Description
Linear regression is a standard machine learning technique, but using common
approaches to perform regression on shared data of multiple parties leaks
information about private inputs. We address the problem of
privacy-preserving linear regression allowing for collaborative computation
while minimizing the information revealed about the individual data sets. We
describe two schemes based on stochastic gradient descent and gaussian
elimination to perform linear regression iteratively and non-iteratively.
Both schemes maintain the privacy of the input data by using multiparty
homomorphic encryption. The evaluation of the approaches on synthetic data
sets shows that both approaches can deliver a sufficiently precise result in
a reasonable time. However, both methods have limitations regarding their
applicability depending on the data set size and distribution.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching