cs249r-project-repo's Introduction

cs249r-project repo

by Nikhil Suri, Thushan Puhalendran, and Samuel Clay

cleverhans-attacks: includes code referenced and used to generate simple adversarial attacks on labeled datasets, such as MNIST. mnist_tutorial_tf.py trains an MNIST classifier and shows the accuracy of the model after training and FSGM attacks. fast_gradient_method.py defines the FastGradientMethod object, which is an Attack object (defined in attack.py). Unfortunately, the cleverhans attacks expect models of a specific format. It was very difficult to adapt the OpenAI baselines agents that we trained to conform to this format.
cleverhans-rl-attack: includes code referenced and used in attempts to generate adversarial attacks on OpenAI agents using cleverhans. enjoy-adv.py is meant to load a pre-trained model and generate an attack for it. train.py can be used to either simply train a model (which can be saved and then attack using enjoy-adv.py) or attack a model during training to use adversarial training. Unfortunately, neither of the scripts (enjoy-adv.py and train.py) worked properly.
openai-baselines-models: includes weights for the Pong DQN agent that we trained using OpenAI baselines. This agent ran successfully on OpenAI gym.
keras-rl: A copy of the keral-rl repository with weights included for trained agents in examples/. Merging this with cleverhans attacks also proved very difficult.
RLMA RLMA code for the paper: Minimalistic Attacks: How Little it Takes to Fool a Deep Reinforcement Learning Policy, which describes methods for implementing attacks with minimal perturbation efforts
atari-rl Alternative implemenations of OpenAI Gym RL Agents for Atari games with additional documentation, providing a source of pre-trained agents to implement and evaluate adversarial attacks on

Recommend Projects