carmenmarcos00 / ml_anonymization

Analyzing the performance of machine learning models on anonymized data

Analysis the performance of machine learning models on anonymized data

The large amount of open data available makes it necessary to study and develop tech- niques that guarantee its security for processing and analysis. Specifically, the study of anonymization techniques focuses on analyzing the distribution of the quasi-identifiers and sensitive attributes in a database. There are numerous techniques that can be applied, each of which can prevent different types of attacks. The present study explores three classical anonymity techniques, their theoretical basis and the kind of attacks they prevent: k-anonymity, ℓ-diversity and t-closeness. Specific- ally, different tools are used to ensure the reliability of these techniques which are applied at various levels on two open-access datasets, after pre-defining different hierarchies for the quasi-identifiers. Next, the performance of a battery of machine learning models applied on the anonymized data is studied. A wide range of experimental results is carried out, varying the anonym- ization technique employed, as well as the level established. All the code developed is written in Python and is distributed through an open source repository. In addition, the datasets were anonymized using the ARX Software. Keywords: anonymization, machine learning, performance analysis, privacy, k-anonymity.