|
|
|
Title: Privacy Preserving Machine Learning: A Theoretically Sound
Approach
Speaker: Wang Liwei(Beijing University)
Abstract: Privacy is an important concern in the big data era. The
potential benefit for scientific study from these data is huge. But
how can the database holder release sensitive data while preserving
individual privacy? In this talk I will review a recent rigorous
definition of privacy: differential privacy. Differential privacy
guarantees that there is almost nothing new can be learned from a
database if an individual contributes her data compared to that can
be learned from the same database except her data is not in; and
thus there is no harm for an individual to contribute her data.
Previous algorithms on differential privacy are usually inefficient.
In fact, it can be shown that answering general queries while
preserving differential privacy is computationally hard. I will give
a very efficient algorithm (sublinear time in many parametric
settings), which can answer a broad class of queries of high
practical interest.
|
|