System and method for machine learning architecture with differential privacy

Abstract

Differential private dictionary learning privatizes input data by training an autoencoder to learn a dictionary, the autoencoder including an encoder and a decoder, and weights of channels in a layer in the decoder defining dictionary atoms forming the dictionary; inputting the input data to the trained autoencoder; projecting, using the encoder, the input data on the learned dictionary to generate a sparse representation of the input data, the sparse representation including coefficients for each dictionary atom; adding noise to the sparse representation to generate a noisy sparse representation; and mapping, using the decoder, the noisy sparse representation to a reconstructed differentially private output.