Reduce the information outputted from a model (confidence values, input size, token limits), reducing the ability for an adversary to extract information and optimize attacks for the model. Such recommendations include confidence rounding, Gradient Masking.

​
Explanation

 

​
How it works

 

​
How to implement