Data anonymization

Data security and confidentiality are at the heart of organizations’ concerns today. The challenge is to find a compromise between valuing data and the risk of using personal data for analytical and artificial intelligence projects.

Data governance shares the responsibility with other teams, namely with cybersecurity experts, to put in place the necessary measures to maintain granular control over the access rights of each employee. For example, it defines access levels by role directly on the data. However, the implementation of these measures and the resulting processes restricts access to the data, regardless of what you want to do with it. This approach also reaches its limits when it comes to giving partial access to databases, as the risks of inference and re-identification increase.

Alternatives exist and constitute very active areas of research. They can be grouped into three categories (not separate):

• Data anonymization techniques (eg “k-anonymization”) aimed at making the original data less sensitive.
• Analysis designed to keep data confidential (eg “Differential Privacy”). We use the original data and we desensitize the calculations on the data.
• Generative synthetic data models

These approaches introduce the concept of confidentiality measurement and therefore a quantity associated with data risk. Given their inherent advantages and disadvantages, what approach(es) should be favored in order to optimize the compromise between the risk of confidentiality and the valuation of data?