Measure performance hit Vs Differential Privacy

[anupamme]

DP gives a good defense but it affects the performance at the main task. We want to measure this relationship.

I tested with DP: Gaussian (0, 0.01) and it acted like a good defense. But we dont know the performance hit on the main task.

The goal of this issue is to track this performance hit Vs the noise we add.

So what we want to track (for varying values of noise):

1. Training Loss
2. Training Accuracy
3. Test Accuracy