Methodological News, June Quarter 2025

As part of our ongoing commitment to protect the privacy of data providers, the ABS is investigating a constrained optimisation approach to tabular suppression as we move to modernise our business statistics production. The ABS currently applies an internally developed graphical method to determine effective suppression patterns, but we are now reassessing its suitability as more contemporary methodologies become available. 

As a potential alternative suppression method, we are considering a software developed by Statistics Canada called G-Confid. G-Confid builds upon the Controlled Tabular Adjustment algorithm (Cox, 2005). For each table cell subject to primary suppression, G-Confid identifies the optimal locations for complementary suppression such that loss of information is minimised, using linear programming techniques to represent table additivity.  

G-Confid offers several potential improvements over the existing methodology, including: 

• the ability to process multi-dimensional collections holistically, which is increasingly important as big data continues to grow; the existing ABS method can only process collections in two-dimensional slices
• the inclusion of a post-treatment procedure to reduce the number of redundant complementary suppressions
• the extensibility of constrained optimisation techniques, such as handling non-linearity, which may allow more complex data to be protected as the demand for more detailed statistics increases. 

Another constrained optimisation method we are exploring is the Attacker Model (Fischetti & Salazar, 2001), as a means of auditing suppression patterns. The model aims to recalculate the possible values of suppressed cells, again using a linear programming framework to describe additivity between table cells. If a cell can be estimated within close range of its true value, it is considered a disclosure risk.  

Using an implementation of the Attacker Model in Python, we are comparing the efficacy of G-Confid against the existing graphical suppression method. In terms of both data utility and protection levels, initial findings are promising. Our next steps are to expand upon test cases and to address computation time and model complexity for large tables. 

For more information, please contact Zanya Barns. 

Differential privacy (DP) is a rigorous privacy framework that quantifies the risk of privacy leakage for database contributors. The ABS has explored the opportunities offered by DP for use in our cross-sectional data products and now looks to explore its suitability for time series data. This will enable us to develop clear confidentiality metrics for contributors to temporal datasets, such as business statistics releases.  

There are many challenges to be overcome when confidentialising temporal data, including: 

• an increase in perturbation required for confidentiality, due to correlation between sequential data points.
• multiple levels of privacy leakage for contributors to time series data - a privacy mechanism may only need to protect a contribution at a single time step or across an arbitrary number of time steps. This affects the perturbation required for confidentiality and introduces a policy decision to the framework.
• time series analysis needing to preserve correlation between data points - this means perturbation must preserve more relationships within the data, in comparison to tabular perturbation.
• the dominance of a small numbers of contributors across some business statistics - this increases the perturbation required for protecting contributors’ confidentiality, reducing the quality of statistics.
• data being subject to revision, which must also be protected by any privacy mechanisms used.  

One differentially private mechanism for time series is the Fourier Perturbation Algorithm (FPA). This mechanism works by first transforming the data using the discrete Fourier transform. The data is then perturbed before being transformed back into its original domain. This process results in a smoother output with less noise and a sinusoidal structure. As such, this algorithm is suitable for seasonal data, which also has sinusoidal structure, but the FPA is not suitable for trended data, since trend preservation requires increased perturbation. 

We are currently investigating the suitability of a privacy mechanism which detrends the time series data before applying FPA. We aim to see if such an approach will preserve utility, including autocorrelation of the data, whilst also providing a suitable level of confidentiality. This will enable us to release more useful information to maximise the value of the data.  

For more information, please contact Matt Giegerl. 

Please email methodology@abs.gov.au to:

• contact authors for further information
• provide comments or feedback
• be added to or removed from our electronic mailing list.

Alternatively, you can post to:

Methodological News Editor
Methodology Division
Australian Bureau of Statistics
Locked Bag No. 10
Belconnen ACT 2617

The ABS Privacy Policy outlines how the ABS will handle any personal information

Releases from June 2021 onwards can be accessed under research.

Releases up to March 2021 can be accessed under past releases.