We propose to iteratively adversarially train an external auto-encoder to protect a vulnerable base classifier.
The algorithm is implemented in four frameworks, source code can be found in the following folders:
- python: using tensorflow v1 and keras, most complete implementation including BPDA
- tf2: using tensorflow v2
- pytorch: using pytorch
- julia: using Julia and Flux, best code quality, recommanded
License of code: MIT