In the cosmetic industry it is crucial to undergo a risk assessment of all products placed onto the market to ensure they are safe for consumers and in particular non-sensitizing. Bayesian networks are increasingly used to predict potential sensitization but only from internal proprietary databases. The goal of this project was to improve prediction performance by pooling information from the main players in cosmetics. However, confidentiality issues arise when sharing internal data. Federated learning with differential privacy (FLDP) is a promising learning approach that addresses privacy and data security concerns. A proof of concept using FLDP in the framework of Bayesian networks methodology has been conducted showing promising results and could potentially become a new standard for assessing the risk of sensitization.
