Obfuscating your SQL Server Data

If you are required to use real production data to test applications, any sensitive data should be "disguised" before loading it into the development environment. Although there are tools to generate convincing test data, it occasionally happens that the variances and frequencies within data cannot easily be simulated. In such cases, the DBA should apply one or more of the obfuscation techniques described in this article, extracted from John Magnabosco's excellent new book, Protecting SQL Server Data.

Halloween is one of my favorite times of the year. On this holiday, the young and young at heart apply make-up, masks, costumes and outfits and wander the streets in search of sweet treats from their neighbors. These costumes are designed to hide the identity of their wearer and grant that person the freedom to shed their everyday demeanor and temporarily adopt the persona of their disguise.

Applying the technique of obfuscation to our sensitive data is somewhat akin to donning a Halloween disguise. By doing so, we mask the underlying data values, hiding their true nature, until the appropriate time to disclose it.

We will explore a handful of obfuscation techniques, which do not require an algorithm, encryption key, decryption key or transformation of data types.

Each of these methods, including character scrambling and masking, numeric variance and nulling, rely on an array of built-in SQL Server system functions that are used for string manipulation.

While these methods of obfuscation will not be used by any federal government to protect nuclear missile launch codes, they can be highly effective when printing documents that contain sensitive data, transferring production data to test environments or presenting data through reporting and warehousing mechanisms.

Development Environment Considerations

Before we proceed with an exploration of obfuscation methods, let’s spend a few moments reviewing a strong candidate for the implementation of the obfuscation methods presented in this article: the development environment.

The database that is utilized for daily business transactions is referred to as the production database. The version of the database that is used to develop and test new functionality is referred to as the development database. These environments are separated so that new development and troubleshooting can occur without having a negative effect on the performance and integrity of the production database.

Any proposed modifications to a production database should be first implemented and tested on a development or test database. In order to ensure the accuracy of this testing, the development database should mimic the production database as closely as possible, in terms of the data it contains and the set of security features it implements.

This means that all of the sensitive data efforts and options noted in this book apply to both environments and that it may be necessary to store sensitive data in both the development and production databases. The difficulty with this is that it is common for developers and testers to be granted elevated permissions within the development database. If the development database contains identical data to that stored in the production database, then these elevated permissions could present a severe and intolerable security risk to the organization and its customers.

In order to mitigate this risk, the Database Administrator responsible for refreshing the contents of the development environment should apply obfuscation methods to hide the actual values that are gleaned from the production environment.

Obfuscation Methods

The word obfuscation is defined by the American Heritage Dictionary as follows:

“To make so confused or opaque as to be difficult to perceive or understand … to render indistinct or dim; darken.”

The word obfuscation, at times, can be used interchangeably with the term obscurity, meaning “the quality or condition of being unknown“. However, there is a subtle difference between the two terms and the former definition is more appropriate since obscurity implies that the hidden condition can be achieved without any additional effort.

Many methods of “disguise”, or obfuscation, are available to the Database Administrator that can contribute a level of control to how sensitive data is stored and disclosed, in both production and development environments. The options that will be discussed in this article are:

  • Character Scrambling
  • Repeating Character Masking
  • Numeric Variance
  • Nulling
  • Artificial Data Generation
  • Truncating
  • Encoding
  • Aggregating.

Many of these methods rely on SQL Server’s built-in system functions for string manipulation, such as SUBSTRING, REPLACE, and REPLICATE. Appendix A of my book provides a syntax reference for these, and other system functions that are useful in obfuscating sensitive data.

Prior to diving into the details of these obfuscation methods we need to explore the unique value of another system function, called RAND.

The Value of RAND

The RAND system function is not one that directly manipulates values for the benefit of obfuscation, but its ability to produce a reasonably random value makes it a valuable asset when implementing character scrambling or producing a numeric variance.

One special consideration of the RAND system function is that when it is included in a user defined function an error will be returned when the user defined function is created.

This can be overcome by creating a view that contains the RAND system function and referencing the view in the user defined function. The script in Listing 8-1 will create a view in the HomeLending database that returns a random value, using the RAND system function. Since this view holds no security threat, we will make this available to the Sensitive_high, Sensitive_medium and Sensitive_low database roles with SELECT permissions on this view.

Listing 8-1: Generating random numbers using RAND.

Now, we can obtain a random number in any user defined function with a simple call to our new view. In Listing 8-2, an example is provided that produces a random number between the values of 1 and 100.

Listing 8-2: Testing the View.

Character Scrambling

Character scrambling is a process by which the characters contained within a given statement are re-ordered in such a way that its original value is obfuscated. For example, the name “Jane Smith” might be scrambled into “nSem Jatih”.

This option does have its vulnerabilities. The process of cracking a scrambled word is often quite straightforward, and indeed is a source of entertainment for many, as evidenced by newspapers, puzzle publications and pre-movie entertainment.

Cracking a scrambled word can be made more challenging by, for example, eliminating any repeating characters and returning only lower case letters. However, not all values will contain repeating values, so this technique may not be sufficient for protecting highly sensitive data.

The Character Scrambling UDF

In the HomeLending database we will create a user defined function called Character_Scramble that performs character scrambling, which is shown in Listing 8-3. It will be referenced as needed in views and stored procedures that are selected to use this method of data obfuscation.

Included in this user defined function is a reference to the vwRandom view that was created in Listing 8-1. In essence, this user defined function will loop through all of the characters of the value that is passed through the @OrigVal argument replacing each character with other randomly selected characters from the same string. For example, the value of “John” may result as “nJho”.

In order for the appropriate users to utilize this user defined function permissions must be assigned. The GRANT EXECUTE command is included in the following script.

Listing 8-3: The character scrambling UDF.

This user defined function takes advantage of system functions such as DATALENGTH which provides the length of a value, SUBSTRING which is used to obtain a portion of a value, REPLACE which replaces a value with another value and LOWER which returns the value in lowercase characters. All are valuable to string manipulation.

This UDF will be referenced in any views and stored procedures that are selected to use this method of data obfuscation, an example of which we’ll see in the next section.

Repeating Character Masking

Over recent years the information that is presented on a credit card receipt has changed. In the past, it was not uncommon to find the entire primary account number printed upon the receipt. Today, this number still appears on credit card receipts; but only a few of the last numbers appear in plain text with the remainder of the numbers being replaced with a series of “x” or “*” characters. This is called a repeating character mask.

This approach provides a level of protection for sensitive data, rendering it useless for transactional purposes, while providing enough information, the number’s last four digits, to identify the card on which the transaction was made.

In the HomeLending database, we will create a user defined function called Character_Mask, as shown in Listing 8-4, which performs repeating character masking, which again can be referenced as needed in views and stored procedures that are selected to use this method of data obfuscation.

This user defined function will modify the value that is passed through the @OrigVal argument and replace all of the characters with the character passed through the @MaskChar argument. The @InPlain argument defines the number of characters that will remain in plain text after this user defined function is executed. For example, the value of “Samsonite” may result in “xxxxxxite”.

In order for the appropriate users to utilize this user defined function permissions must be assigned. The GRANT EXECUTE command is included in the script.

Listing 8-4: The Character_Mask UDF.

This user defined function takes advantage of system functions such as DATALENGTH, which provides the length of a value, and REPLICATE which is used to repeat a given character for a defined number of iterations. Both of these are valuable to string manipulation.

To illustrate the use of this, and our previous Character_Scramble user defined function, to present data in a masked format to the user, we will create a view in the HomeLending database, called vwLoanBorrowers, for the members of the Sensitive_high and Sensitive_medium database roles.

This view, shown in Listing 8-5, will present to the lender case numbers, using the Character_Mask user defined function, and the borrower names using the Character_Scramble user defined function.

Listing 8-5: The vwLoanBorrowers View.

The vwLoanBorrowers view, without the use of the masking user defined functions, would have returned the data set shown in Table 8-1.

Lender Case Number

Borrower Name

9646384387HSW

Damion Booker

8054957254EZE

Danny White

Table 8-1: The non-obfuscated result set.

However, with the user defined functions in place the masked data set shown in Table 8-2 is returned:

Lender Case Number

Borrower Name

XXXXXXXXX7HSW

o akdenbimr

XXXXXXXXX4EZE

ni ahtydwe

Table 8-2: The results returned after character masking and scrambling.

Numeric Variance

Numeric variance is a process in which the numeric values that are stored within a development database can be changed, within a defined range, so as not to reflect their actual values within the production database. By defining a percentage of variance, say within 10% of the original value, the values remain realistic for development and testing purposes. The inclusion of a randomizer to the percentage that is applied to each row will prevent the disclosure of the actual value, through identification of its pattern.

In the HomeLending database, we will create a user defined function called Numeric_Variance that increases or decreases the value of the value passed to it by some defined percent of variance, also passed as a parameter to the function. For example, if we want the value to change within 10% of its current value we would pass the value of 10 in the @ValPercent argument.

A randomizer is added through the use of the vwRandom view that we created earlier in this article. This will vary the percent variance on a per execution basis. For example, the first execution may change the original value by 2%, while the second execution may change it by 6%.

The script to create this Numeric_Variance function, which can be referenced as needed in other views and stored procedures, is shown in Listing 8-6.

Listing 8-6: The Numeric_Variance UDF.

To employ this method of masking in a development database, simply use an UPDATE statement to change the column’s value to a new value, using our Numeric_Variance function, as shown in Listing 8-7.

Listing 8-7: Updating a development database to use numeric variance.

Nulling

The process of nulling is the replacement of sensitive data with a NULL value, thus rendering the sensitive data unavailable in the development database. While this certainly protects the sensitive data, since the values are no longer known in the database, it does present issues if there are dependencies upon this data or constraints that do not permit a NULL value. Also, use of nulling can also present difficulties when trying to troubleshoot issues that specifically involve sensitive data.

To employ this method of masking in a development database, simply use an UPDATE statement to set the column’s value to NULL, as shown in Listing 8-8.

Listing 8-8: Nulling a database column.

Truncation

Truncation is a method of protecting sensitive data where a portion of its value is removed. The concept is very similar to the repeating character masking covered earlier except that rather than replacing values with a “mask”, such as an “x” or “*”, truncating simply discards those values. For example, a Social Security Number, “555-86-1234”, that is stored in plain text might be truncated to the value of “1234”.

One way to apply this method is to permanently modify the stored value in the database by executing an UPDATE statement using the LEFT, RIGHT or SUBSTRING system function to define the remaining portion of the value.

For example, the script in Listing 8-9 uses the LEFT function to truncate all but the last four digits from the Identification_Value column.

Listing 8-9: Permanently truncating the Identification_Value column.

Alternatively, in order to maintain the original value but perform the truncation for viewing, we can simply reference the column in views and stored procedures that use the LEFT, RIGHT or SUBSTRING system functions to define the remaining portion of the value. For example, Listing 8-10 returns only the last four digits of the values of the Identification_Value column:

Listing 8-10: Returning a truncated value.

Encoding

Encoding is a technique in which a series of characters is used to represent another value. This technique can be used to camouflage sensitive data, since the code used has no meaning outside the system in which the code is defined.

NOTE:
There are many benefits to encoding, beyond securing sensitive data, such as overcoming language barriers when working in an international environment and providing an expedient means of entering data.

Encoding is a practice that is found in abundance in the health care industry. The World Health Organization maintains the International Classification of Diseases (ICD), which is an industry standard that defines codes that represent diseases and health problems. These codes are used in health records and death certificates. For example, the ICD code for bacterial pneumonia is J15.9.

In the establishment of foreign keys in the HomeLending database we have, at a basic level, implemented encoding. The Loan table, for example, contains two columns that are named Purpose_Type_ID and Mortgage_Type_ID as illustrated in Figure 8-1

845-image001.png

Figure 8-1: Loan table with Foreign Key Relationships

The Purpose_Type_ID and Mortgage_Type_ID columns are foreign keys to the Purpose_Type table and Mortgage_Type table. These tables contain, respectively, the list of potential purposes for a loan and the types of available mortgage, as defined in the Uniform Residential Loan Application, developed by the Federal National Mortgage Association, commonly known as Fannie Mae.

In the Purpose_Type table, we have used a sequence of numbers to indicate these purposes. So, for example, when a new loan record is created, the value of “2” is captured instead of the value “Refinance”.

To further enhance this encoding, we may choose to either utilize a higher starting number in our sequence, such as “5000”, so that the options can be organized into logical groups. For example, we may have various types of refinance options for our borrowers. Through a higher starting number we could use the value range of 5000 through 5100 to represent the available refinance options, while construction loans might be found in the 2000 through 2100 range.

Aggregation

Aggregation is a technique in which identifying details of data are obfuscated through its provision in a summarized format. A few examples of presenting data as an aggregation are as follows:

  • As an average: 40% of the loans originated in the HomeLending database during the past quarter were refinance loans.
  • As a calculated sum: $2.5 million in loans were originated in the HomeLending during the past quarter.
  • As a geographical statistic: The median home value in the city of Indianapolis, Indiana is $150,000.

Aggregating is a common technique used to populate data warehouses for data analysis. This not only protects the underlying sensitive data, but also reduces the storage requirements for the data.

An advantage of this approach is that the data that is provided to the user is only that which they need for their reporting and analysis requirements, so the potential for the leakage of sensitive data is greatly reduced.

A disadvantage to this approach is that if the aggregations are determined to be inaccurate, the detail data is not available to identify the cause. Another challenge to this approach is that a given aggregation may not meet everyone’s needs, resulting in requests for different views of the same aggregated data, which increases your maintenance footprint.

Within the HomeLending database, aggregation may be beneficial in the collection of the borrower’s liabilities. The current design of the Borrower_Liability table requires the capture of the monthly payment amount and remaining balance. As shown in Figure 8-2, the Borrower_Liability table is related to the Liability_Account table, which reveals the creditor and account number of the liability.

845-image003.png

Figure 8-2: The Borrower_Liability and Liability_Account Tables.

An alternate approach would be to dispose of the Liability_Account table and simply capture a single record for the loan application, indicating the sum of their monthly payments and remaining balances for all liabilities, as shown in Figure 8-3.

845-image005.png

Figure 8-3: Alternate approach for Borrower_Liability table.

This level of detail would suffice for most users of this database and would protect this sensitive information from being inappropriately disclosed. The Underwriters, who may need access to the detailed liability data for qualification purposes, would refer to the credit report data, which is stored in a separate database, to determine whether or not the borrower can be approved for the loan.

Artificial Data Generation

As an alternative to these obfuscating techniques, you can generate data using third party tools, such as Red Gate’s SQL Data Generator which can produce large volumes of artificial data, based upon regular expressions and pre-defined ranges of values.

This provides protection of sensitive data due to the artificial nature of the data that is generated. In many cases, these third party tools will not execute internal encryption and obfuscation methods. This should not discourage the consideration of artificial data generation. This is noted simply to note that you may need to consider these special methods in a separate process from the general artificial data generation process.

Summary

In this article, we explored some options that are available to protect sensitive data through obfuscation. The system functions that are available in SQL Server provide us with the tools that can be employed in our views, stored procedures and user defined functions to create an effective line of defense.

Once our sensitive data has been secured we will want to identify when an attacker has gained access to our system and is snooping around for sensitive data. In addition, we will want to identify when valid users of our database are taking actions that may be suspicious. This can be done through the auditing feature of SQL Server 2008, and through a practice known as honeycombing.

Tags: , , , , ,

  • 86997 views

  • Rate
    [Total: 3    Average: 3.7/5]
  • anabelfre

    Character_Scramble
    I think that will be more of a general solution if Character_Scramble accepts nvarchar and use LEN.

  • Urvashi Saxena

    Data masking tool
    Another way to obfuscate SQL Server, other DB, and flat-file data with various masking functions is with a third-party data masking tool. The one I worked with, IRI FieldShield (www.iri.com/products/fieldshield), uses an Eclipse GUI to connect to the DB, and job wizards to select the columns and protection functions per the compliance regs for each. Resulting job scripts run from the GUI or command line, or per-column functions called via a .net SDK.

  • datasunrise

    Two types of data obfuscation or masking: static and dynamic.

    Static when creating a functional copy of production data with some of the columns could be masked or obfuscated.

    The dynamic data masking (DDM) is when data is not changed in the database but only masked when someone unauthorized trying to retrieve it. In run time

    https://www.datasunrise.com/blog/data-masking-made-simple/