Adversarial Training for Free!

Ali Shafahi Mahyar Najibi Amin Ghiasi

University of Maryland University of Maryland University of Maryland
[email protected] [email protected] [email protected]

Zheng Xu John Dickerson Christoph Studer

University of Maryland University of Maryland Cornell University
[email protected] [email protected] [email protected]

Larry S. Davis Gavin Taylor Tom Goldstein

University of Maryland United States Naval Academy University of Maryland
[email protected] [email protected] [email protected]

Abstract

Adversarial training, in which a network is trained on adversarial examples, is one

of the few defenses against adversarial attacks that withstands strong attacks. Un-
fortunately, the high cost of generating strong adversarial examples makes standard
adversarial training impractical on large-scale problems like ImageNet. We present
an algorithm that eliminates the overhead cost of generating adversarial examples
by recycling the gradient information computed when updating model parameters.
Our “free” adversarial training algorithm achieves comparable robustness to PGD
adversarial training on the CIFAR-10 and CIFAR-100 datasets at negligible addi-
tional cost compared to natural training, and can be 7 to 30 times faster than other
strong adversarial training methods. Using a single workstation with 4 P100 GPUs
and 2 days of runtime, we can train a robust model for the large-scale ImageNet
classification task that maintains 40% accuracy against PGD attacks.

1 Introduction

Deep learning has been widely applied to various computer vision tasks with excellent performance.
Prior to the realization of the adversarial example phenomenon by Biggio et al. [2013], Szegedy et al.
[2013], model performance on clean examples was the the main evaluation criteria. However, in
security-critical applications, robustness to adversarial attacks has emerged as a critical factor.
A robust classifier is one that correctly labels adversarially perturbed images. Alternatively, robustness
may be achieved by detecting and rejecting adversarial examples [Ma et al., 2018, Meng and Chen,
2017, Xu et al., 2017]. Recently, Athalye et al. [2018] broke a complete suite of allegedly robust
defenses, leaving adversarial training, in which the defender augments each minibatch of training
data with adversarial examples [Madry et al., 2017], among the few that remain resistant to attacks.
Adversarial training is time-consuming—in addition to the gradient computation needed to update
the network parameters, each stochastic gradient descent (SGD) iteration requires multiple gradient
computations to produce adversarial images. In fact, it takes 3-30 times longer to form a robust
network with adversarial training than forming a non-robust equivalent. Put simply, the actual
slowdown factor depends on the number of gradient steps used for adversarial example generation.
The high cost of adversarial training has motivated a number of alternatives. Some recent works
replace the perturbation generation in adversarial training with a parameterized generator network

33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.
[Baluja and Fischer, 2018, Poursaeed et al., 2018, Xiao et al., 2018]. This approach is slower than
standard training, and problematic on complex datasets, such as ImageNet, for which it is hard to
produce highly expressive GANs that cover the entire image space. Another popular defense strategy
is to regularize the training loss using label smoothing, logit squeezing, or a Jacobian regularization
[Shafahi et al., 2019a, Mosbach et al., 2018, Ross and Doshi-Velez, 2018, Hein and Andriushchenko,
2017, Jakubovitz and Giryes, 2018, Yu et al., 2018]. These methods have not been applied to
large-scale problems, such as ImageNet, and can be applied in parallel to adversarial training.
Recently, there has been a surge of certified defenses [Wong and Kolter, 2017, Wong et al., 2018,
Raghunathan et al., 2018a,b, Wang et al., 2018]. These methods were mostly demonstrated for
small networks, low-res datasets, and relatively small perturbation budgets (). Lecuyer et al. [2018]
propose randomized smoothing as a certified defense and which was later improved by Li et al.
[2018a]. Cohen et al. [2019] prove a tight robustness guarantee under the `2 norm for smoothing
with Gaussian noise. Their study was the first certifiable defense for the ImageNet dataset [Deng
et al., 2009]. They claim to achieve 12% robustness against non-targeted attacks that are within an
`2 radius of 3 (for images with pixels in [0, 1]). This is roughly equivalent to an `∞ radius of  = 2
when pixels lie in [0, 255] .
Adversarial training remains among the most trusted defenses, but it is nearly intractable on large-
scale problems. Adversarial training on high-resolution datasets, including ImageNet, has only been
within reach for research labs having hundreds of GPUs1 . Even on reasonably-sized datasets, such as
CIFAR-10 and CIFAR-100, adversarial training is time consuming and can take multiple days.

Contributions

We propose a fast adversarial training algorithm that produces robust models with almost no extra
cost relative to natural training. The key idea is to update both the model parameters and image
perturbations using one simultaneous backward pass, rather than using separate gradient computations
for each update step. Our proposed method has the same computational cost as conventional natural
training, and can be 3-30 times faster than previous adversarial training methods [Madry et al.,
2017, Xie et al., 2019]. Our robust models trained on CIFAR-10 and CIFAR-100 achieve accuracies
comparable and even slightly exceeding models trained with conventional adversarial training when
defending against strong PGD attacks.
We can apply our algorithm to the large-scale ImageNet classification task on a single workstation
with four P100 GPUs in about two days, achieving 40% accuracy against non-targeted PGD attacks.
To the best of our knowledge, our method is the first to successfully train a robust model for ImageNet
based on the non-targeted formulation and achieves results competitive with previous (significantly
more complex) methods [Kannan et al., 2018, Xie et al., 2019].

2 Non-targeted adversarial examples

Adversarial examples come in two flavors: non-targeted and targeted. Given a fixed classifier with
parameters θ, an image x with true label y, and classification proxy loss l, a bounded non-targeted
attack sneaks an example out of its natural class and into another. This is done by solving
max l(x + δ, y, θ), subject to ||δ||p ≤ , (1)
δ

where δ is the adversarial perturbation, ||.||p is some `p -norm distance metric, and  is the adversarial
manipulation budget. In contrast to non-targeted attacks, a targeted attack scooches an image into a
specific class of the attacker’s choice.
In what follows, we will use non-targeted adversarial examples both for evaluating the robustness of
our models and also for adversarial training. We briefly review some of the closely related methods
for generating adversarial examples. In the context of `∞ -bounded attacks, the Fast Gradient Sign
Method (FGSM) by Goodfellow et al. [2015] is one of the most popular non-targeted methods that
uses the sign of the gradients to construct an adversarial example in one iteration:
xadv = x +  · sign(∇x l(x, y, θ)). (2)
1
Xie et al. [2019] use 128 V100s and Kannan et al. [2018] use 53 P100s for targeted adv training ImageNet.

2
The Basic Iterative Method (BIM) by Kurakin et al. [2016a] is an iterative version of FGSM. The
PGD attack is a variant of BIM with uniform random noise as initialization, which is recognized by
Athalye et al. [2018] to be one of the most powerful first-order attacks. The initial random noise was
first studied by Tramèr et al. [2017] to enable FGSM to attack models that rely on “gradient masking.”
In the PGD attack algorithm, the number of iterations K plays an important role in the strength
of attacks, and also the computation time for generating adversarial examples. In each iteration, a
complete forward and backward pass is needed to compute the gradient of the loss with respect to the
image. Throughout this paper we will refer to a K-step PGD attack as PGD-K.

3 Adversarial training

Adversarial training can be traced back to [Goodfellow et al., 2015], in which models were hardened
by producing adversarial examples and injecting them into training data. The robustness achieved
by adversarial training depends on the strength of the adversarial examples used. Training on fast
non-iterative attacks such as FGSM and Rand+FGSM only results in robustness against non-iterative
attacks, and not against PGD attacks [Kurakin et al., 2016b, Madry et al., 2017]. Consequently, Madry
et al. [2017] propose training on multi-step PGD adversaries, achieving state-of-the-art robustness
levels against `∞ attacks on MNIST and CIFAR-10 datasets.
While many defenses were broken by Athalye et al. [2018], PGD-based adversarial training was
among the few that withstood strong attacks. Many other defenses build on PGD adversarial training
or leverage PGD adversarial generation during training. Examples include Adversarial Logit Pairing
(ALP) [Kannan et al., 2018], Feature Denoising [Xie et al., 2019], Defensive Quantization [Lin et al.,
2019], Thermometer Encoding [Buckman et al., 2018], PixelDefend [Song et al., 2017], Robust
Manifold Defense [Ilyas et al., 2017], L2-nonexpansive nets [Qian and Wegman, 2018], Jacobian
Regularization [Jakubovitz and Giryes, 2018], Universal Perturbation [Shafahi et al., 2018], and
Stochastic Activation Pruning [Dhillon et al., 2018].
We focus on the min-max formulation of adversarial training [Madry et al., 2017], which has been
theoretically and empirically justified. This widely used K-PGD adversarial training algorithm has
an inner loop that constructs adversarial examples by PGD-K, while the outer loop updates the model
using minibatch SGD on the generated examples. In the inner loop, the gradient ∇x l(xadv , y, θ) for
updating adversarial examples requires a forward-backward pass of the entire network, which has
similar computation cost as calculating the gradient ∇θ l(xadv , y, θ) for updating network parameters.
Compared to natural training, which only requires ∇θ l(x, y, θ) and does not have an inner loop,
K-PGD adversarial training needs roughly K + 1 times more computation.

4 “Free” adversarial training

K-PGD adversarial training [Madry et al., 2017] is generally slow. For example, the 7-PGD training
of a WideResNet [Zagoruyko and Komodakis, 2016] on CIFAR-10 in Madry et al. [2017] takes about
four days on a Titan X GPU. To scale the algorithm to ImageNet, Xie et al. [2019] and Kannan et al.
[2018] had to deploy large GPU clusters at a scale far beyond the reach of most organizations.
Here, we propose free adversarial training, which has a negligible complexity overhead compared to
natural training. Our free adversarial training algorithm (alg. 1) computes the ascent step by re-using
the backward pass needed for the descent step. To update the network parameters, the current training
minibatch is passed forward through the network. Then, the gradient with respect to the network
parameters is computed on the backward pass. When the “free” method is used, the gradient of the
loss with respect to the input image is also computed on this same backward pass.
Unfortunately, this approach does not allow for multiple adversarial updates to be made to the same
image without performing multiple backward passes. To overcome this restriction, we propose a
minor yet nontrivial modification to training: train on the same minibatch m times in a row. Note
that we divide the number of epochs by m such that the overall number of training iterations remains
constant. This strategy provides multiple adversarial updates to each training image, thus providing
strong/iterative adversarial examples. Finally, when a new minibatch is formed, the perturbation
generated on the previous minibatch is used to warm-start the perturbation for the new minibatch.

3
Algorithm 1 “Free” Adversarial Training (Free-m)
Require: Training samples X, perturbation bound , learning rate τ , hop steps m
1: Initialize θ
2: δ ← 0
3: for epoch = 1 . . . Nep /m do
4: for minibatch B ⊂ X do
5: for i = 1 . . . m do
6: Update θ with stochastic gradient descent
7: gθ ← E(x,y)∈B [∇θ l(x + δ, y, θ)]
8: gadv ← ∇x l(x + δ, y, θ)]
9: θ ← θ − τ gθ
10: Use gradients calculated for the minimization step to update δ
11: δ ← δ +  · sign(gadv )
12: δ ← clip(δ, −, )
13: end for
14: end for
15: end for

(a) CIFAR-10 sensitivity to m (b) CIFAR-100 sensitivity to m

Figure 1: Natural validation accuracy of Wide Resnet 32-10 models using varied mini-batch replay
parameters m. Here m = 1 corresponds to natural training. For large m’s, validation accuracy drops
drastically. However, small m’s have little effect. For reference, CIFAR-10 and CIFAR-100 models
that are 7-PGD adversarially trained have natural accuracies of 87.25% and 59.87%, respectively.

The effect of mini-batch replay on natural training

While the hope for alg. 1 is to build robust models, we still want models to perform well on natural
examples. As we increase m in alg. 1, there is risk of increasing generalization error. Furthermore,
it may be possible that catastrophic forgetting happens. Consider the worst case where all the
“informative” images of one class are in the first few mini-batches. In this extreme case, we do not see
useful examples for most of the epoch, and forgetting may occur. Consequently, a natural question is:
how much does mini-batch replay hurt generalization?
To answer this question, we naturally train wide-resnet 32-10 models on CIFAR-10 and CIFAR-100
using different levels of replay. Fig. 1 plots clean validation accuracy as a function of the replay
parameter m. We see some dropoff in accuracy for small values of m. Note that a small compromise
in accuracy is acceptable given a large increase in robustness due to the fundamental tradeoffs
between robustness and generalization [Tsipras et al., 2018, Zhang et al., 2019a, Shafahi et al.,
2019b]. As a reference, CIFAR-10 and CIFAR-100 models that are 7-PGD adversarially trained
have natural accuracies of 87.25% and 59.87%, respectively. These same accuracies are exceeded by
natural training with m = 16. We see in section 5 that good robustness can be achieved using “free”
adversarial training with just m ≤ 10.

4
 Table 1: Validation accuracy and robustness of CIFAR-10 models trained with various methods.
Evaluated Against Train
Training
10 restart Time
Nat. Images PGD-20 PGD-100 CW-100
PGD-20 (min)
Natural 95.01% 0.00% 0.00% 0.00% 0.00% 780
Free m = 2 91.45% 33.92% 33.20% 34.57% 33.41% 816
Free m = 4 87.83% 41.15% 40.35% 41.96% 40.73% 800
Free m = 8 85.96% 46.82% 46.19% 46.60% 46.33% 785
Free m = 10 83.94% 46.31% 45.79% 45.86% 45.94% 785
7-PGD trained 87.25% 45.84% 45.29% 46.52% 45.53% 5418

Table 2: Validation accuracy and robustness of CIFAR-100 models trained with various methods.
Evaluated Against Training Time
Training
Natural Images PGD-20 PGD-100 (minutes)
Natural 78.84% 0.00% 0.00% 811
Free m = 2 69.20% 15.37% 14.86% 816
Free m = 4 65.28% 20.64% 20.15% 767
Free m = 6 64.87% 23.68% 23.18% 791
Free m = 8 62.13% 25.88% 25.58% 780
Free m = 10 59.27% 25.15% 24.88% 776
Madry et al. (2-PGD trained) 67.94% 17.08% 16.50% 2053
Madry et al. (7-PGD trained) 59.87% 22.76% 22.52% 5157

5 Robust models on CIFAR-10 and 100

In this section, we train robust models on CIFAR-10 and CIFAR-100 using our “free” adversarial
training ( alg. 1) and compare them to K-PGD adversarial training 23 . We find that free training is
able to achieve state-of-the-art robustness on the CIFARs without the overhead of standard PGD
training.

CIFAR-10

We train various CIFAR-10 models using the Wide-Resnet 32-10 model and standard hyper-
parameters used by Madry et al. [2017]. In the proposed method (alg. 1), we repeat (i.e. replay) each
minibatch m times before switching to the next minibatch. We present the experimental results for
various choices of m in table 1. Training each of these models costs roughly the same as natural
training since we preserve the same number of iterations. We compare with the 7-PGD adversarially
trained model from Madry et al. [2017] 4 , whose training requires roughly 7× more time than all of
our free training variations. We attack all models using PGD attacks with K iterations on both the
cross-entropy loss (PGD-K) and the Carlini-Wagner loss (CW-K) [Carlini and Wagner, 2017]. We
test using the PGD-20 attack following Madry et al. [2017], and also increase the number of attack
iterations and employ random restarts to verify robustness under stronger attacks. To measure the
sensitivity of our method to initialization, we perform five trials for the Free-m = 8 case and find that
our results are insensitive. The natural accuracy is 85.95±0.14 and robustness against a 20-random
restart PGD-20 attack is 46.49±0.19. Note that gradient free-attacks such as SPSA will result in
inferior results for adversarially trained models in comparison to optimization based attacks such
as PGD as noted by Uesato et al. [2018]. Gradient-free attacks are superior in settings where the
defense works by masking or obfuscating the gradients.

2
Adversarial Training for Free code for CIFAR-10 in Tensorflow can be found here: https://github.
com/ashafahi/free_adv_train/
3
ImageNet Adversarial Training for Free code in Pytorch can be found here: https://github.com/
mahyarnajibi/FreeAdversarialTraining
4
Results based on the “adv_trained” model in Madry’s CIFAR-10 challenge repo.

5
 plane cat dog cat ship cat dog car horse

car dog cat deer cat frog bird frog dog

car dog cat deer cat horse cat plane car

Figure 2: Attack images built for adversarially trained models look like the class into which they get
misclassified. We display the last 9 CIFAR-10 clean validation images (top row) and their adversarial
examples built for a 7-PGD adversarially trained (middle) and our “free” trained (bottom) models.

Our “free training” algorithm successfully reaches robustness levels comparable to a 7-PGD ad-
versarially trained model. As we increase m, the robustness is increased at the cost of validation
accuracy on natural images. Additionally note that we achieve reasonable robustness over a wide
range of choices of the main hyper-parameter of our model, 10 ≥ m > 2, and the proposed method
is significantly faster than 7-PGD adversarial training. Recently, a new method called YOPO [Zhang
et al., 2019b] has been proposed for speeding up adversarial training, in their CIFAR-10 results they
use a wider networks (WRN-34-10) with larger batch-sizes (256). As shown in our supplementary,
both of these factors increase robustness. To do a direct comparison, we a train WRN-34-10 using
m = 10 and batch-size=256. We match their best reported result (48.03% against PGD-20 attacks
for “Free” training v.s. 47.98% for YOPO 5-3).

CIFAR-100

We also study the robustness results of “free training” on CIFAR-100 which is a more difficult dataset
with more classes. As we will see in sec. 4, training with large m values on this dataset hurts the
natural validation accuracy more in comparison to CIFAR-10. This dataset is less studied in the
adversarial machine learning community and therefore for comparison purposes, we adversarially
train our own Wide ResNet 32-10 models for CIFAR-100. We train two robust models by varying K
in the K-PGD adversarial training algorithm. One is trained on PGD-2 with a computational cost
almost 3× that of free training, and the other is trained on PGD-7 with a computation time roughly
7× that of free training. We adopt the code for adversarial training from Madry et al. [2017], which
produces state-of-the-art robust models on CIFAR-10. We summarize the results in table. 2.
We see that “free training” exceeds the accuracy on both natural images and adversarial images when
compared to traditional adversarial training. Similar to the effect of increasing m, increasing K in
K-PGD adversarial training results in increased robustness at the cost of clean validation accuracy.
However, unlike the proposed “free training” where increasing m has no extra cost, increasing K for
standard K-PGD substantially increases training time.

6 Does “free” training behave like standard adversarial training?

Here, we analyze two properties that are associated with PGD adversarially trained models: The
interpretability of their gradients and the flattness of their loss surface. We find that “free” training
enjoys these benefits as well.

Generative behavior for largely perturbed examples

Tsipras et al. [2018] observed that hardened classifiers have interpretable gradients; adversarial
examples built for PGD trained models often look like the class into which they get misclassified.
Fig. 2 plots “weakly bounded” adversarial examples for the CIFAR-10 7-PGD adversarially trained
model [Madry et al., 2017] and our free m = 8 trained model. Both models were trained to resist `∞
attacks with  = 8. The examples are made using a 50 iteration BIM attack with  = 30 and s = 2.

6
 (a) Free m = 8 (b) 7-PGD adv trained

(c) Free m = 8 both rad (d) 7-PGD adv trained both rad
Figure 3: The loss surface of a 7-PGD adversarially trained model and our “free” trained model for
CIFAR-10 on the first 2 validation images. In (a) and (b) we display the cross-entropy loss projected
on one random (Rademacher) and one adversarial direction. In (c) and (d) we display the the cross
entropy loss projected along two random directions. Both training methods behave similarly and do
not operate by masking the gradients as the adversarial direction is indeed the direction where the
cross-entropy loss changes the most.

“Free training” maintains generative properties, as our model’s adversarial examples resemble the
target class.

Smooth and flattened loss surface

Another property of PGD adversarial training is that it flattens and smoothens the loss landscape.
In contrast, some defenses work by “masking” the gradients, i.e., making it difficult to identify
adversarial examples using gradient methods, even though adversarial examples remain present.
Reference Engstrom et al. [2018] argues that gradient masking adds little security. We show in fig. 3a
that free training does not operate by masking gradients using a rough loss surface. In fig. 3 we
plot the cross-entropy loss projected along two directions in image space for the first few validation
examples of CIFAR-10 [Li et al., 2018b]. In addition to the loss of the free m = 8 model, we plot the
loss of the 7-PGD adversarially trained model for comparison.

7 Robust ImageNet classifiers

ImageNet is a large image classification dataset of over 1 million high-res images and 1000 classes
(Russakovsky et al. [2015]). Due to the high computational cost of ImageNet training, only a
few research teams have been able to afford building robust models for this problem. Kurakin
et al. [2016b] first hardened ImageNet classifiers by adversarial training with non-iterative attacks.5
Adversarial training was done using a targeted FGSM attack. They found that while their model
became robust against targeted non-iterative attacks, the targeted BIM attack completely broke it.
Later, Kannan et al. [2018] attempted to train a robust model that withstands targeted PGD attacks.
They trained against 10 step PGD targeted attacks (a process that costs 11 times more than natural
training) to build a benchmark model. They also generated PGD targeted attacks to train their
adversarial logit paired (ALP) ImageNet model. Their baseline achieves a top-1 accuracy of 3.1%
against PGD-20 targeted attacks with  = 16. Very recently, Xie et al. [2019] trained a robust
ImageNet model against targeted PGD-30 attacks, with a cost 31× that of natural training. Training
this model required a distributed implementation on 128 GPUs with batch size 4096. Their robust
ResNet-101 model achieves a top-1 accuracy of 35.8% on targeted PGD attacks with many iterations.

5
Training using a non-iterative attack such as FGSM only doubles the training cost.

7
Table 3: ImageNet validation accuracy and robustness of ResNet-50 models trained with various
replay parameters and  = 2.
Evaluated Against
Training
Natural Images PGD-10 PGD-50 PGD-100
Natural 76.038% 0.166% 0.052% 0.036%
Free m = 2 71.210% 37.012% 36.340% 36.250%
Free m = 4 64.446% 43.522% 43.392% 43.404%
Free m = 6 60.642% 41.996% 41.900% 41.892%
Free m = 8 58.116% 40.044% 40.008% 39.996%

(a) Clean (b) PGD-100

Figure 4: The effect of the perturbation bound  and the mini-batch replay hyper-parameter m on the
robustness achieved by free training.

Free training results

Our alg. 1 is designed for non-targeted adversarial training. As Athalye et al. [2018] state, defending
on this task is important and more challenging than defending against targeted attacks, and for this
reason smaller  values are typically used. Even for  = 2 (the smallest  we consider defensing
against), a PGD-50 non-targeted attack on a natural model achieves roughly 0.05% top-1 accuracy.
To put things further in perspective, Uesato et al. [2018] broke three defenses for  = 2 non-targeted
attacks on ImageNet [Guo et al., 2017, Liao et al., 2018, Xie et al., 2017], degrading their performance
below 1%. Our free training algorithm is able to achieve 43% robustness against PGD attacks bounded
by  = 2. Furthermore, we ran each experiment on a single workstation with four P100 GPUs. Even
with this modest setup, training time for each ResNet-50 experiment is below 50 hours.
We summarize our results for various ’s and m’s in table 3 and fig. 4. To craft attacks, we used a
step-size of 1 and the corresponding  used during training. In all experiments, the training batch
size was 256. Table 3 shows the robustness of Resnet-50 on ImageNet with  = 2. The validation
accuracy for natural images decreases when we increase the minibatch replay m, just like it did for
CIFAR in section 5.
The naturally trained model is vulnerable to PGD attacks (first row of table 3), while free training
produces robust models that achieve over 40% accuracy vs PGD attacks (m = 4, 6, 8 in table 3).
Attacking the models using PGD-100 does not result in a meaningful drop in accuracy compared to
PGD-50. Therefore, we did not experiment with increasing the number of PGD iterations further.
Fig. 4 summarizes experimental results for robust models trained and tested under different per-
turbation bounds . Each curve represents one training method (natural training or free training)
with hyperparameter choice m. Each point on the curve represents the validation accuracy for an
-bounded robust model. These results are also provided as tables in the appendix. The proposed
method consistently improves the robust accuracy under PGD attacks for  = 2 − 7, and m = 4
performs the best. It is difficult to train robust models when  is large, which is consistent with
previous studies showing that PGD-based adversarial training has limited robustness for ImageNet
[Kannan et al., 2018].

8
Table 4: Validation accuracy and robustness of “free” and 2-PGD trained ResNet-50 models – both
trained to resist `∞  = 4 attacks. Note that 2-PGD training time is 3.46× that of “free” training.
Evaluated Against Train time
Model & Training
Natural Images PGD-10 PGD-50 PGD-100 (minutes)
RN50 – Free m = 4 60.206% 32.768% 31.878% 31.816% 3016
RN50 – 2-PGD trained 64.134% 37.172% 36.352% 36.316% 10,435

Table 5: Validation accuracy and robustness of free-m = 4 trained ResNets with various capacities.
Evaluated Against
Architecture
Natural Images PGD-10 PGD-50 PGD-100
ResNet-50 60.206% 32.768% 31.878% 31.816%
ResNet-101 63.340% 35.388% 34.402% 34.328%
ResNet-152 64.446% 36.992% 36.044% 35.994%

Comparison with PGD-trained models

We compare “free” training to a more costly method using 2-PGD adversarial examples  = 4.
We run the conventional adversarial training algorithm and set s = 2,  = 4, and K = 2. All
other hyper-parameters were identical to those used for training our “free” models. Note that in our
experiments, we do not use any label-smoothing or other common tricks for improving robustness
since we want to do a fair comparison between PGD training and our “free” training. These extra
regularizations can likely improve results for both approaches.
We compare our “free trained” m = 4 ResNet-50 model and the 2-PGD trained ResNet-50 model in
table 4. 2-PGD adversarial training takes roughly 3.4× longer than “free training” and only achieves
slightly better results (≈4.5%). This gap is less than 0.5% if we free train a higher capacity model
(i.e. ResNet-152, see below).

Free training on models with more capacity

It is believed that increased network capacity leads to greater robustness from adversarial training
[Madry et al., 2017, Kurakin et al., 2016b]. We verify that this is the case by “free training” ResNet-
101 and ResNet-152 with  = 4. The comparison between ResNet-152, ResNet-101, and ResNet-50
is summarized in table 5. Free training on ResNet-101 and ResNet-152 each take roughly 1.7×
and 2.4× more time than ResNet-50 on the same machine, respectively. The higher capacity model
enjoys a roughly 4% boost to accuracy and robustness.

8 Conclusions
Adversarial training is a well-studied method that boosts the robustness and interpretability of neural
networks. While it remains one of the few effective ways to harden a network to attacks, few can
afford to adopt it because of its high computation cost. We present a “free” version of adversarial
training with cost nearly equal to natural training. Free training can be further combined with
other defenses to produce robust models without a slowdown. We hope that this approach can put
adversarial training within reach for organizations with modest compute resources.
Acknowledgements: Goldstein and his students were supported by DARPA GARD, DARPA QED
for RML, DARPA L2M, and the YFA program. Additional support was provided by the AFOSR
MURI program. Davis and his students were supported by the Office of the Director of National
Intelligence (ODNI), and IARPA (2014-14071600012). Studer was supported by Xilinx, Inc. and
the US NSF under grants ECCS-1408006, CCF-1535897, CCF-1652065, CNS-1717559, and ECCS-
1824379. Taylor was supported by the Office of Naval Research (N0001418WX01582) and the
Department of Defense High Performance Computing Modernization Program. The views and
conclusions contained herein are those of the authors and should not be interpreted as necessarily
representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA,
or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for
Governmental purposes notwithstanding any copyright annotation thereon.

9
References
Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio
Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML-PKDD,
pages 387–402. Springer, 2013.
Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow,
and Rob Fergus. Intriguing properties of neural networks. ICLR, 2013.
Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck,
Dawn Song, Michael E Houle, and James Bailey. Characterizing adversarial subspaces using local
intrinsic dimensionality. arXiv preprint arXiv:1801.02613, 2018.
Dongyu Meng and Hao Chen. Magnet: a two-pronged defense against adversarial examples. In
Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security,
pages 135–147. ACM, 2017.
Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep
neural networks. arXiv preprint arXiv:1704.01155, 2017.
Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of
security: Circumventing defenses to adversarial examples. ICML, 2018.
Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu.
Towards deep learning models resistant to adversarial attacks. ICLR, 2017.
Shumeet Baluja and Ian Fischer. Adversarial transformation networks: Learning to generate adver-
sarial examples. AAAI, 2018.
Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturba-
tions. CVPR, 2018.
Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial
examples with adversarial networks. IJCAI, 2018.
Ali Shafahi, Amin Ghiasi, Furong Huang, and Tom Goldstein. Label smoothing and logit squeezing:
A replacement for adversarial training?, 2019a.
Marius Mosbach, Maksym Andriushchenko, Thomas Trost, Matthias Hein, and Dietrich Klakow.
Logit pairing methods can fool gradient-based attacks. arXiv preprint arXiv:1810.12042, 2018.
Andrew Slavin Ross and Finale Doshi-Velez. Improving the adversarial robustness and interpretability
of deep neural networks by regularizing their input gradients. In AAAI, 2018.
Matthias Hein and Maksym Andriushchenko. Formal guarantees on the robustness of a classifier
against adversarial manipulation. In NeurIPS, pages 2266–2276, 2017.
Daniel Jakubovitz and Raja Giryes. Improving dnn robustness to adversarial attacks using jacobian
regularization. In ECCV, pages 514–529, 2018.
Fuxun Yu, Chenchen Liu, Yanzhi Wang, and Xiang Chen. Interpreting adversarial robustness: A
view from decision surface in input space. arXiv preprint arXiv:1810.00144, 2018.
Eric Wong and J Zico Kolter. Provable defenses against adversarial examples via the convex outer
adversarial polytope. ICML, 2017.
Eric Wong, Frank Schmidt, Jan Hendrik Metzen, and J Zico Kolter. Scaling provable adversarial
defenses. In NeurIPS, pages 8400–8409, 2018.
Aditi Raghunathan, Jacob Steinhardt, and Percy S Liang. Semidefinite relaxations for certifying
robustness to adversarial examples. In NeurIPS, pages 10877–10887, 2018a.
Aditi Raghunathan, Jacob Steinhardt, and Percy Liang. Certified defenses against adversarial
examples. arXiv preprint arXiv:1801.09344, 2018b.

10
Shiqi Wang, Yizheng Chen, Ahmed Abdou, and Suman Jana. Mixtrain: Scalable training of formally
robust neural networks. arXiv preprint arXiv:1811.02625, 2018.
Mathias Lecuyer, Vaggelis Atlidakis, Roxana Geambasu, Daniel Hsu, and Suman Jana. Certified
robustness to adversarial examples with differential privacy. arXiv preprint arXiv:1802.03471,
2018.
Bai Li, Changyou Chen, Wenlin Wang, and Lawrence Carin. Second-order adversarial attack and
certifiable robustness. CoRR, abs/1809.03113, 2018a. URL http://arxiv.org/abs/1809.
03113.
Jeremy M Cohen, Elan Rosenfeld, and J Zico Kolter. Certified adversarial robustness via randomized
smoothing. arXiv preprint arXiv:1902.02918, 2019.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical
Image Database. In CVPR09, 2009.
Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan Yuille, and Kaiming He. Feature denoising for
improving adversarial robustness. CVPR, 2019.
Harini Kannan, Alexey Kurakin, and Ian Goodfellow. Adversarial logit pairing. arXiv preprint
arXiv:1803.06373, 2018.
Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial
examples. ICLR, 2015.
Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world.
arXiv preprint arXiv:1607.02533, 2016a.
Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick Mc-
Daniel. Ensemble adversarial training: Attacks and defenses. arXiv preprint arXiv:1705.07204,
2017.
Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. ICLR,
2016b.
Ji Lin, Chuang Gan, and Song Han. Defensive quantization: When efficiency meets robustness.
ICLR, 2019.
Jacob Buckman, Aurko Roy, Colin Raffel, and Ian Goodfellow. Thermometer encoding: One hot
way to resist adversarial examples. ICLR, 2018.
Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend:
Leveraging generative models to understand and defend against adversarial examples. arXiv
preprint arXiv:1710.10766, 2017.
Andrew Ilyas, Ajil Jalal, Eirini Asteri, Constantinos Daskalakis, and Alexandros G Dimakis.
The robust manifold defense: Adversarial training using generative models. arXiv preprint
arXiv:1712.09196, 2017.
Haifeng Qian and Mark N Wegman. L2-nonexpansive neural networks. arXiv preprint
arXiv:1802.07896, 2018.
Ali Shafahi, Mahyar Najibi, Zheng Xu, John Dickerson, Larry S Davis, and Tom Goldstein. Universal
adversarial training. arXiv preprint arXiv:1811.11304, 2018.
Guneet S Dhillon, Kamyar Azizzadenesheli, Zachary C Lipton, Jeremy Bernstein, Jean Kossaifi, Aran
Khanna, and Anima Anandkumar. Stochastic activation pruning for robust adversarial defense.
arXiv preprint arXiv:1803.01442, 2018.
Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146,
2016.
Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry.
Robustness may be at odds with accuracy. ICLR, 1050:11, 2018.

11
Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan.
Theoretically principled trade-off between robustness and accuracy. ICML, 2019a.
Ali Shafahi, W Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial
examples inevitable? ICLR, 2019b.
Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017
IEEE Symposium on Security and Privacy (SP), pages 39–57. IEEE, 2017.
Jonathan Uesato, Brendan O’Donoghue, Aaron van den Oord, and Pushmeet Kohli. Adversarial risk
and the dangers of evaluating against weak attacks. arXiv preprint arXiv:1802.05666, 2018.
Dinghuai Zhang, Tianyuan Zhang, Yiping Lu, Zhanxing Zhu, and Bin Dong. You only propagate
once: Painless adversarial training using maximal principle. arXiv preprint arXiv:1905.00877,
2019b.
Logan Engstrom, Andrew Ilyas, and Anish Athalye. Evaluating and understanding the robustness of
adversarial logit pairing. arXiv preprint arXiv:1807.10272, 2018.
Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape
of neural nets. In NeurIPS, pages 6389–6399, 2018b.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang,
Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition
challenge. IJCV, 115(3):211–252, 2015.
Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial
images using input transformations. arXiv preprint arXiv:1711.00117, 2017.
Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against
adversarial attacks using high-level representation guided denoiser. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 1778–1787, 2018.
Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects
through randomization. arXiv preprint arXiv:1711.01991, 2017.

12