 may store various data 146 (data representing plaintext values, pseudonym values, ciphertext, ciphertext representing encrypted plaintext values, ciphertext representing encrypted ancillary data associated with the plaintext values, intermediate results pertaining to the pseudonymization process, and so forth). The memory 138 may store instructions 142 that, when executed by one or multiple processors 134, cause the processor(s) 134 to form one or multiple components of the processing node 110, such as, for example, the pseudonymization engine 122.
In accordance with some implementations, the pseudonymization engine 122 may be implemented at least in part by a hardware circuit that does not include a processor executing machine executable instructions. In this regard, in accordance with some implementations, the pseudonymization engine 122 may be formed in whole or in part by a hardware processor that does not execute machine executable instructions, such as, for example, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), and so forth. Thus, many implementations are contemplated, which are within the scope of the appended claims.
In accordance with example implementations, the encrypted ancillary data may be used to impart variances to the pseudonym values. The ancillary data may be used as a tweak parameter to an FPE or encryption cipher. Moreover, the pseudonymization process may be reversible, i.e., it may be possible to recover the original plaintext value from the pseudonym value.
More specifically, in accordance with example implementations, the encrypted plaintext value forms a base value, and the associated encrypted ancillary data determines how much to vary the base value, i.e., determines a variance to the base value for purposes of forming the corresponding pseudonym value. A determination may first be made regarding a variance range, i.e., the amount that the base value is allowed to vary. Secondly, a determination may be made of corresponding soft and hard limits to ensure that the pseudonym value does not vary outside of an absolute range.
As a more specific example, the plaintext values may be, for example, dates in the format of MM/DD/YYYY (where “M” represents a month digit, “D” represents a day digit and “Y” represents a year digit). For example, a set of dates may be soft limited to the range of Jan. 1, 1900 to Dec. 31, 2010, and a variance range of plus or minus five years may be imposed, resulting in a hard limit of Jan. 1, 1895 to Dec. 31, 2015. The pseudonymization process involves encrypting the plaintext dates to derive base dates, and varying the base dates based on the encrypted ancillary data.
For example, the date of “Dec. 3, 1965” may be encrypted, using an FPE cipher, to produce a ciphertext representing a date of “Jul. 30, 1948.” In this manner, using an FPE cipher, “12,” “03,” and “1965” may be encrypted to “07,” “30,” and “1948,” respectively. The “Jul. 30, 1948” base date may then be varied based on the encryption of the associated ancillary data. For example, the Dec. 3, 1965 plaintext date may be associated with an index number of 132 and a sample size of 2000. The ancillary data for this example may be the modulus, or remainder, which is produced by dividing the sample size by the associated index number. In other words, here, 2000 modulo 132, produces a remainder of 20. The remainder of 20 may be encrypted using FPE, which means that the corresponding ciphertext may be in a range of 0 to 99. This range, in turn may be partitioned into subranges, and each subrange may corresponding to a different variance to be added to the base date of Jul. 20, 1948. The variance may be positive or negative. As examples, a value “0” from the encrypted remainder may correspond to a variance of −1825 days (i.e., five years is subtracted from the base date); a value of “99” may correspond to a variance of +1825 days; a value of “20” (the ciphertext corresponding to the encryption of the ancillary data for this example) may correspond to a variance of −1095 days; and so forth. For this example, 1095 days are subtracted from the base date of Jul. 20, 1948 to produce a final pseudonym date of Jul. 21, 1945.
The above-described pseudonymization process is reversible (with knowledge of the encryption key), in accordance with example implementations. For the above example, it is known that the index of 132 is associated the pseudonym date of Jul. 21, 1945, and it is further known that the sample size is 2000. Therefore, reversing the pseudonymization for the above-described example, the variance is first removed. In other words, FPE encryption may be applied to the remainder of 20 (2000 modulo 132) to derive the variance of −1095. Correspondingly, a variance of +1095 days may be added to the pseudonym date of “Jul. 21, 1945” to produce a date of “Jul. 20, 1948.” FPE decryption may then be applied to the date of “Jul. 20, 1948” to derive the original plaintext date of “Dec. 3, 1965.”
Other techniques may be used to derive a variance from the encrypted ancillary data, in accordance with further implementations.
Thus, as depicted in
Next, the pseudonymization engine 128 may perform the following for each plaintext value. Pursuant to the technique 200, the pseudonymization engine 128 may apply (block 212) FPE to an index associated with the plaintext value to provide an associated variance within the variance range. As noted above, in accordance with some implementations, this may involve determining a result of the function (sample size) modulo index value to determine a remainder that is associated with the plaintext value. Moreover, the technique 200 includes applying the FPE to the plaintext value, pursuant to block 216, to determine a base pseudonym value. Pursuant to block 220, the base pseudonym value may then be adjusted, based on the determined variance, to determine a pseudonym value for the plaintext value. If, pursuant to decision block 224, there is another plaintext value to process, then, control may return to block 212. It is noted that although
In accordance with further example implementations, the pseudonymization engine 122 may encrypt the plaintext values and the associated ancillary data for purposes of generating, for each plaintext value, a pair of random values. The random value pairs, in turn, serve as inputs to a statistical transformation, such as a Marsaglia polar transformation or a Box-Muller transformation. The statistical transformation produces a set of pseudonyms that have specific statistical distributions. The statistical transformation may be used to produce a dataset representing a set of pseudonym values that have a normal distribution, a log-normal distribution, and so forth, to reflect the statistical distribution of the corresponding set of plaintext values.
As a more specific example, in accordance with example implementations, the pseudonymization engine 122 may use a Box-Muller transformation to determine pseudonym values from plaintext values, and the pseudonymization engine 122 may use an inverse Box-Muller transformation to determine plaintext values from pseudonym values. The Box-Muller transformation takes two independent uniformly distributed sets of variables and returns values with a standard normal distribution.
Mathematically, the Box-Muller transformation may be described as follows:
Z
0
=R cos(θ)=√{square root over (−2 ln U1)}cos(2πU2), Eq. 1,
where “Zo” represents the transformed set of values having a normal distribution; and “U1” and “U2” represent the two uniformly distributed variables, where each variable is in the range of 0 to 1.
When data undergoes FPE, the data is effectively permuted within its data set. If a different tweak is used on each point, then the data set is approximately uniformly distributed. An example of a tweak may simply be the index of the data point within the set. There are advantages in returning a uniform distribution, as it increases entropy of the data reducing an attacker's ability to determine the underlying values. However, it may be beneficial to maintain some statistical properties, such as a normal or a log-normal distribution. The Box-Muller transform may be used to achieve this from the ciphertext that is produced by the FPE. The FPE ciphertext provides a first random uniformly distributed variable for the Box-Muller transformation. The second uniformly distributed variable may be provided, by, for example, performing an FPE permute operation on the index value itself (or on any associated unique data for each point, if available).
As depicted in
In accordance with example implementations, the pseudonymization process that is depicted in
U
1
=e
((Z
/cos(2πU2))
/−2), Eq 2
The pseudonymization engine 122 may decrypt the (a0, a1, a2 . . . an) FPE values to provide the corresponding plaintext values.
In accordance with example implementations, the pseudonymization engine 122 may perform input data conversion as follows. The values for the U1 and U2 variables may each have a range between 0 and 1. These values are generated using the range of the true input data. For example, if there are 1000 data points that range between 0 to 99 then, the pseudonymization engine 122 generates the values for the U1 and U2 variables as follows:
U
1=data point/100, and Eq. 3
U
2=index/1000, Eq. 4
In accordance with example implementations, the sinusoidal term, cos(2 πU2), is not allowed to be zero, as the term is used in the recovery function (Eq. 2). Therefore, in accordance with example implementations, the pseudonym ization engine 122 may apply a constant offset to the sinusoidal term to ensure the term is not zero.
The pseudonymization engine 122 may also adjust the √{square root over (−2 lnU1)} term of Eq. 1, as this term effectively squeezes the U1 term into a smaller set, and reversing this would lead to collisions. Therefore, in accordance with example implementations, the pseudonymization engine 122 applies a scale factor to the √{square root over (−2 lnU1)} term to “spread,” or de-squeeze,” the values out, such that recovery of exact values may be achieved.
Thus, referring to
Referring to
Referring to
While the present disclosure has been described with respect to a limited number of implementations, those skilled in the art, having the benefit of this disclosure, will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations
