How to De-identify Data Under the Australian Privacy Act

De-identify data the OAIC way: remove direct identifiers, then alter postcodes, ages and rare values or control access until re-identification is unlikely.

By · · 11 min read

Illustration for How to De-identify Data Under the Australian Privacy Act

Australia edition. Written for Australia: the Privacy Act 1988 test of a "reasonably identifiable" individual, the OAIC's two-step de-identification process and motivated intruder test, the OAIC and Data61 Decision-Making Framework, and a worked example with Victorian postcodes and age ranges. US edition

You have a customer or patient extract with names, Medicare numbers, postcodes and dates of birth, and someone wants it for research, a vendor or an open-data release. Under the Privacy Act 1988 (Cth), information that has been properly de-identified is no longer personal information, so the Australian Privacy Principles stop applying to it. The Office of the Australian Information Commissioner (OAIC) sets the bar: the risk of an individual being re-identified must be very low in the context where the data will be released, meaning there is no reasonable likelihood of re-identification. Deleting the name column does not get you there, and hashing an identifier only obscures it.

This edition follows the OAIC's guidance step by step with a small table and the Find Duplicate Entries tool. It is practical guidance, not legal advice; the OAIC itself recommends specialist help for rich datasets, public releases and multi-entity data sharing.

Personal or de-identified: the Australian test

The Privacy Act defines personal information as information "about an identifiable individual, or an individual who is reasonably identifiable". De-identified information is information that has been through a de-identification process and no longer falls within that definition. The OAIC's guidance turns this into two conditions. An individual is still reasonably identifiable where:

  1. it is technically possible to re-identify them, from the data itself or combined with other information available in the data access environment; and
  2. there is a reasonable likelihood of that happening.

So the same table can be personal information when published openly and de-identified when shared inside a secure research environment. The OAIC says it is not always possible to draw a bright line, and calls de-identification "a risk management exercise, not an exact science". Australian practice also uses other words for the same idea, such as anonymisation and the Australian Bureau of Statistics' "confidentialisation", so agree on terms before you agree on a release.

The OAIC's two steps

The OAIC describes de-identification as two steps:

  1. Remove direct identifiers: name, address, Medicare number, email, phone, customer number.
  2. Then do one or both of:
    • alter the data: remove or change information that could identify someone because of a rare characteristic or an unusual combination; and/or
    • control the environment: put safeguards on who can access the data, what they can do with it, where and how.

The OAIC notes that environment controls can reduce re-identification risk more effectively than modifying the data, and keep more of its usefulness. Its examples of data techniques include:

TechniqueWhat the OAIC describes
SamplingRelease only a fraction of records, so nobody can be sure a given person is in it
Choice of variablesDrop quasi-identifiers such as significant dates, profession or income
RoundingPut values into categories, such as age ranges instead of single years, with open-ended top and bottom ranges
PerturbationChange values slightly so totals hold but originals cannot be known
SwappingExchange a rare value between similar people in different areas
Synthetic dataGenerate new records that keep the patterns but describe nobody real
Encryption or hashing of identifiersObscure an identifier, usually to link datasets, rather than remove it

The last row matters. A hashed Medicare number or email is still the same value every time, so anyone who can guess the input can match it. The OAIC lists hashing as a way to obscure identifiers for linkage, not as de-identification on its own.

The motivated intruder test

The OAIC says that, as a minimum, you must apply the motivated intruder test: could a reasonably competent, motivated person with no specialist skills, who has access to the data, the internet and all public documents and makes reasonable enquiries, identify anyone? For data released publicly it adds a stronger test, re-identification "in the round": could anyone in the world, including experts or people with special knowledge, identify someone from the data on its own or combined with other data?

Worked example: postcodes and age ranges

Eight rows from a Melbourne clinic, with names and Medicare numbers already deleted. The quasi-identifiers are postcode, year of birth and sex; the condition is the sensitive value.

PostcodeBornSexCondition
30561984FAsthma
30571986FDiabetes
30651981FAsthma
30561983MHypertension
30681987MAsthma
30571972MDiabetes
30651975MMigraine
30561978FHypertension

Measure the groups. Write each row's quasi-identifiers on one line, for example 3056, 1984, F, paste the eight lines into Find Duplicate Entries, turn off "Only duplicates" and keep "Show count in result" on. Every line appears once, so every person is unique: k = 1.

Round, the OAIC way. Turn year of birth into an age range for 2026 (ages 39 to 43 become 35-44, 45 to 54 become 45-54) and keep only the first three digits of the postcode (305*, 306*). Run the tool again:

GroupRows
305*, 35-44, F2
306*, 45-54, F1
305*, 35-44, M1
306*, 35-44, M1
305*, 45-54, M1
306*, 45-54, M1
305*, 45-54, F1

Six of seven groups still hold a single person. In a real suburb, someone who knows a 45-to-54-year-old woman in 306* attends this clinic could read off her condition.

Choose which variable to give up. Two options:

  • Drop sex, keep the postcode prefix: groups of 3, 2, 2 and 1. One person, a man aged 35-44 in 306*, is still unique, so k = 1.
  • Keep sex, replace postcode with the state (VIC): four groups of 2 each, so k = 2, and no rows are lost.

Here the postcode, not sex, is what splits the groups, so it goes. Whether a group of 2 is enough depends on the release: for a public file it rarely is, while inside a controlled environment with access agreements it may be. That is the OAIC's point about context.

Check the sensitive column. Each group of two holds two different conditions (asthma and diabetes, asthma and hypertension, hypertension and asthma, diabetes and migraine), so knowing someone's group does not reveal their condition. If a group shared one condition, the OAIC's "attribute disclosure" risk would apply even with k above 1.

The framework to use for real releases

For anything beyond a small internal share, the OAIC points to the De-Identification Decision-Making Framework, released in September 2017 by the OAIC and CSIRO's Data61 as an adaptation of the UK's Anonymisation Decision-Making Framework. The OAIC notes parts of it may now be out of date, and also points to the Guide to Getting Started with Anonymisation, published on 1 August 2025 by privacy authorities from Australia, Hong Kong, Macao, Canada, Japan, South Korea, New Zealand and Singapore, which sets out a five-step approach.

Two Australian obligations sit alongside: under APP 11.2, an entity that no longer needs personal information must take reasonable steps to destroy or de-identify it, and APP 11.3, added by the Privacy and Other Legislation Amendment Act 2024, says reasonable steps to secure personal information include technical and organisational measures; the OAIC's APP 11 guidelines say it applies to personal information held from 11 December 2024.

These duties bind APP entities: most Australian Government agencies, organisations with an annual turnover of more than $3 million, and some other organisations, as the OAIC's Privacy Act page puts it. A small business should check whether it is covered; de-identifying well is good practice either way.

Tools for this job

  • Find Duplicate Entries: count how many rows share each quasi-identifier combination, which is the k in k-anonymity.
  • Hash Generator: see why a hashed identifier is consistent and therefore linkable, not de-identified.
  • Remove Duplicate Lines: list the distinct groups once you have rounded the data.

For why hashes are one-way but not secret, read hash algorithms compared.

Frequently asked questions

Is de-identified data covered by the Privacy Act?

No. The OAIC says information that has gone through an appropriate and thorough de-identification process is not personal information and is not subject to the Privacy Act. The catch is that the test is contextual: the same data can be personal information if released somewhere re-identification becomes reasonably likely.

Is hashed data de-identified in Australia?

Not on its own. The OAIC lists encryption or hashing of identifiers as a technique that obscures an identifier, usually to link datasets, rather than removing it. If the hash can be matched back to a person, the data is still about a reasonably identifiable individual.

What is the motivated intruder test?

It asks whether a reasonably competent, motivated person with no specialist skills, with access to the data, the internet and public documents, could identify someone. The OAIC treats it as the minimum test; for public releases it also asks whether anyone, experts included, could do so.

Do I have to de-identify data I no longer need?

Under APP 11.2, an APP entity that no longer needs personal information for any permitted purpose must take reasonable steps to destroy or de-identify it, though the OAIC notes this does not apply to personal information in a Commonwealth record.

Sources and further reading

Browse all tools