This data was de-identified, not anonymized. Just taking the identifiers out of a data set often isn't sufficient to anonymize it.
But there *is* anonymized data. Even leaving techniques like differential privacy aside, we can find simple and useful examples. (Thread)
But there *is* anonymized data. Even leaving techniques like differential privacy aside, we can find simple and useful examples. (Thread)
Let's take aggregate analytics data. In order to figure out how much server power I'm going to need to serve my site next month, I need to know about how traffic I have now. Say today I had 55 million people. That number is made from user data, but is anonymous.
Anonymization is a valuable technique in our technical privacy arsenal, but not a simple one.
De-identification is, perhaps surprisingly, also useful in certain situations. In some cases one needs the data to remain non-anonymous (for example, I really do not want Google to anonymize which apps I bought in the Play Store. I would like to use those apps TYVM 😉)
... but that doesn't mean that everything that is done with those records needs to have all the information in them. When training a ML model to recommend apps it doesn't necessarily need to know who I am, just what stuff I've got. De-identification is good practice there.
