Trade-off evaluation between privacy and data utility through the application of Fully Homomorphic Encryption during ML models training for classifying misogyny content

In the context of machine learning (ML), the training of models on sensitive data can produce significant privacy concerns. Specifically, these concerns derive from the inherent nature of datasets, which often hold confidential information. Even with stringent safeguards in place, the formidable processing capabilities of ML algorithms can inadvertently lead to the identification of individuals or, in the most severe scenarios, catastrophic data breaches. In this work, we delve into the capabilities of the cryptography library ConcreteML, a groundbreaking tool designed to uphold privacy throughout the ML training by harnessing the power of Fully Homomorphic Encryption (FHE). Moreover, through a series of rigorous experiments, we explore the delicate balance between privacy preservation and data utility using real-world datasets. Our findings reveal that ConcreteML not only maintains robust privacy protections but also delivers encouraging results in terms of data utility, striking an impressive balance that showcases its potential in safeguarding sensitive information while maximizing analytical insights.