Adversarially trained auto-encoders

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