diffix / test-syndiffix Goto Github PK

@pdobacz before you start digging into anonymeter, I think it would be good to run the attack that Cristian mentioned in the meeting today.

You can find a description of the attack in section 5.20, page 35, of https://arxiv.org/pdf/2201.04351.pdf.

What you'll have to do is create a bunch of datasets with two columns each. One column is categorical with a relatively small number of distinct values (you can test 2, 5, and 10 values). The other column is the AID column. It has multiple rows per entity, but all rows for a given entity have the same categorical value. Select each AID's category randomly from the set of categories. (This selection can also be skewed, but let's not worry about that now.)

All entities are assigned some number of rows taken from a uniform distribution of some range (say between 50 and 100 rows). One outlier entity, however, is assigned substantially more rows, like 200 or even more.

Actually, you'll need to build two datasets for each test. One dataset is as described above. The other has one row per entity (and the same category assignment).

Build synthetic datasets from both tables. From the one-row-per-AID table, you estimate the number of distinct AIDs per category, and the total number of distinct AIDs. From the multi-row table, you determine to total number of rows.

You then estimate the expect number of rows per category on the assumption that all entities have the same number of rows. You then select the category that exceeds that estimate by the largest amount. (If the categories are uniformly assigned to AIDs, then this is simply the category with the most rows.)

You then guess that this category is the one assigned to the outlier AID.

Do this for a few hundred different datasets, and record the fraction of times the guess is right.

Problem is that we are pulling the elapsedTime measures from the ML measures, and many 2dim runs don't have any ML measures.