Making Personal Data Anonymous

Governments hold lots of personally identifiable and commercially sensitive information. This sensitive information necessarily restricts agencies’ ability to share this information as open data. This article will go some way to introducing you to the option that you have to make it anonymous. Sometimes, just removing fields from a row is insufficient. It can be possible to use statistical techniques to identify individuals. This is especially the case when your data are combined with other sources of information.

Defining personal information

What counts?

This question is more complicated than what one would initially believe. Every jurisdiction has its own requirements and and every culture has its own norms about what counts as personal and private.

Things to look out for


Any number or other value that is used by a computer to identify individuals can be examined by an attacker. Typical identifiers include registration numbers, ID numbers, passport numbers, credit card numbers, IP addresses, and order numberr.

Very rare characteristics

Outliers are very easy to identify in a statistical manner.

Very specific descriptions

Specific descriptions make it very easy to identify individuals. As specificity increases, the number of individuals possessing a characteristic decreases. For example, there are fewer French than Europeans. While the specific information can be valuable, be wary of its effects in very small population segments.


This can include photos of individuals, their property and other items of interest.

Biometric data

Biometric data are collected specifically to identify individuals. Therefore, it is highly likely that they will be sensitive. Biometric data includes fingerprints, retina, DNA, height, body markings. It may also include samples of handwriting.

Free text

Respondents’ free text responses very can often identify who has spoken. Free text can be subject to word frequency analysis and other computational linguistic analysis.

Strategies for making data anonymous


When data are aggregated, they unable to be used to identify the sources of the data. That is, if we provide the mean household income for a street, we will not be able to identify the household income of any particular family. The downside of this approach is that aggregating data too far will impair analysts’ ability to interpret the data.


A very simple approach to privacy protection. Here, we simply remove some field of interest that was originally collected. For example, we could omit gender, age, location or any other variable that has been collected.


Dithering results means to add a variation to every value within a sample, while attempting to maintain the integrity of the aggregate values. The goal is to prevent the the true value for any specific value to be deduced, but to enable statistical analysis to be carried out. For example, for geographic data, you could move points of interest to a random location within a given radius.

Top and Bottom Coding

Top and bottom coding means to replace extreme values of sensitive numerical variables with the weighted group mean for those values in order to mask outlying values which are potentially identifying. For example if the data is about the airline industry in New Zealand results from Air New Zealand might be replaced with the weighted group mean for the values in the group of data in order to ensure that the results from Air New Zealand could not be identified.

From the OECD:

“It consists in setting top-codes or bottom-codes on quantitative variables. A top-code for a variable is an upper limit on all published values of that variable. Any value greater than this upper limit is replaced by the upper limit or is not published on the microdata file at all. Similarly, a bottom-code is a lower limit on all published values for a variable. Different limits may be used for different quantitative variables, or for different subpopulations.”


We can group multiple values together to protect individuals’ privacy. Consider the following table, with the transformation appearing in the right-hand column.

132 cm 139.67 cm
143 cm 139.67 cm
144 cm 139.67 cm
144 cm 152 cm
153 cm 152 cm
159 cm 152 cm
161 cm 164 cm
167 cm 164 cm

There may be problems with this approach, as you will impact on the median and other percentile values.

Hash digests

Using a crypographic hash of a string can make it impossible for someone to determine what the original string was, while being able allow 3rd parties to check if strings they have are included. As they can apply the hash function to their own values, they can undertake comparisons without being able to access data that they don’t already have. The transformation looks something like this: c242dbe863aa0a38eacc72888fd41804 a99650df0d55169e0d9f1dc17194830f