MMOOZZGG.com

Glossary

What is data anonymization?

In short

Data anonymization is the processing of information about people so that nobody can tell whom it concerns without extra information.

Security and law · anonymization

MMOOZZGG.com editorsChecked: 1 min read

For a child

A teacher wants to show the class the most common mistakes in a dictation but does not want to shame anyone. The teacher covers the names on the notebooks and hands them out mixed up. You can see the mistakes, but you cannot guess whose work it is. People do the same with information about people when they give it to a smart program: they remove everything that could identify a person.

In plain words

To train a model or send a conversation to a neural network, you often do not need names and phone numbers. So everything that points to a person is removed or replaced: full name, phone number, address, contract number. Labels or general values take their place, for example an age group instead of an exact date of birth. The key that turns labels back into names is stored separately and locked away. Removing the name alone is not enough: a person is sometimes recognized by a rare mix of city, job title and purchase date.

For business

It is the main way to let AI work with customer data without breaking 152-FZ, Russia’s personal data law. You can analyze, label and send anonymized records to a model with less risk. Companies use it in request analytics, model training, test databases for contractors and work with external neural networks. The cost depends on the volume and type of data. Tables are easier to process, while free text in emails and call recordings is harder and needs a separate tool to find names and numbers. The main risk is re-identification, when someone recognizes a person by a mix of the remaining fields or by matching with another database. Before launch, check who holds the key to the labels and whether anyone can restore identities from an export. These checks are part of a review of data leaks through neural networks.

For a specialist

Under clause 9 of article 3 of 152-FZ, depersonalization means actions after which, without extra information, it is impossible to tell which person the personal data belongs to. In data protection terms, this is closer to pseudonymization than to irreversible anonymization. Methods: removing direct identifiers, replacing them with tokens tied to a key, generalizing and suppressing quasi-identifiers, shuffling and adding noise. Formal models include k-anonymity, where each record cannot be told apart from at least k-1 others by quasi-identifiers. Another is differential privacy, which limits the effect of one record on a query result with calibrated noise (Dwork, Roth, 2014). Narayanan and Shmatikov used the Netflix Prize dataset, with ratings by about 500,000 subscribers, to show that a small amount of information about a person finds that person’s record when matched with public ratings on IMDb. For texts and dialogs, named entity recognition (NER) and masking run before data goes to an external API.

Example

A call center wants a neural network to find the common reasons for complaints in call transcripts. Before sending, a program replaces names, phone numbers and addresses with labels, and the mapping table stays on the company’s server.

Related terms

Questions and answers

How is depersonalization under Russian law different from strict anonymization? Under 152-FZ, Russia’s personal data law, depersonalized data can be linked back to a person with extra information, such as the key to the labels. Anonymization in the strict sense cannot be reversed: you cannot link the record to a person at all.
Can you send anonymized data to an external neural network? The risk is much lower, but check that the text has no indirect clues, such as a rare job title or an address. The key to the labels must never leave together with the data.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us