Perform quantitative tests

[TTitcombe]

Prior work introduced a quantitative test for model inversion attack:

• train another classifier on a different subset of the data (unseen by any other model at any point)
• Attack success accuracy is accuracy of the "comparitor model" on recreations