Lab Machine learning on encrypted data

This course is listed

• in Aachen RWTHonline as Lab Machine learning on encrypted data,
• in Bonn Basis as MA-INF 4322 - Lab Machine Learning on encrypted data.

Lecture

Michael Nüsken

Time & Place

Send an email to apply for participation in this lab.

We meet by appointment.

• Tuesday, 14⁰⁰-16⁰⁰, digital seminar room .
• Friday, 10⁰⁰-12⁰⁰, digital seminar room.

Kick-off meeting: Tuesday, 19 October 2021, 14⁰⁰, digital seminar room.

Presentations

For reports and presentations see Seminar Advanced Topics in Cryptography.

Contents

With the rise of more and more mechanisms and installations of data science methodology to automatically analyze large amounts of possibly privacy infringing data we have to carefully understand how to protect our data. Also more and more fake data shows up and we have to find ways to distinguish faked from trustable data. At the same time we want to allow insightful research and life-easing analyzes to be possible. This seeming contradiction has lead to various efforts for unifying both: protecting data and allowing analyzes, at least to some extent and possibly under some restrictions. See Munn et al. (2019) for a review on challenges and options.

Some methods use supervised machine learning where test data is classified manually and then used as training data. Other use unsupervised machine learning where data is classified without any manual training. The latter include clustering and outlier detection methods.

• Jäschke & Armknecht (2018) employ fully homomorphic encryption (FHE) to analyze data with real valued dimensions and perform a classical clustering algorithm on encrypted data. This is a first major step to employ unsupervised machine learning methods in a fully privacy-protected scenario. A variety of other clustering methods have not yet been investigated at all. The aim of this lab is to understand the given approach and apply it to further methods.
• Crawford, Gentry, Halevi, Platt & Shoup (2018) show how to get the best out of present FHE.
• Results from last winter.
• tbc.

The target of the lab is to understand how unsupervised machine learning on encrypted data may work. Ideally, we can come up with a novel solution for performing an unconsidered algorithm. We study the tasks and tools, select a clustering algorithm, find a protocol, prototype an implemention, perform a security analysis, present an evaluation, ... The named core papers and the experience from last winter's lab shall serve as guide how to proceed.

We will plan and distribute work shares together in our meetings.

Prerequisites

Basic knowledge of data science and privacy is helpful.
A fast understanding of mathematical and computer science topics is required.

Allocation

Lab.

• Master in Media Informatics: PR, ? ECTS.
  Students have to register this course in RWTHonline.
• Master in Computer Science at University of Bonn: MA-INF 4322, 9 CP.
  Students have to register this course in BASIS.