Settings

The dataset

The file you are about to publish, or hand over, or attach to an answer. It is read by this tab and goes nowhere, which is the only sensible way to handle a file precisely because you are not yet sure what is in it.

Is anybody alone

The combinations too few people share

What to do about it

The check, written out

What you marked as what, what came back, and what you were offered. This is the thing to keep on the file: a release nobody can show they checked is indistinguishable from one nobody checked.

How it works

CROWD checks a CSV for re-identification risk using k-anonymity, which finds people who are alone in their group, and l-diversity, which finds groups where everyone shares the same sensitive value. A good score reduces risk but does not prove a release is safe. Your data stays in this tab.

The guide: Check a dataset before you publish it

Questions people ask about CROWD

What is k-anonymity, in plain terms?

Group every row in the table by the combination of the columns an outsider could already know about a person: their age, their postcode, their sex, a date. The number of people in the smallest group is k. If k is one, at least one person is unique on that combination and can be picked out by anyone holding it, along with everything else on their row. If k is five, the smallest crowd anybody is standing in has five people in it and the best an attacker can do from those columns is narrow it to five. The probability of picking the right person out of the smallest group is one in k, which is the whole idea.

How do I know which columns are quasi-identifiers?

Ask what somebody trying to find a particular person could already know or look up about them. Age, date of birth, postcode, sex, ethnicity, job title, employer, the date they were admitted or arrested or hired, height, nationality. It is not about whether the column is sensitive; it is about whether it is knowable. Getting this wrong in either direction is the main way the answer becomes meaningless, which is why the page guesses and then asks you to check rather than deciding quietly.

What is l-diversity and why do I need it as well?

k tells you the crowd is big enough; it says nothing about whether the crowd is varied. If all eight people in a crowd share one diagnosis, an attacker who narrows somebody to that crowd has learned the diagnosis without needing to identify them at all: the homogeneity attack. l-diversity counts the distinct values of the sensitive column inside each crowd and reports the smallest count. An l of one means at least one crowd gives the answer away, however large k is.

What thresholds should I use?

There is no universal right answer, which is why the target is yours to set rather than fixed. Three is sometimes used for internal sharing, five is a common general figure, and some regimes use higher. The number that matters is the one your own guidance, your information rights team or your data protection officer sets, and this tool is there to tell you honestly whether you are meeting it, not to tell you what it should be.

Does a good k mean the release is safe?

No, and the page says so in as many words. k says nothing about a release that still contains a name or a reference number; nothing about what an attacker knows from a source outside your table; nothing about what two of your own releases reveal when compared; and nothing about hidden sheets, pivot table caches, or the author name in the document properties, which is a different and very common way that releases go wrong. This is one check done properly, not a certificate.

What about spreadsheets rather than CSV?

Save the sheet you are releasing as CSV and check that. That is also good practice in itself: a great many disclosure failures are not about the visible data at all but about what an .xlsx file carries around with it invisibly, and converting to a flat format is one of the standard ways of shedding it. For the hidden content of a spreadsheet or document, use the tools here that look specifically for it.

Is anything sent anywhere?

No. The file is read by the browser’s own file reader, all the counting happens in the page, and the download is written by the tab. There is no account and no server. The dataset is not kept across a reload (only your column choices are) because it is not ours to keep.

Doing this often? OBSCURA Gnosis analyses the data set before and after release: descriptives, tests and reliability, with tables saved to Word. On its own, US$2 a month; with everything else, US$5 a month or US$48 a year.