Towards Robust Stacked Capsule Autoencoder with Hybrid Adversarial Training

This is the official source code of the paper: "Towards Robust Stacked Capsule Autoencoder with Hybrid Adversarial Training".

• Author: Jiazhu Dai, Siwei Xiong
• Institution: Shanghai University
• Email: daijz@shu.edu.cn (J. Dai)

Executable .py files in *_SCAE folders:

• train_*.py: Train the SCAE model with the specified defense method and save it under *_SCAE/checkpoints folder. Dataset cache files are saved under SCAE/datasets folder.
• test_*.py: Test the model under *_SCAE/checkpoints folder, generate and save the K-Means classifier at the same place.
• attack_opt_*.py: Launch the attack with OPT algorithm using the model under *_SCAE/checkpoints folder. Results are saved under *_SCAE/results/opt folder.

Please feel free to use it as you like.