Feature Request: Adversarial Examples on Discrete Sequences about foolbox

[jonasrauber]

Contributions in that direction are welcome.

Some aspects are already possible with Foolbox, e.g. attacking non-differentiable models using our decision-based attacks and enforcing discrete inputs (by adding a rounding layer before the actual input).

Attacking only a subset of the input is also already possible by modifying the model such that it get's a constant input (that's part of the model) and a modifiable input that will be fed to the attack. It might not be the most convenient way of doing this, but it should work well.

[jonasrauber]

Closing this because there are no plans to work on this and it's somewhat unclear what exactly would be needed.