Implementation of orthogonal/inverted matrix-based homomorphic encrpytion for somewhat-encrypyted machine learning
For experimental purposes only
Taking the UCI credit default dataset, we built a benchmark classification model (~75%).
Then encrypted the dataset using a set of matrix transformations based on the homomorphic encryption schemata here.
Running a backpropagation neural network model on encrypted data yielded similar accuracy (~74%) to the vanilla model on non-encrypted data, indicating no loss of insight/pattern during encryption.
This is a CUDA implementation