Fig. 6.

Illustration of surrogate interpretability methods, where g is a white box, which is trained to mimic the behavior of the “black-box”, so that $g\left(y\right)\approx f\left(x\right)$ [30]