Skip to content

Create Natural Adversarial Observations for Atari games

Notifications You must be signed in to change notification settings

phanfeld/nao-atari

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

nao-atari

Create Natural Adversarial Observations for Atari games

CD-UAP on Enduro

With the notebook provided in this repository, you will be able to recreate three attacks from my master thesis "Introducing natural adversarial observations to a Deep Reinforcement Learning agent for Atari games".

So what do we see in the images? To the left, the plotted input the Atari agent will be presented with; a stack of 4 consecutive, grayscaled images from the game Enduro. We add the middle image, the calculated perturbation, and receive the adversarial observation to the right. Adversarial observation and original input are indistinguishable for humans, but the agent playing the game will be tricked into breaking and loosing score points.

The repository comes with three already trained agents for Enduro, Road Runner and Breakout. All of them outperform a human expert player.

Requirements

The attacks were tested with the following packages installed

About

Create Natural Adversarial Observations for Atari games

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published