Method and system for adaptively reducing feature bit-size for homomorphically encrypted data sets used to train machine learning models

A system, comprising:

• a processor;
• and
• a memory having instructions stored thereon which, when executed by the processor, performs an operation for training a machine learning model using a homomorphically encrypted reduced-size data set, the operation comprising:
• receiving a data set to be used in training a machine learning model and information defining properties of the machine learning model to be trained;
• determining, based on values of one or more features of the received data set and the properties of the machine learning model, a reduced number of bits to represent, in a reduced bit-size data set, one or more features of the received data set, the reduced number of bits being smaller than a number of bits representing the one or more features of the received data set;
• generating the reduced bit-size data set, the generating comprising reducing a bit size of the one or more features in the data set according to the reduced number of bits;
• encrypting the reduced bit-size data set using a homomorphic encryption scheme;
• and
• training the machine learning model based on the encrypted reduced bit-size data set.

A method for generating predictions from a machine learning model trained using a reduced-size homomorphically encrypted data set, comprising:

• receiving an encrypted input on which a machine learning model is to generate a predicted value, the machine learning model trained using a homomorphically encrypted data set having a reduced size relative to an uncompressed data set;
• generating a reduced bit-size prediction based on the received encrypted input;
• decompressing the reduced bit-size prediction based on a compression applied to one or more features in the homomorphically encrypted data set;
• and
• outputting the decompressed prediction as the predicted value from the machine learning model.