Security overview and performance assessment of fully homomorphic encryption in machine learning

In the era of cloud computing and third-party data processing, protecting data during computation remains a major challenge, as traditional encryption secures data only at rest and in transit. Fully homomorphic encryption (FHE) addresses this limitation by enabling computation directly on encrypted data. This paper evaluates the performance of FHE for machine learning using concrete ML, a library designed for FHE-based models. Ten machine learning and deep learning algorithms were implemented using both scikit-learn and concrete ML to compare training time, execution time, and accuracy. While FHE enables execution on encrypted data, only one of the tested algorithms supports training on encrypted data; the remaining models require plaintext training. Results indicate a significant computational overhead for FHE-based models, whereas accuracy remained comparable, with deviations typically below 1% and never exceeding 5%. Despite current performance limitations, FHE offers inherent resistance to quantum attacks and strong potential for privacy-preserving machine learning.