Data Science Fellows Projects 2019

Machine learning algorithm adaptation for homomorphically-encrypted data and models

Project by Gilad

Abstract

The project goal was to efficiently apply machine-learning techniques on homomorphically-encrypted data and/or models.

Metrics used were ROC, maximal absolute relative difference in prediction probability values, algorithm run time and core utilization.

 

Challenges

  • Learning how to design algorithms abstractly
  • Parallelizing algorithms across multiple cores (Python’s multiprocessing module would not work with Duality’s algorithms).
  • Dealing with large quantities of data.

 

Achievements (according to KPIs)

  • Run-times were significantly reduced
  • The algorithms are now parallelized (where possible) and modulated.

ROC and maximal absolute relative difference in prediction probability values remained the same despite decreased run times

Share this post

Share on facebook
Share on twitter
Share on linkedin
Share on email