DashanGao / SPAA

SPAA: Semantic-Preserving Adversarial Attack

---

In this project, we propose an approach to harness the SPA attack via learning to learning approach.

---

Data Pre-processing

MNIST-ROT dataset generation: run adv_defence/process_mnist_rot.py

Model Training

Model definition: adv_defence\models.py

Configuration: adv_defence\config.py

Model training code: adv_defence\trainer_adv_example_2.py

To start model training, run main.py

The a demo of training script is saved in script.sh as follows:

python main.py --dataset mnist_rot --f_pretrain True --save_step 100 \
--g_optimizer adam --g_deeper_layer True --g_base_channel_dim 8 --img_size 32 \
--adv_loss_lambda 0.1 --g_msp_lambda 0.0001 --g_vae_lambda 0.0 \
--g_rec_lambda 0.001 --comment Rece-3Adv01Mspe-4