Create an aggregated embedding for membership inference hot spots

[suhacker1]

Is your feature request related to a problem? Please describe.
We need a clarifying visual to showcase the privacy risks of membership inference, especially as it varies between classes.

Describe the solution you'd like.
An aggregated embedding similar to Figure 4 of Understanding and Visualizing Data Iteration in Machine Learning would be useful. Highlight the worst case.

Describe alternatives you've considered.
Please comment any alternatives.