Adversarial Attacks on Neural Network Policies

Sandy Huang† , Nicolas Papernot‡ , Ian Goodfellow§ , Yan Duan†§ , Pieter Abbeel†§
†
University of California, Berkeley, Department of Electrical Engineering and Computer Sciences
‡
Pennsylvania State University, School of Electrical Engineering and Computer Science
§
OpenAI
arXiv:1702.02284v1 [cs.LG] 8 Feb 2017

Abstract

Machine learning classifiers are known to be vulnerable to inputs maliciously

constructed by adversaries to force misclassification. Such adversarial examples
have been extensively studied in the context of computer vision applications. In this
work, we show adversarial attacks are also effective when targeting neural network
policies in reinforcement learning. Specifically, we show existing adversarial exam-
ple crafting techniques can be used to significantly degrade test-time performance
of trained policies. Our threat model considers adversaries capable of introducing
small perturbations to the raw input of the policy. We characterize the degree of
vulnerability across tasks and training algorithms, for a subclass of adversarial-
example attacks in white-box and black-box settings. Regardless of the learned
task or training algorithm, we observe a significant drop in performance, even with
small adversarial perturbations that do not interfere with human perception. Videos
are available at http://rll.berkeley.edu/adversarial.

1 Introduction

Recent advances in deep learning and deep reinforcement learning (RL) have made it possible to learn
end-to-end policies that map directly from raw inputs (e.g., images) to a distribution over actions
to take. Deep RL algorithms have trained policies that achieve superhuman performance on Atari
games [15, 19, 16] and Go [21], perform complex robotic manipulation skills [13], learn locomotion
tasks [19, 14], and drive in the real world [7].
These policies are parametrized by neural networks, which have been shown to be vulnerable to
adversarial attacks in supervised learning settings. For example, for convolutional neural networks
trained to classify images, perturbations added to the input image can cause the network to classify
the adversarial image incorrectly, while the two images remain essentially indistinguishable to
humans [22]. In this work, we investigate whether such adversarial examples affect neural network
policies, which are trained with deep RL. We consider a fully trained policy at test time, and allow
the adversary to make limited changes to the raw input perceived from the environment before it is
passed to the policy.
Unlike supervised learning applications, where a fixed dataset of training examples is processed
during learning, in reinforcement learning these examples are gathered throughout the training
process. In other words, the algorithm used to train a policy, and even the random initialization of
the policy network’s weights, affects the states and actions encountered during training. Policies
trained to do the same task could conceivably be significantly different (e.g., in terms of the high-level
features they extract from the raw input), depending on how they were initialized and trained. Thus,
particular learning algorithms may result in policies more resistant to adversarial attacks. One could
also imagine that the differences between supervised learning and reinforcement learning might
prevent an adversary from mounting a successful attack in the black-box scenario, where the attacker
does not have access to the target policy network.
 action taken: down action taken: noop
original input adversarial input

action taken: up action taken: down

original input adversarial input

Figure 1: Two approaches for generating adversarial examples, applied to a policy trained using
DQN [15] to play Pong. The dotted arrow starts from the ball and denotes the direction it is traveling
in, and the green rectangle highlights the action that maximizes the Q-value, for the given input. In
both cases, the policy chooses a good action given the original input, but the adversarial perturbation
results in missing the ball and losing the point. Top: This adversarial example is computed using the
fast gradient sign method (FGSM) [10] with an `∞ -norm constraint on the adversarial perturbation;
the adversarial input is equivalent to the original input when converted to 8-bit image encodings, but is
still able to harm performance. Bottom: FGSM with an `1 -norm constraint; the optimal perturbation
is to create a “fake” ball lower than the position of the actual ball.

Our main contribution is to characterize how the effectiveness of adversarial examples is impacted by
two factors: the deep RL algorithm used to learn the policy, and whether the adversary has access
to the policy network itself (white-box vs. black-box). We first analyze three types of white-box
attacks on four Atari games trained with three deep reinforcement learning algorithms (DQN [15],
TRPO [19], and A3C [16]). We show that across the board, these trained policies are vulnerable to
adversarial examples. However, policies trained with TRPO and A3C seem to be more resistant to
adversarial attacks. Fig. 1 shows two examples of adversarial attacks on a Pong policy trained with
DQN, each at a specific time step during test-time execution.
Second, we explore black-box attacks on these same policies, where we assume the adversary has
access to the training environment (e.g., the simulator) but not the random initialization of the
target policy, and additionally may not know what the learning algorithm is. In the context of
computer vision, Szegedy et al. [22] observed the transferability property: an adversarial example
designed to be misclassified by one model is often misclassified by other models trained to solve
the same task. We observe that the cross-dataset transferability property also holds in reinforcement
learning applications, in the sense that an adversarial example designed to interfere with the operation
of one policy interferes with the operation of another policy, so long as both policies have been
trained to solve the same task. Specifically, we observe that adversarial examples transfer between
models trained using different trajectory rollouts and between models trained with different training
algorithms.

2
2 Related Work
Adversarial machine learning [2], and more generally the security and privacy of machine learn-
ing [18], encompasses a line of work that seeks to understand the behavior of models and learning
algorithms in the presence of adversaries. Such malicious individuals can target machine learning
systems either during learning by tampering with the training data [6], or during inference by ma-
nipulating inputs on which the model is making predictions [5]. Among the perturbations crafted at
test time, a class of adversarial inputs known as adversarial examples was introduced by [22]. This
first demonstration of the vulnerability of — then state-of-the-art — architectures to perturbations
indistinguishable to the human eye led to a series of follow-up work showing that perturbations
could be produced with minimal computing resources [10] and/or with access to the model label
predictions only (thus enabling black-box attacks) [17], and that these perturbations can also be
applied to physical objects [11, 20].
Most work on adversarial examples so far has studied their effect on supervised learning algorithms.
A recent technical report studied the scenario of an adversary interfering with the training of an agent,
with the intent of preventing the agent from learning anything meaningful [3]. Our work is the first
to study the ability of an adversary to interfere with the operation of an RL agent by presenting
adversarial examples at test time.

3 Preliminaries
In this section, we describe technical background on adversarial example crafting and deep reinforce-
ment learning, which are used throughout the paper.

3.1 Adversarial Example Crafting with the Fast Gradient Sign Method

Techniques for crafting adversarial examples generally focus on maximizing some measure of harm
caused by an adversarial perturbation, constrained by some limit on the size of the perturbation
intended to make it less noticeable to a human observer. A range of crafting techniques exist,
allowing the attacker to choose an attack that makes the right tradeoff between computational cost
and probability of success.
For exploratory research purposes, it is common to use a computationally cheap method of generating
adversarial perturbations, even if this reduces the attack success rate somewhat. We therefore use the
Fast Gradient Sign Method (FGSM) [10], an existing method for efficiently generating adversarial
examples in the context of computer vision classification. The FGSM is fast because it makes a
linear approximation of a deep model and solves the maximization problem analytically, in closed
form. Despite this approximation, it is still able to reliably fool many classifiers for computer vision
problems, because deep models often learn piece-wise linear functions with surprisingly large pieces.
FGSM focuses on adversarial perturbations where each pixel of the input image is changed by no
more than . Given a linear function g(x) = w> x, the optimal adversarial perturbation η that satisfies
kηk∞ <  is
η =  sign(w), (1)
since this perturbation maximizes the change in output for the adversarial example x̃, g(x̃) =
w> x + w> η.
Given an image classification network with parameters θ and loss J(θ, x, y), where x is an image
and y is a distribution over all possible class labels, linearizing the loss function around the input x
results in a perturbation of
η =  sign(∇x J(θ, x, y)). (2)

3.2 Deep Reinforcement Learning

Reinforcement learning algorithms train a policy π to optimize the expected cumulative reward
received over time. For a given state space S and action space A, the policy may be a deterministic
function mapping each state to an action: π : S → A, or it may be a stochastic function mapping
each state to a distribution over actions: π : S → ∆A , where ∆A is the probability simplex on A.
Here, the state space may consist of images or low-dimensional state representations. We choose to

3
represent π by a function parametrized by θ, for instance θ may be a weighting on features of the
state [1]. In the case of deep reinforcement learning, θ are the weights of a neural network. Over
the past few years, a large number of algorithms for deep RL have been proposed, including deep
Q-networks (DQN) [15], trust region policy optimization (TRPO) [19], and asynchronous advantage
actor-critic (A3C) [16]. We compare the effectiveness of adversarial examples on feed-forward
policies trained with each of these three algorithms.

3.2.1 Deep Q-Networks

Instead of modeling the policy directly, a DQN [15] approximately computes, for each state, the
Q-values for the available actions to take in that state. The Q-value Q∗ (s, a) for a state s and action a
is the expected cumulative discounted reward obtained by taking action a in state s, and following
the optimal policy thereafter. A DQN represents the Q-value function via a neural network trained to
minimize the squared Bellman error, using a variant of Q-learning. As this is off-policy learning, it
employs an -greedy exploration strategy. To reduce the variance of Q-learning updates, experience
replay is used: samples are randomly drawn from a replay buffer (where all recent transitions are
stored) so that they are not correlated due to time. The corresponding policy for a DQN is obtained
by choosing the action with the maximum Q-value for each state, hence it is deterministic.

3.2.2 Trust Region Policy Optimization

TRPO [19] is an on-policy batch learning algorithm. At each training iteration, whole-trajectory
rollouts of a stochastic policy are used to calculate the update to the policy parameters θ, while
controlling the change in the policy as measured by the KL divergence between the old and new
policies.

3.2.3 Asynchronous Advantage Actor-Critic

A3C [16] uses asynchronous gradient descent to speed up and stabilize learning of a stochastic
policy. It is based on the actor-critic approach, where the actor is a neural network policy π(a|s; θ)
and the critic is an estimate of the value function V (s; θv ). During learning, small batches of on-
policy samples are used to update the policy. The correlation between samples is reduced due to
asynchronous training, which stabilizes learning.

4 Adversarial Attacks
In our work, we use FGSM both as a white-box attack to compute adversarial perturbations for a
trained neural network policy πθ whose architecture and parameters are available to the adversary,
and as a black-box attack by computing gradients for a separately trained policy πθ0 to attack πθ using
adversarial example transferability [22, 10, 17].

4.1 Applying FGSM to Policies

FGSM requires calculating ∇x J(θ, x, y), the gradient of the cost function J(θ, x, y) with respect to
the input x. In reinforcement learning settings, we assume the output y is a weighting over possible
actions (i.e., the policy is stochastic: πθ : S → ∆A ). When computing adversarial perturbations with
FGSM for a trained policy πθ , we assume the action with the maximum weight in y is the optimal
action to take: in other words, we assume the policy performs well at the task. Thus, J(θ, x, y) is
the cross-entropy loss between y and the distribution that places all weight on the highest-weighted
action in y.1
Of the three learning algorithms we consider, TRPO and A3C both train stochastic policies. However,
DQN produces a deterministic policy, since it always selects the action that maximizes the computed
Q-value. This is problematic because it results in a gradient ∇x J(θ, x, y) of zero for almost all inputs
x. Thus, when calculating J(θ, x, y) for policies trained with DQN, we define y as a softmax of
the computed Q-values (with a temperature of 1). Note that we only do this for creating adversarial
examples; during test-time execution, policies trained with DQN are still deterministic.
1
Functionally, this is equivalent to a technique introduced in the context of image classification, to generate
adversarial examples without access to the true class label [12].

4
4.2 Choosing a Norm Constraint

Let η be the adversarial perturbation. In certain situations, it may be desirable to change all input
features by no more than a tiny amount (i.e., constrain the `∞ -norm of η), or it may be better to
change only a small number of input features (i.e., constrain the `1 -norm of η). Thus we consider
variations of FGSM that restrict the `1 - and `2 -norm of η, as well as the original version of FGSM
that restricts the `∞ -norm (Sec. 3.1).
Linearizing the cost function J(θ, x, y) around the current input x, the optimal perturbation for each
type of norm constraint is:

 sign(∇x J(θ, x, y)) for constraint kηk∞ ≤ 
√d ∗ ∇x J(θ,x,y)



k∇x J(θ,x,y)k2 for constraint kηk2 ≤ k1d k2
η= (3)


 maximally perturb highest-impact dimensions with budget d
for constraint kηk1 ≤ k1d k1


where d is the number of dimensions of input x.

Note that the `2 -norm and `1 -norm constraints have  adjusted to be the `2 - and `1 -norm of the vector
1d , respectively, since that is the amount of perturbation under the `∞ -norm constraint. In addition,
the optimal perturbation for the `1 -norm constraint either maximizes or minimizes the feature value
at dimensions i of the input, ordered by decreasing |∇θ J(θ, x, y)i |. For this norm, the adversary’s
budget — the total amount of perturbation the adversary is allowed to introduce in the input — is d.

5 Experimental Evaluation

We evaluate our adversarial attacks on four Atari 2600 games in the Arcade Learning Environment [4]:
Chopper Command, Pong, Seaquest, and Space Invaders. We choose these games to encompass a
variety of interesting environments; for instance, Chopper Command and Space Invaders include
multiple enemies.

5.1 Experimental Setup

We trained each game with three deep reinforcement learning algorithms: A3C [16], TRPO [19], and
DQN [15].
For DQN, we use the same pre-processing and neural network architecture as in [15] (Appendix A).
We also use this architecture for the stochastic policies trained by A3C and TRPO. Specifically, the
input to the neural network policy is a concatenation of the last 4 images, converted from RGB to
luminance (Y) and resized to 84 × 84. Luminance values are rescaled to be from 0 to 1. The output
of the policy is a distribution over possible actions.
For each game and training algorithm, we train five policies starting from different random initial-
izations. For our experiments, we focus on the top-performing trained policies, which we define as
all policies that perform within 80% of the maximum score for the last ten training iterations. We
cap the number of policies at three for each game and training algorithm. Certain combinations (e.g.,
Seaquest with A3C) had only one policy meet these requirements.
In order to reduce the variance of our experimental results, the average return for each result reported
is the average cumulative reward across ten rollouts of the target policy, without discounting rewards.

5.2 Vulnerability to White-Box Attacks

First, we are interested in how vulnerable neural network policies are to white-box adversarial-
example attacks, and how this is affected by the type of adversarial perturbation and by how the policy
is trained. If these attacks are effective, even small adversarial perturbations (i.e., small  for FGSM)
will be able to significantly lower the performance of the target trained network, as observed in [10]
for image classifiers. We evaluate multiple settings of  across all four games and three training
algorithms, for the three types of norm-constraints for FGSM.

5
 Chopper Command, A3C 30 Pong, A3C Seaquest, A3C Space Invaders, A3C
7000 3000 1000

6000 20 2500
800
5000
10 2000

Average Return
Average Return

Average Return

Average Return
4000 600
0 1500
3000 400
10 1000
2000
20 200
1000 500

0 30 0 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, TRPO 30 Pong, TRPO Seaquest, TRPO Space Invaders, TRPO
8000 1100 1400
7000 1000 1200
20
6000 900
1000
10 800

Average Return
Average Return

Average Return

Average Return
5000
700 800
4000 0
600 600
3000
10 500
400
2000 400
20 200
1000 300
0 30 200 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, DQN 20 Pong, DQN Seaquest, DQN 800 Space Invaders, DQN
2200 2500
2000 15 700
1800 10 2000
600
1600 5

Average Return
Average Return

500
Average Return

Average Return
1500
1400 0
400
1200 5
1000 300
1000 10
15 200
800 500
600 20 100
400 25 0 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²

Figure 2: Comparison of the effectiveness of `∞ , `2 , and `1 FGSM adversaries on four Atari games
trained with three learning algorithms. The average return is taken across ten trajectories. Constraint
on FGSM perturbation: `∞ -norm `2 -norm `1 -norm

5.2.1 Observations

We find that regardless of which game the policy is trained for or how it is trained, it is indeed possible
to significantly decrease the policy’s performance through introducing relatively small perturbations
in the inputs (Fig. 2).
Notably, in many cases an `∞ -norm FGSM adversary with  = 0.001 decreases the agent’s per-
formance by 50% or more; when converted to 8-bit image encodings, these adversarial inputs are
indistinguishable from the original inputs.
In cases where it is not essential for changes to be imperceptible, using an `1 -norm adversary may
be a better choice: given the same , `1 -norm adversaries are able to achieve the most significant
decreases in agent performance. They are able to sharply decrease the agent’s performance just by
changing a few pixels (by large amounts).
We see that policies trained with A3C, TRPO, and DQN are all susceptible to adversarial inputs. Inter-
estingly, policies trained with DQN are more susceptible, especially to `∞ -norm FGSM perturbations
on Pong, Seaquest, and Space Invaders.

5.3 Vulnerability to Black-Box Attacks

In practice, it is often the case that an adversary does not have complete access to the neural network
of the target policy [17]. This threat model is frequently referred to as a black-box scenario. We
investigate how vulnerable neural network policies are to black-box attacks of the following two
variants:

1. The adversary has access to the training environment and knowledge of the training algorithm
and hyperparameters. It knows the neural network architecture of the target policy network,
but not its random initialization. We will refer to this as transferability across policies.
2. The adversary additionally has no knowledge of the training algorithm or hyperparameters.
We will refer to this as transferability across algorithms.

6
 Chopper Command, A3C, `∞ norm 30 Pong, A3C, `∞ norm Seaquest, A3C, `∞ norm Space Invaders, A3C, `∞ norm
7000 3000 1000
900
6000 20 2500
800
5000
10 2000 700

Average Return
Average Return

Average Return

Average Return
4000 600
0 1500
3000 500
10 1000 400
2000
300
1000 20 500
200
0 30 0 100
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, A3C, `1 norm 30 Pong, A3C, `1 norm Seaquest, A3C, `1 norm Space Invaders, A3C, `1 norm
6000 3000 1000

5000 20 2500
800

4000 10 2000

Average Return
Average Return

Average Return

Average Return
600
3000 0 1500
400
2000 10 1000

200
1000 20 500

0 30 0 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, A3C, `2 norm 30 Pong, A3C, `2 norm Seaquest, A3C, `2 norm Space Invaders, A3C, `2 norm
8000 3000 1000

7000 900
20 2500
6000 800
10 2000 700
Average Return

5000
Average Return

Average Return

Average Return
600
4000 0 1500
500
3000
10 1000 400
2000 300
20 500
1000 200
0 30 0 100
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²

Figure 3: Transferability of adversarial inputs for policies trained with A3C. Type of transfer:
algorithm policy none

5.3.1 Transferability Across Policies

To explore transferability of adversarial examples across policies, we generate adversarial pertur-
bations for the target policy using one of the other top-performing policies trained with the same
algorithm for the same task. We test all adversary-target combinations of top-performing policies
trained with the same algorithm, for each combination of task, learning algorithm, and type of
adversary.

5.3.2 Transferability Across Training Algorithms

To explore transferability of adversarial examples across training algorithms, we generate adversarial
perturbations for the target policy using one of the top-performing policies trained with a different
algorithm. Similarly, we test all adversary-target combinations of top-performing policies trained
with different algorithms, for each combination of task and type of adversary.

5.3.3 Observations
As one might expect, we find that the less the adversary knows about the target policy, the less
effective the adversarial examples are (Fig. 3, 4, 5). Transferability across algorithms is less effective
at decreasing agent performance than transferability across policies, which is less effective than
when the adversary does not need to rely on transferability (i.e., the adversary has full access to the
target policy network). However, for most games, transferability across algorithms is still able to
significantly decrease the agent’s performance, especially for larger values of .
Notably for `1-norm adversaries, transferability across algorithms is nearly as effective as no transfer-
ability, for most game and algorithm combinations.

6 Discussion and Future Work

This direction of work has significant implications for both online and real-world deployment of
neural network policies. Our experiments show it is fairly easy to confuse such policies with
computationally-efficient adversarial examples, even in black-box scenarios. Based on [11], it
is possible that these adversarial perturbations could be applied to objects in the real world, for

7
 Chopper Command, TRPO, `∞ norm 30 Pong, TRPO, `∞ norm Seaquest, TRPO, `∞ norm Space Invaders, TRPO, `∞ norm
8000 1100 1400

7000 20 1000
1200
6000 900
10 1000

Average Return
Average Return 5000 800

Average Return

Average Return
4000 0 700 800

3000 600
10 600
2000 500
20 400
1000 400

0 30 300 200
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, TRPO, `1 norm 30 Pong, TRPO, `1 norm Seaquest, TRPO, `1 norm Space Invaders, TRPO, `1 norm
8000 1400

7000 1000
20 1200
6000
800 1000
10

Average Return
5000
Average Return

Average Return

Average Return
800
4000 0 600
600
3000
10
400 400
2000
20 200
1000 200
0 30 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, TRPO, `2 norm 30 Pong, TRPO, `2 norm Seaquest, TRPO, `2 norm Space Invaders, TRPO, `2 norm
8000 1100 1400
1000
7000 20 1200
900
6000 1000
10 800
Average Return
Average Return

Average Return

Average Return
5000 700 800
0
4000 600 600
10 500
3000 400
400
2000 20 200
300
1000 30 200 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²

Figure 4: Transferability of adversarial inputs for policies trained with TRPO. Type of transfer:
algorithm policy none

Chopper Command, DQN, `∞ norm 20 Pong, DQN, `∞ norm Seaquest, DQN, `∞ norm 800 Space Invaders, DQN, `∞ norm
2200 2500
2000 15 700
1800 10 2000
600
1600 5
500

Average Return
Average Return
Average Return

Average Return

1500
1400 0
400
1200 5
1000 300
1000 10
15 200
800 500
600 20 100

400 25 0 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, DQN, `1 norm 20 Pong, DQN, `1 norm Seaquest, DQN, `1 norm 700 Space Invaders, DQN, `1 norm
2200 2500
2000 15
600
1800 10 2000
500
1600 5
Average Return
Average Return
Average Return

Average Return

1500 400
1400 0

1200 5 300
1000
1000 10
200
800 15 500
100
600 20
400 25 0 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²
Chopper Command, DQN, `2 norm 20 Pong, DQN, `2 norm Seaquest, DQN, `2 norm 700 Space Invaders, DQN, `2 norm
2200 2500

2000 15
600
10 2000
1800 500
5
Average Return
Average Return

1600
Average Return

Average Return

1500 400
0
1400
5 300
1200 1000
10
200
1000 15 500
800 100
20
600 25 0 0
0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008
² ² ² ²

Figure 5: Transferability of adversarial inputs for policies trained with DQN. Type of transfer:
algorithm policy none

example adding strategically-placed paint to the surface of a road to confuse an autonomous car’s
lane-following policy.
Thus, an important direction of future work is developing defenses against adversarial attacks. This
could involve adding adversarially-perturbed examples during training time (as in [10]), or it could
involve detecting adversarial input at test time, to be able to deal with it appropriately.

8
References
[1] P. Abbeel and A. Y. Ng. Apprenticeship learning via inverse reinforcement learning. In
Proceedings of the Twenty-First International Conference on Machine Learning, 2004.
[2] M. Barreno, B. Nelson, R. Sears, A. D. Joseph, and J. D. Tygar. Can machine learning be secure?
In Proceedings of the 2006 ACM Symposium on Information, Computer and Communications
Security, pages 16–25, 2006.
[3] V. Behzadan and A. Munir. Vulnerability of deep reinforcement learning to policy induction
attacks. arXiv preprint arXiv:1701.04143, 2017.
[4] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An
evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279,
06 2013.
[5] B. Biggio, I. Corona, D. Maiorca, B. Nelson, N. Šrndić, P. Laskov, G. Giacinto, and F. Roli.
Evasion attacks against machine learning at test time. In Machine Learning and Knowledge
Discovery in Databases, pages 387–402, 2013.
[6] B. Biggio, B. Nelson, and L. Pavel. Poisoning attacks against support vector machines. In
Proceedings of the Twenty-Ninth International Conference on Machine Learning, 2012.
[7] M. Bojarski, D. D. Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, L. D. Jackel, M. Mon-
fort, U. Muller, J. Zhang, X. Zhang, J. Zhao, and K. Zieba. End to end learning for self-driving
cars. arXiv preprint arXiv:1604.07316, 2016.
[8] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba.
Openai gym, 2016.
[9] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel. Benchmarking deep reinforcement
learning for continuous control. In Proceedings of the Thirty-Third International Conference on
Machine Learning, 2016.
[10] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. In
Proceedings of the Third International Conference on Learning Representations, 2015.
[11] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. arXiv
preprint arXiv:1607.02533, 2016.
[12] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. Proceedings
of the Fifth International Conference on Learning Representations, 2017.
[13] S. Levine, C. Finn, T. Darrell, and P. Abbeel. End-to-end training of deep visuomotor policies.
Journal of Machine Learning Research, 17(39):1–40, 2016.
[14] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra.
Continuous control with deep reinforcement learning. In Proceedings of the Fourth International
Conference on Learning Representations, 2016.
[15] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller.
Playing atari with deep reinforcement learning. In NIPS Workshop on Deep Learning, 2013.
[16] V. Mnih, A. Puigdomenech Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver, and
K. Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of
the Thirty-Third International Conference on Machine Learning, 2016.
[17] N. Papernot, P. McDaniel, I. Goodfellow, S. Jha, Z. B. Celik, and A. Swami. Practical
black-box attacks against deep learning systems using adversarial examples. arXiv preprint
arXiv:1602.02697, 2016.
[18] N. Papernot, P. McDaniel, A. Sinha, and M. Wellman. Towards the science of security and
privacy in machine learning. arXiv preprint arXiv:1611.03814, 2016.

9
[19] J. Schulman, S. Levine, P. Moritz, M. I. Jordan, and P. Abbeel. Trust region policy optimization.
In Proceedings of the Thirty-Second International Conference on Machine Learning, 2015.
[20] M. Sharif, S. Bhagavatula, L. Bauer, and M. K. Reiter. Accessorize to a crime: Real and
stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM SIGSAC
Conference on Computer and Communications Security, pages 1528–1540, 2016.
[21] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser,
I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalch-
brenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis.
Mastering the game of go with deep neural networks and tree search. Nature, 529:484–503,
2016.
[22] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus.
Intriguing properties of neural networks. In Proceedings of the Second International Conference
on Learning Representations, 2014.

A Experimental Setup
We set up our experiments within the rllab [9] framework. We use a parallelized version of the
rllab implementation of TRPO, and integrate outside implementations of DQN2 and A3C3 . We use
OpenAI Gym environments [8] as the interface to the Arcade Learning Environment [4].
The policies use the network architecture from [15]: a convolutional layer with 16 filters of size 8 × 8
with a stride of 4, followed by a convolutional layer with 32 filters of size 4 × 4 with a stride of 2.
The last layer is a fully-connected layer with 256 hidden units. All hidden layers are followed by a
rectified nonlinearity.
For all games, we set the frame skip to 4 as in [15]. The frame skip specifies the number of times the
agent’s chosen action is repeated.

A.1 Training

We trained policies with TRPO and A3C on Amazon EC2 c4.8xlarge machines. For each policy, we
ran TRPO for 2,000 iterations of 100,000 steps each, which took 1.5 to 2 days. We set the bound on
the KL divergence to 0.01, as in [19].
For A3C, we used 18 actor-learner threads and a learning rate of 0.0004. As in [16], we use an
entropy regularization weight of 0.01, use RMSProp for optimization with a decay factor of 0.99,
update the policy and value networks every 5 time steps, and share all weights except the output layer
between the policy and value networks. For each policy, we ran A3C for 200 iterations of 1,000,000
steps each, which took 1.5 to 2 days.
For DQN, we trained policies on Amazon EC2 p2.xlarge machines. We used 100,000 steps per epoch
and trained for two days.

2
github.com/spragunr/deep_q_rl
3
github.com/muupan/async-rl

10