What does DP Wizard do?

DP Wizard guides you through the application of differential privacy. After selecting a local CSV, you'll be prompted to describe the analysis you need. Output options include:

• A Jupyter notebook which demonstrates how to use the OpenDP Library.
• A plain Python script.
• Text and CSV reports.

What kind of data can DP Wizard handle?

DP Wizard (or differential privacy more generally) is not an all-purpose privacy filter: It requires data of a particular form for its privacy guarantees to be valid.

• Tabular data: While DP on other kinds of data is an active area of research (graphs or text, for example), DP Wizard is limited to CSVs.
• Field / record structure: Rows should represent records, and columns should represent fields.
• Individuals, not aggregates: Further, each row should be related to a single entity whose privacy you want to protect: If the data has already been aggregated to a higher level, DP Wizard is not the right tool.
• Tables are tall...: With less than than a hundred rows, it is difficult to return useful results, while also preserving the privacy of contributors. A thousand rows or more is better.
• ... And not wide: Differential privacy assumes a finite privacy budget, and each query consumes a portion of that budget. Statistics for more columns means each statistic is less accurate. (The DP Wizard interface is another constraint: It starts to get awkward with more than half a dozen columns.)
• Know the big picture (but not the details): How many rows can an individual contribute? What are typical ranges for numeric fields? What are expected values for enumerated fields? While accurate answers are important, it's also important not to look through the data to get the answers.

Does differential privacy provide anonymization?

This is one definition of anonymization:

Anonymisation is the way in which you turn personal data into anonymous information, so that it then falls outside the scope of data protection law. You can consider data to be effectively anonymised if people are not (or are no longer) identifiable.

We use the broad term ‘anonymisation’ to cover the techniques and approaches you can use to prevent identifying people that the data relates to.

(From the UK Information Commissioner's Office Introduction to Anonymisation)

Differential privacy, used appropriately, may provide anonymization, but just as with other anonymization techniques it needs to be applied carefully.

It may be helpful to compare the vulnerabilities of different anonymization techniques: If indentifying values are dropped from datasets for anonymization, that may not be sufficient if other combinations of individual characteristics are still distinctive. In contrast, differential privacy does not have this particular vulnerability, but it does present other challenges: successive DP releases on the same dataset will expose incrementally more information and consume a larger privacy budget.

Anonymization requirements vary across legal regimes, data domains, and end uses, and any approach to anonymization should be reviewed by an expert who understands your particular requirements.

How do I choose a privacy budget?

It may be helpful to make transparent the competing needs for accuracy and privacy. Accuracy may be the easier one: The confidence margins that result from additional noise can often be precisely characterized. You should not use a larger budget than necessary for your analysis.

The privacy requirements on the other hand can not be reduced to a number. Although the epsilon value has a formal definition as a probability ratio, it embodies worst case assumptions, and doesn't take into account legal requirements, or possible harms to the individuals involved.

These tradeoffs are not unique to differential privacy: Any disclosure of information carries with it a risk of revealing personal information. One strength of differential privacy is that it makes this tradeoff more clear.

Can synthetic data from DP Wizard be used for analysis?

Synthetic data has certainly been used for analysis when individual privacy needs to be preserved. The 2020 US Census is a notable example of a synthetic data product.

However, conducting an analysis on synthetic data is not the only possibility: It can also be used as a proxy. Analyses can be prepared against the synthetic dataset, and then once the bugs have been worked out, the analysis can be run against the real data in a protected environment. Whether this is feasible will depend on your organizational constraints.

One thing to remember is that the synthetic data is derived from the contingency table, and if a question can be answered directly from the contingency table, that is usually better. Contingency tables also make assumptions like bin ranges transparent to the user.

Where can I learn more?

One place to begin is this slideshow: it provides an introduction to differential privacy and DP Wizard, with some group exercises along the way. At the end there are suggestions for further reading.

How can I report bugs or request features?

If you are not on github, feel free to email the maintainer directly. If you are on github, please file an issue. Thank you for the feedback!

If you are on Github, the form below captures environment information and pre-fills a Github issue: