Kurzbeschreibung
This thesis pays attention to the scenarios where the user data are distributed among
multiple parties (e.g., companies and institutes). Each party possesses a different set
of attributes of the same individuals. The parties aim to collaborate with others to pub-
lish a private joint dataset for better decision-making or data-driven analysis. Motivated
by the challenge of prohibition of illegal collection of private data, this thesis intends
to provide a novel deep-learning-based solution for the privacy-preserving publication
of vertically partitioned data. More specifically, this study proposes a so-called Verti-
cally Distributed General Adversarial Network (VDGAN) framework, which would be
trained jointly among multiple parties and generate synthetic joint data to replace the
integrated real one for data mining tasks. What’s more, two Differential Privacy (DP)
protocols are introduced to the framework to provide strong privacy guarantees. Ex-
tensive experiments are conducted to evaluate the proposed framework in terms of
data utility and privacy preservability with the help of four public datasets with differ-
ent attribute domains. The results present an apparent improvement compared with
the state-of-the-art tree-based models, while some limitations for practical applications
should also be considered.
Betreuer
Michael Dötzer
Dateien
| View file | ||||
|---|---|---|---|---|
|