Page 762

Ming Ding | Information Security and

Privacy Group | Data61, CSIROVisual experiments

•What if we don’t have access to the

original dataset? Are the synthetic data

still useful?

•Explainable AI : Visual analysis on the

trained classifier using the synthetic

data only

➢Local explanation: Why do you think

this image is a “2”?

➢Global explanation: What does a typical

“2” look like?

https://www.youtube.com/watch?v=elPTCt45ZGo&list=PL9Hr9s

NUjfsmEu1ZniY0XpHSzl5uihcXZ&index=19MLP trained on synthetic data, DP -Copula -Conditional , ε=10

Good trade -off

between privacy

and utility

MLP trained on synthetic data, DP -Copula -Conditional , ε=5

MLP trained on synthetic data, DP -Copula -Conditional , ε=1MLP trained on original data