Full Program
Summary:
In recent years we have witnessed a rise in the popularity of cloud-based Machine Learning that comes with indisputable benefits but also poses security risks. Privacy-Preserving Machine Learning is possible through Multi-Key Homomorphic Encryption (MKHE) that allows computations on data encrypted by different parties under different keys. The MKHE methods proposed in our previous works allow the evaluation of encrypted ensemble models over encrypted feature vectors in order to perform classification tasks. We assumed that all features in the feature vectors to be classified are encrypted but in many real-world scenarios only some features are sensitive while the others can be disclosed. In this work we enhance our previous methods to evaluate encrypted models over feature vectors containing both encrypted and plaintext values. We perform experiments on benchmark UCI datasets and demonstrate that by hiding only the sensitive features the classification time decreases by one to two orders of magnitude.Author(s):
Diana-Elena Petrean
Technical University of Cluj-Napoca
Romania
Rodica Potolea
Technical University of Cluj-Napoca
Romania