Replace $\max_\delta \mathcal L(f_\theta(x + \delta), y)$ with $\max_{g\in \mathcal G} L(f_\theta(g(x), y))$ Example: filter based models where $g$ is a piece-wise linear function.
Replace$\max_\delta \mathcal L(f_\theta(x + \delta), y)$ with $\max_{g\in \mathcal G} L(f_\theta(g(x), y))$
Example: filter based models where$g$ is a piece-wise linear function.