RL-VAEGAN

Using style transfer for adversarial defense on RL agents.

Train RL agent

python -m agent.main --env-name 'PongDeterministic-v4'

Train RL-VAEGAN

• Use the state returned by the agent interacting with the environment to train the RL agent and RL-VAEGAN.

• Or, use the state returned by the well trained agent interacting with the environment to train RL-VAEGAN.

python -m rl_vaegan.train --env-name 'PongDeterministic-v4'

Defense with RL-VAEGAN

python -m attack.main --env-name 'PongDeterministic-v4' --which-epoch '00380000' --test-attacker 'fgsm'  --test-epsilon-adv 0.003