Main question: does slightly perturbing the input data yield a drastically different risk or a different risk group. If so
- there is an additional attack vector because people could learn decision boundaries and by slightly tweaking features that do not require exact entry (like estimated miles per year) get into a better category, thus being able to hack the system
- high local variation hints towards undetected overfitting
- high local variation makes it likely that retraining with new data will yield a completely new model which also requires new interpretation etc.
Links:
Main question: does slightly perturbing the input data yield a drastically different risk or a different risk group. If so
Links: