OpenAI, UC Berkeley
Most existing adversarial defenses only measure robustness to Lp adversarial attacks. Not only are adversaries unlikely to exclusively create small Lp perturbations, adversaries are unlikely to remain fixed. Adversaries adapt and evolve their attacks; hence adversarial defenses must be robust to a broad range of unforeseen attacks. We address this discrepancy between research and reality by proposing a new evaluation framework called ImageNet-UA. Our framework enables the research community to test ImageNet model robustness against attacks not encountered during training. To create ImageNet-UA’s diverse attack suite, we introduce a total of four novel adversarial attacks. We also demonstrate that, in comparison to ImageNet-UA, prevailing L robustness assessments give a narrow account of adversarial robustness. By evaluating current defenses with ImageNet-UA, we find they provide little robustness to unforeseen attacks. We hope the greater variety and realism of ImageNet-UA enables development of more robust defenses which can generalize beyond attacks seen during training.
Neural networks perform well on many datasets  yet can be consistently fooled by minor adversarial distortions . The research community has responded by quantifying and developing adversarial defenses against such attacks , yet these defenses and metrics have two key limitations.
First, the vast majority of existing defenses exclusively defend against and quantify robustness to Lp-constrained attacks [33, 11, 43, 58]. Though real-world adversaries are not Lp constrained  and can attack with diverse distortions [5, 49], the literature largely ignores this and evaluates against the Lp adversaries already seen during training [33, 58], resulting in optimistic robustness assessments. The attacks outside the Lp threat model that have been proposed [51, 42, 14, 61, 15, 48] are not intended for general defense evaluation and suffer from narrow dataset applicability, difficulty of optimization, or fragility of auxiliary generative models.
Second, existing defenses assume that attacks are known in advance  and use knowledge of their explicit form during training . In practice, adversaries can deploy unforeseen attacks not known to the defense creator. For example, online advertisers use attacks such as perturbed pixels in ads to defeat ad blockers trained only on the previous generation of ads in an ever-escalating arms race . However, current evaluation setups implicitly assume that attacks encountered at test-time are the same as those seen at train-time, which is unrealistic. The reality that future attacks are unlike those encountered during training is akin to a train-test distribution mismatch—a problem studied outside of adversarial robustness [45, 25]—but we now bring this idea to the adversarial setting.
The present work addresses these limitations by proposing an evaluation framework ImageNet-UA to measure robustness against unforeseen attacks. ImageNet-UA assesses a defense which may have been created with knowledge of the commonly used L or L2 attacks with six diverse attacks (four of which are novel) distinct from L or L2. We intend these attacks to be used at test-time only and not during training. Performing well on ImageNet-UA thus demonstrates generalization to a diverse set of distortions not seen during defense creation. While ImageNet-UA does not provide an exhaustive guarantee over all conceivable attacks, it evaluates over a diverse unforeseen test distribution similar to those used successfully in other studies of distributional shift [44, 25, 45]. ImageNet-UA works for ImageNet models and can be easily used with our code available at https://github.com/ddkang/advex-uar.
Designing ImageNet-UA requires new attacks that are strong and varied, since real-world attacks are diverse in structure. To meet this challenge, we contribute four novel and diverse adversarial attacks, in contrast to prior papers offering only one [4, 1, 14, 57]. Our new attacks produce distortions with occlusions, spatial similarity, and simulated weather, all of which are absent in previous attacks. Performing well on ImageNet-UA thus demonstrates that a defense generalizes to a diverse set of distortions distinct from the commonly used L∞ or L2.
With ImageNet-UA, we show marked weaknesses in existing evaluation practices and defenses through a study of 8 attacks against 48 models adversarially trained on ImageNet-100, a 100-class subset of ImageNet . While most adversarial robustness evaluations use only L attacks, ImageNet-UA reveals that models with high L attack robustness can remain susceptible to other attacks. This implies that L evaluations are a narrow measure of robustness, even though much of the literature treats this evaluation as comprehensive [33, 40, 47, 60]. We address this deficiency by using the novel attacks in ImageNet-UA to evaluate robustness to a more diverse set of unforeseen attacks. Moreover, our results demonstrate that L adversarial training, the current state-of-the-art defense, has limited generalization to unforeseen adversaries, and is not easily improved by training against more attacks. This adds to the evidence that achieving robustness against a few train-time attacks is insufficient to impart robustness to unforeseen test-time attacks [29,30,53].
In summary, we propose the framework ImageNet-UA to measure robustness to a diverse set of attacks, made possible by our four new adversarial attacks. Since existing defenses scale poorly to multiple attacks [30, 53], finding defense techniques which generalize to unforeseen attacks is crucial to create robust models. We suggest ImageNet-UA as a way to measure progress towards this goal.
2 Related Work
Adversarial robustness is notoriously difficult to correctly evaluate [39, 2]. To that end, Carlini et al.  provide extensive guidance for sound adversarial robustness evaluation. By measuring attack success rates across several distortion sizes and using a broader threat model with diverse differentiable attacks, ImageNet-UA has several of their recommendations built-in. Previous work on
evaluation considers small sets of fixed attacks. DeepFool  and CLEVER  estimate empirical robustness, the expected minimum ε needed to successfully attack an image. They apply only to attacks optimizing over an Lp-ball of radius ε, and CLEVER is susceptible to gradient masking . Wu et al.  evaluate against physically-realizable attacks from Evtimov et al.  and Sharif et al. , thus using a threat model restricted to occlusion attacks on narrow datasets.
Prior attacks outside the Lp threat model exist, but lack the general applicability and fast optimization of ours. Song et al.  attack using variational autoencoders, yet the attacks are weak and require simple image distributions suitable for VAEs. Qiu et al.  create adversarial images with a StarGAN, which is subject to GAN instabilities. Engstrom et al.  apply Euclidean transformations determined by brute-force search. Zhao et al.  use perceptual color distances to align human perception and L2 perturbations. Evtimov et al.  and Sharif et al.  attack stop signs and face- recognition systems with carefully placed patches or modified eyeglass frames, requiring physical object creation and applying only to specific image types. In contrast, our attacks are fast by virtue of differentiability, broadly applicable, and independent of auxiliary generative models.
3 New Attacks for a Broader Threat Model
There are few diverse, easily optimizable, plug-and-play adversarial attacks in the current literature; outside of Elastic , most are Lp attacks such as L , L2 [52, 6], L1 . We rectify this deficiency with four novel adversarial attacks: JPEG, Fog, Snow, and Gabor. Our attacks are differentiable and fast, while optimizing over enough parameters to be strong. We show example adversarial images in Figure 1 and compare stochastic and adversarial distortions in Figure 2.
Our novel attacks provide a broad range of test-time adversaries distinct from L or L2 attacks. They are intended as unforeseen attacks not used during training, allowing them to evaluate whether a de- fense can generalize from L or L2 to a much more varied set of distortions than current evaluations. Though our attacks are not exhaustive, performing well against them already demonstrates robustness to occlusion, spatial similarity, and simulated weather, all of which are absent from previous attacks.
Our attacks create an adversarial image xi from a clean image x with true label y. Let model f map images to a softmax distribution, and let £(f (x), y) be the cross-entropy loss. Given a target class yi /= y, our attacks attempt to find a valid image xi such that (1) the attacked image xi is obtained
by applying a distortion (of size controlled by a parameter ε) to x, and (2) the loss £(f (xi), yi) is minimized. An unforeseen adversarial attack is a white- or black-box adversarial attack unknown to the defense designer which does not change the true label of x according to an oracle or human.
3.1 Four New Unforeseen Attacks
JPEG. JPEG applies perturbations in a JPEG-encoded space of compressed images rather than raw pixel space. After color-space conversion, JPEG encodes small image patches using the discrete cosine transform. It then uses projected gradient descent to find an L -constrained adversarial perturbation in the resulting frequency space.The perturbed frequency coefficients are quantized and reverse-transformed to obtain the image in pixel space. We use ideas from Shin and Song  to make this differentiable. The resulting attack is conspicuously distinct from Lp attacks.
Fog. Fog simulates worst-case weather conditions. Robustness to adverse weather is a safety critical priority for autonomous vehicles, and Figure 2 shows Fog provides a more rigorous stress-test than stochastic fog . Fog creates adversarial fog-like occlusions by adversarially optimizing parameters in the diamond-square algorithm  typically used to render stochastic fog effects.
Snow. Snow simulates snowfall with occlusions of randomly located small image regions rep- resenting snowflakes. It adversarially optimizes their intensity and direction. Making Snow fast and differentiable is non-trivial and hinges on the use of an exponential distribution for snowflake intensities. Compared to synthetic stochastic snow , our adversarial snow is faster and includes snowflakes at differing angles instead of one fixed angle. Figure 2 shows adversarial snow exposes model weaknesses more effectively than the easier stochastic, average-case snow.
Gabor. Gabor spatially occludes the image with visually diverse Gabor noise . Gabor adversarially optimizes semantically meaningful parameters (orientation, bandwidth, etc.) to create different Gabor kernels used in Gabor noise. While rendering Gabor noise, we use spectral variance normalization  and initialize our optimization parameters with a sparse random matrix.
3.2 Improving Existing Attacks
Elastic modifies the attack of Xiao et al. ; it warps the image by distortions xi = Flow(x, V ), where V : 1, . . . , 224 2 R2 is a vector field on pixel space, and Flow sets the value of pixel (i, j) to the bilinearly interpolated original value at (i, j) + V (i, j). We construct V by smoothing a vector field W by a Gaussian kernel (size 25 × 25, σ ≈ 3 for a 224 × 224 image) and optimize W under IW (i, j)I∞ ≤ ε for all i, j. The resulting attack is suitable for large-scale images.
The other three attacks are L1, L2, L attacks, but we improve the L1 attack. For L and L2 constraints, we use randomly-initialized projected gradient descent (PGD), which applies gradient descent and projection to the L and L2 balls . Projection is difficult for L1, and previous L1 attacks resort to heuristics [9, 53]. We replace PGD with the Frank-Wolfe algorithm , which optimizes a linear function instead of projecting at each step (pseudocode in Appendix D). This makes our L1 attack more principled than previous implementations.
4 ImageNet-UA: Measuring Robustness to Unforeseen Attacks
We propose the framework ImageNet-UA and its CIFAR-10 analogue CIFAR-10-UA to measure and summarize model robustness while fulfilling the following desiderata:
- Defenses should be evaluated against a broad threat model through a diverse set of attacks.
- Defenses should exhibit generalization to attacks not exactly identical to train-time attacks.
The range of distortion sizes used for an attack must be wide enough to avoid misleading conclusions caused by overly weak or strong versions of that attack (Figure 3).
The ImageNet-UA evaluation framework aggregates robustness information into a single measure, the mean Unforeseen Adversarial Robustness (mUAR). The mUAR is an average over six different attacks of the Unforeseen Adversarial Robustness (UAR), a metric which assesses the robustness of a defense against a specific attack by using a wide range of distortion sizes. UAR is normalized using a measure of attack strength, the ATA, which we now define.
Adversarial Training Accuracy (ATA). The Adversarial Training Accuracy ATA(A, ε) estimates the strength of an attack A against adversarial training , one of the strongest currently known defense methods. For a distortion size ε, it is the best adversarial test accuracy against A achieved by adversarial training against A. We allow a possibly different distortion size εi during training, since this sometimes improves accuracy, and we choose a fixed architecture for each dataset.
For ImageNet-100, we choose ResNet-50 for the architecture, and for CIFAR-10 we choose ResNet-56. When evaluating a defense with architecture other than ResNet-50 or ResNet-56, we recommend using ATA values computed with these architectures to enable consistent comparison. To estimate ATA(A, ε) in practice, we evaluate models adversarially trained against distortion size εi for εi in a large range (we describe this range at this section’s end).
UAR: Robustness Against a Single Attack. The UAR, a building block for the mUAR, averages a model’s robustness to a single attack over six distortion sizes ε1, . . . , ε6 chosen for each attack (we describe the selection procedure at the end of this section). It is defined as
where Acc(A, εk, M ) is the accuracy Acc(A, εk, M ) of a model M after attack A at distortion size εk. The normalization in (1) makes attacks of different strengths more commensurable in a stable way. We give values of ATA(A, εk) and εk for our attacks on ImageNet-100 and CIFAR-10 in Tables 4 and 5 (Appendix B), allowing computation of UAR of a defense against a single attack with six adversarial evaluations and no adversarial training.
mUAR: Mean Unforeseen Attack Robustness. We summarize a defense’s performance on ImageNet-UA with the mean Unforeseen Attack Robustness (mUAR), an average of UAR scores for the L1, Elastic, JPEG, Fog, Snow, and Gabor attacks:
Our measure mUAR estimates robustness to a broad threat model containing six unforeseen attacks at six distortion sizes each, meaning high mUAR requires generalization to several held-out attacks. In particular, it cannot be achieved by the common practice of engineering defenses to a single attack, which Figure 4 shows does not necessarily provide robustness to different attacks.
Our four novel attacks play a crucial role in mUAR by allowing us to estimate robustness to a sufficiently large set of adversarial attacks. As is customary when studying train-test mismatches and distributional shift, we advise against adversarially training with these six attacks when evaluating ImageNet-UA to preserve the validity of mUAR, though we encourage training with other attacks.
Distortion Size Choices. We explain the εi values used to estimate ATA and the choice of ε1, . . . , ε6 used to define UAR. This calibration of distortion sizes adjusts for the fact (Figure 3) that adversarial robustness against an attack may vary drastically with distortion size. Further, the relation between
distortion size and attack strength varies between attacks, so too many or too few εk values in a certain range may cause an attack to appear artificially strong or weak according to UAR.
We choose distortion sizes between minimum and maximum values εmin and εmax defined as follows:
- The minimum distortion size εmin is the largest ε for which the adversarial accuracy of an adversarially trained model at distortion size ε is comparable to that of a model trained and evaluated on unattacked data (for ImageNet-100, within 3 of 87).
- The maximum distortion size εmax is the smallest ε which either reduces the adversarial accuracy of an adversarially trained model at distortion size ε below 25 or yields images which confuse humans (adversarial accuracy can remain non-zero in this case).
As is typical in recent work on adversarial examples [3, 15, 13, 41], our attacks are perceptible at large distortion sizes, reflecting the perceptibility of attacks in real world threat models suggested by Gilmer et al. .
For ATA, we evaluate against models adversarially trained with εi increasing geometrically from εmin to εmax by factors of 2. We then choose εk as follows: We compute ATA at ε increasing geometrically from εmin to εmax by factors of 2 and take the size-6 subset whose ATA values have minimum £1-distance to the ATA values of the L attack in Table 4 (Appendix B.1). For example,
for Gabor, (εmin, εmax) = (6.25, 3200), so we compute ATAs at the 10 values ε = 6.25, . . . , 3200. Viewing size-6 subsets of the ATAs as vectors with decreasing coordinates, we select εk for Gabor corresponding to the vector with minimum £1-distance to the ATA vector for L∞.
5 New Insights From ImageNet–UA
We use ImageNet-UA to assess existing methods for adversarial defense and evaluation. First, ImageNet-UA reveals that L trained defenses fail to generalize to different attacks, indicating substantial weakness in current L adversarial robustness evaluation. We establish a baseline for ImageNet-UA using L2 adversarial training which is difficult to improve upon by adversarial training alone. Finally, we show non-adversarially trained models can still improve robustness on ImageNet-UA over standard models and suggest this as a direction for further inquiry.
5.1 Experimental Setup
We adversarially train 48 models against the 8 attacks from Section 3 and evaluate against targeted attacks. We use the CIFAR-10 and ImageNet-100 datasets for ImageNet-UA and CIFAR-10-UA. ImageNet-100 is a 100-class subset of ImageNet-1K  containing every tenth class by WordNet ID order; we use a subset of ImageNet-1K due to the high compute cost of adversarial training
Table 1: Clean Accuracy, UAR, and mUAR scores for models adversarially trained against L and L2 attacks. L training, the most popular defense, provides less robustness than L2 training. Comparing the highest mUAR achieved to individual UAR values in Figure 4 indicates a large robustness gap.
Table 2: Clean Accuracy, UAR, and mUAR scores for models jointly trained against (L , L2). Joint training does not provide much additional robustness.
To adversarially train  against attack A, at each mini-batch we select a uniform random (incorrect) target class for each training image. For maximum distortion size ε, we apply targeted attack A to the current model with distortion size εi Uniform(0, ε) and take a SGD step using only the attacked images. Randomly scaling εi improves performance against smaller distortions.
We train on 10-step attacks for attacks other than Elastic, where we use 30 steps due to a harder optimization. For Lp, JPEG, and Elastic, we use step size ε/ steps; for Fog, Gabor, and Snow, we use step size 0.001/steps because the latent space is independent of ε. These choices have optimal rates for non-smooth convex functions [36, 37]. We evaluate on 200-step targeted attacks with uniform random (incorrect) target, using more steps for evaluation than training per best practices .
5.2 ImageNet-UA Reveals Weaknessess in L∞ Training and Testing
We use ImageNet-UA to reveal weaknesses in the common practices of L∞ robustness evaluation and L∞ adversarial training. We compute the mUAR and UAR(L∞) for models trained against the L∞ attack with distortion size ε and show results in Figure 5. For small ε ≤ 4, mUAR and UAR(L∞) increase together with ε. For larger ε ≥ 8, UAR(L∞) continues to increase with ε, but the mUAR decreases, a fact which is not apparent from L∞ evaluation.
The decrease in mUAR while UAR(L ) increases suggests that L adversarial training begins to heavily fit L distortions at the expense of generalization at larger distortion sizes. Thus, while it is the most commonly used defense procedure, L training may not lead to improvements on other attacks or to real-world robustness.
Worse, L evaluation against L adversarial training at higher distortions indicates higher robustness. In contrast, mUAR reveals that L adversarial training at higher distortions in fact hurts robustness against a more diverse set of attacks. Thus, L evaluation gives a misleading picture of robustness. This is particularly important because L evaluation is the most ubiquitous measure of robustness in deep learning [22,33,58].
Table 3: Non-adversarial defenses can noticeably improve ImageNet-UA performance. ResNeXt- 101 (32 8d) + WSL is a ResNeXt-101 trained on approximately 1 billion images . Stylized ImageNet is trained on a modification of ImageNet using style transfer . Patch Gaussian augments using Gaussian distortions on small portions of the image . AugMix mixes simple random augmentations of the image . These results suggest a complementary avenue toward ImageNet-UA performance may be through non-adversarial defenses.
5.3 Limits of Adversarial Training for ImageNet-UA
We establish a baseline on ImageNet-UA using L2 adversarial training but show a significant perfor- mance gap even for more sophisticated existing adversarial training methods. To do so, we evaluate several adversarial training methods on ImageNet-UA and show results in Table 1.
Our results show that L2 trained models outperform L trained models and have significantly improved absolute performance, increasing mUAR from 14.0 to 50.7 compared to an undefended model. The individual UAR values in Figure 7 (Appendix E.1) improve substantially against all attacks other than Fog, including several (Elastic, Gabor, Snow) of extremely different nature to L2.
This result suggests pushing adversarial training further by training against multiple attacks simulta- neously via joint adversarial training [30, 53] detailed in Appendix C. Table 2 shows that, despite using twice the compute of L2 training, (L , L2) joint training only improves the mUAR from 50.7 to 50.9. We thus recommend L2 training as a baseline for ImageNet-UA, though there is substantial room for improvement compared to the highest UARs against individual attacks in Figure 4, which are all above 80 and often above 90.
5.4 ImageNet-UA Robustness through Non-Adversarial Defenses
We find that methods can improve robustness to unforeseen attacks without adversarial training. Table 3 shows mUAR for diverse architectures including SqueezeNet , ResNeXts , and ResNets. For ImageNet-1K models, we predict ImageNet-100 classes by masking 900 logits.
A popular defense against average case distortions  is Stylized ImageNet , which modifies training images using image style transfer in hopes of making networks rely less on textural features. Table 3 shows it provides some improvement on ImageNet-UA. More recently, Lopes et al.  propose to train against Gaussian noise applied to small image patches, improving the mUAR by 3% over the ResNet-50 baseline. The second largest mUAR improvement comes from training a ResNeXt on approximately 1 billion images . This three orders of magnitude increase in training data yields a 5.4% mUAR increase over a vanilla ResNeXt baseline. Finally, Hendrycks et al.  create AugMix, which randomly mixes stochastically generated augmentations. Although AugMix did not use random nor adversarial noise, it improves robustness to unforeseen attacks by 10%.
These results imply that defenses not relying on adversarial examples can improve ImageNet-UA performance. They indicate that training on more data only somewhat increases robustness on ImageNet-UA, quite unlike many other robustness benchmarks [25, 26] where more data helps tremendously . While models with lower clean accuracy including SqueezeNet and ResNet-18 oddly have higher UAR(L ) and UAR(L2) than many other models, there is no clear difference in mUAR. Last, these non-adversarial defenses do not come at a large cost to accuracy on clean examples, unlike adversarial defenses. Much remains to explore, and we hope non-adversarial defenses will be a promising avenue toward adversarial robustness.
This work proposes a framework ImageNet-UA to evaluate robustness of a defense against unforeseen attacks. Because existing adversarial defense techniques do not scale to multiple attacks, developing models which can defend against attacks not seen at train-time is essential for robustness. Our results using ImageNet-UA show that the common practice of L training and evaluation fails to achieve or measure this broader form of robustness. As a result, it can provide a misleading sense of robustness. By incorporating our 4 novel and strong adversarial attacks, ImageNet-UA enables evaluation on the diverse held-out attacks necessary to measure progress towards robustness more broadly.
D. K., Y. S., and J. S. were supported by a grant from the Open Philanthropy Project. D. K. was supported by NSF Grant DGE-1656518. Y. S. was supported by a Junior Fellow award from the Simons Foundation and NSF Grant DMS-1701654. D. H. was supported by NSF Frontier Award 1804794. Work by D. K. and Y. S. was partially done at OpenAI.
-  A. Athalye, L. Engstrom, A. Ilyas, and K. Kwok. Synthesizing robust adversarial examples. CoRR, abs/1707.07397, 2017. URL http://arxiv.org/abs/1707.07397.
-  A. Athalye, N. Carlini, and D. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. arXiv preprint arXiv:1802.00420, 2018.
-  A. Athalye, L. Engstrom, A. Ilyas, and K. Kwok. Synthesizing robust adversarial examples. In J. Dy and A. Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 284–293, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http://proceedings.mlr.press/v80/athalye18b.html.
-  W. Brendel, J. Rauber, and M. Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248, 2017.
-  T. B. Brown, D. Mané, A. Roy, M. Abadi, and J. Gilmer. Adversarial patch. CoRR, abs/1712.09665, 2017. URL http://arxiv.org/abs/1712.09665.
-  N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pages 39–57. IEEE, 2017.
-  N. Carlini, A. Athalye, N. Papernot, W. Brendel, J. Rauber, D. Tsipras, I. G. Goodfellow, and A. Madry. On evaluating adversarial robustness: Principles of rigorous evaluations. 2019.
-  N. Carlini, A. Athalye, N. Papernot, W. Brendel, J. Rauber, D. Tsipras, I. J. Goodfellow, A. Madry, and A. Kurakin. On evaluating adversarial robustness. CoRR, abs/1902.06705, 2019. URL http://arxiv.org/abs/1902.06705.
-  P.-Y. Chen, Y. Sharma, H. Zhang, J. Yi, and C.-J. Hsieh. EAD: Elastic-net attacks to deep neural networks via adversarial examples. In Thirty-second AAAI conference on artificial intelligence, 2018.
-  K. T. Co, L. Muñoz-González, and E. C. Lupu. Sensitivity of deep convolutional networks to Gabor noise. CoRR, abs/1906.03455, 2019. URL http://arxiv.org/abs/1906.03455.
-  J. M. Cohen, E. Rosenfeld, and J. Z. Kolter. Certified adversarial robustness via randomized smoothing. CoRR, abs/1902.02918, 2019. URL http://arxiv.org/abs/1902.02918.
-  J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009.
-  Y. Dong, T. Pang, H. Su, and J. Zhu. Evading defenses to transferable adversarial examples by translation- invariant attacks. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2019.
-  L. Engstrom, B. Tran, D. Tsipras, L. Schmidt, and A. Madry. A rotation and a translation suffice: Fooling CNNs with simple transformations. arXiv preprint arXiv:1712.02779, 2017.
-  I. Evtimov, K. Eykholt, E. Fernandes, T. Kohno, B. Li, A. Prakash, A. Rahmati, and D. X. Song. Robust physical-world attacks on deep learning models. 2017.
-  A. Fournier, D. Fussell, and L. Carpenter. Computer rendering of stochastic models. Commun. ACM, 25 (6):371–384, June 1982. ISSN 0001-0782. doi: 10.1145/358523.358553. URL http://doi.acm.org/ 10.1145/358523.358553.
-  M. Frank and P. Wolfe. An algorithm for quadratic programming. Naval research logistics quarterly, 3 (1-2):95–110, 1956.
-  R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Con- ference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bygh9j09KX.
-  J. Gilmer, R. P. Adams, I. J. Goodfellow, D. Andersen, and G. E. Dahl. Motivating the rules of the game for adversarial example research. ArXiv, abs/1807.06732, 2018.
-  I. Goodfellow. Gradient masking causes CLEVER to overestimate adversarial perturbation size. arXiv preprint arXiv:1804.07870, 2018.
-  I. J. Goodfellow. A research agenda: Dynamic models to defend against correlated attacks. ArXiv, abs/1903.06293, 2019.
-  I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014.
-  P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski, A. Kyrola, A. Tulloch, Y. Jia, and K. He. Accurate, large minibatch SGD: Training ImageNet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.
-  K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630–645. Springer, 2016.
-  D. Hendrycks and T. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019.
-  D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt, and D. Song. Natural adversarial examples. arXiv preprint arXiv:1907.07174, 2019.
-  D. Hendrycks, N. Mu, E. D. Cubuk, B. Zoph, J. Gilmer, and B. Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. Proceedings of the International Conference on Learning Representations (ICLR), 2020.
-  F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer. Squeezenet: AlexNet-level accuracy with 50x fewer parameters and <1mb model size. ArXiv, abs/1602.07360, 2017.
-  J.-H. Jacobsen, J. Behrmannn, N. Carlini, F. Tramèr, and N. Papernot. Exploiting excessive invariance caused by norm-bounded adversarial robustness, 2019.
-  M. Jordan, N. Manoj, S. Goel, and A. G. Dimakis. Quantifying perceptual distortion of adversarial examples. arXiv e-prints, art. arXiv:1902.08265, Feb 2019.
-  A. Lagae, S. Lefebvre, G. Drettakis, and P. Dutré. Procedural noise using sparse Gabor convolution. ACM Trans. Graph., 28(3):54:1–54:10, July 2009. ISSN 0730-0301. doi: 10.1145/1531326.1531360. URL http://doi.acm.org/10.1145/1531326.1531360.
-  R. G. Lopes, D. Yin, B. Poole, J. Gilmer, and E. D. Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation. ArXiv, abs/1906.02611, 2019.
-  A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017.
-  D. Mahajan, R. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li, A. Bharambe, and L. van der Maaten. Exploring the limits of weakly supervised pretraining. In V. Ferrari, M. Hebert, C. Sminchisescu, and Y. Weiss, editors, Computer Vision – ECCV 2018, pages 185–201, Cham, 2018. Springer International Publishing. ISBN 978-3-030-01216-8.
-  S.-M. Moosavi-Dezfooli, A. Fawzi, and P. Frossard. DeepFool: a simple and accurate method to fool deep neural networks. arXiv preprint arXiv:1511.04599, 2015.
-  A. Nemirovski and D. Yudin. On Cezari’s convergence of the steepest descent method for approximating saddle point of convex-concave functions. In Soviet Math. Dokl, volume 19, pages 258–269, 1978.
-  A. Nemirovski and D. Yudin. Problem Complexity and Method Efficiency in Optimization. Intersci. Ser. Discrete Math. Wiley, New York, 1983.
-  A. E. Orhan. Robustness properties of Facebook’s ResNeXt WSL models. ArXiv, abs/1907.07640, 2019.
-  N. Papernot, P. McDaniel, I. Goodfellow, S. Jha, Z. B. Celik, and A. Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pages 506–519. ACM, 2017.
-  H. Qian and M. N. Wegman. L2-nonexpansive neural networks. In International Conference on Learning Representations (ICLR), 2019. URL https://openreview.net/forum?id=ByxGSsR9FQ.
-  C. Qin, J. Martens, S. Gowal, D. Krishnan, K. Dvijotham, A. Fawzi, S. De, R. Stanforth, and P. Kohli. Adversarial robustness through local linearization, 2019.
-  H. Qiu, C. Xiao, L. Yang, X. Yan, H. Lee, and B. Li. Semanticadv: Generating adversarial examples via attribute-conditional image editing. ArXiv, abs/1906.07927, 2019.
-  E. Raff, J. Sylvester, S. Forsyth, and M. McLean. Barrage of random transforms for adversarially robust defense. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6528–6537, 2019.
-  P. Rajpurkar, R. Jia, and P. Liang. Know what you don’t know: Unanswerable questions for SQuAD. In Association for Computational Linguistics (ACL), 2018.
-  B. Recht, R. Roelofs, L. Schmidt, and V. Shankar. Do imagenet classifiers generalize to imagenet? In ICML, 2019.
-  O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. S. Bernstein, A. C. Berg, and F.-F. Li. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115:211–252, 2014.
-  L. Schott, J. Rauber, W. Brendel, and M. Bethge. Towards the first adversarially robust neural network model on MNIST. May 2019. URL https://arxiv.org/pdf/1805.09190.pdf.
-  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 23rd ACM SIGSAC Conference on Computer and Communications Security, 2016.
-  M. Sharif, S. Bhagavatula, L. Bauer, and M. K. Reiter. A general framework for adversarial examples with objectives. ACM Transactions on Privacy and Security (TOPS), 22(3):1–30, 2019.
-  R. Shin and D. Song. JPEG-resistant adversarial images. In NIPS 2017 Workshop on Machine Learning and Computer Security, 2017.
-  Y. Song, R. Shu, N. Kushman, and S. Ermon. Constructing unrestricted adversarial examples with generative models. In NeurIPS, 2018.
-  C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.
-  F. Tramèr and D. Boneh. Adversarial training and robustness for multiple perturbations. arXiv e-prints, art. arXiv:1904.13000, Apr 2019.
-  F. Tramèr, P. Dupré, G. Rusak, G. Pellegrino, and D. Boneh. Ad-versarial: Defeating perceptual ad-blocking. CoRR, abs/1811.03194, 2018. URL http://arxiv.org/abs/1811.03194.
-  T.-W. Weng, H. Zhang, P.-Y. Chen, J. Yi, D. Su, Y. Gao, C.-J. Hsieh, and L. Daniel. Evaluating the robustness of neural networks: An extreme value theory approach. arXiv preprint arXiv:1801.10578, 2018.
-  T. Wu, L. Tong, and Y. Vorobeychik. Defending against physically realizable attacks on image classification. In International Conference on Learning Representations, 2020. URL https://openreview.net/ forum?id=H1xscnEKDr.
-  C. Xiao, J.-Y. Zhu, B. Li, W. He, M. Liu, and D. Song. Spatially transformed adversarial examples. arXiv preprint arXiv:1801.02612, 2018.
-  C. Xie, Y. Wu, L. v. d. Maaten, A. Yuille, and K. He. Feature denoising for improving adversarial robustness. arXiv preprint arXiv:1812.03411, 2018.
-  S. Xie, R. B. Girshick, P. Dollár, Z. Tu, and K. He. Aggregated residual transformations for deep neural networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5987–5995, 2016.
-  H. Zhang, Y. Yu, J. Jiao, E. Xing, L. E. Ghaoui, and M. Jordan. Theoretically principled trade-off between robustness and accuracy. In K. Chaudhuri and R. Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 7472–7482, Long Beach, California, USA, 09–15 Jun 2019. PMLR. URL http://proceedings. mlr.press/v97/zhang19p.html.
-  Z. Zhao, Z. Liu, and M. Larson. Towards large yet imperceptible adversarial image perturbations with perceptual color distance. ArXiv, abs/1911.02466, 2019.
A Training hyperparameters
For ImageNet-100, we trained on machines with 8 NVIDIA V100 GPUs using standard data aug- mentation . Following best practices for multi-GPU training , we ran synchronized SGD for 90 epochs with batch size 32 8 and a learning rate schedule with 5 “warm-up” epochs and a decay at epochs 30, 60, and 80 by a factor of 10. Initial learning rate after warm-up was 0.1, momentum was 0.9, and weight decay was 10−4. For CIFAR-10, we trained on a single NVIDIA V100 GPU for
200 epochs with batch size 32, initial learning rate 0.1, momentum 0.9, and weight decay 10−4. We
decayed the learning rate at epochs 100 and 150.
B Calibration of ImageNet-UA and CIFAR-10-UA
B.1 Calibration for ImageNet-UA
Calibrated distortion sizes and ATA values are in Table 4.
B1 Calibration for CIFAR-10-UA
The ε calibration procedure for CIFAR-10 was similar to that used for ImageNet-100. We started with small εmin values and increased ε geometrically with ratio 2 until adversarial accuracy of an adversarially trained model dropped below 40. Note that this threshold is higher for CIFAR-10 than ImageNet-100 because there are fewer classes. The resulting ATA values for CIFAR-10 are shown in Table 5.
C Joint adversarial training
Our joint adversarial training procedure for two attacks A and Ai is as follows. At each training step, we compute the attacked image under both A and Ai and backpropagate with respect to gradients induced by the image with greater loss. This corresponds to the “max” loss of . We train ResNet-50 models for (L∞, L2), (L∞, L1), and (L∞, Elastic) on ImageNet-100.
Table 6 shows training against (L , L1) is worse than training against L1 at the same distortion size and performs particularly poorly at large distortion sizes. Table 7 shows joint training against (L , Elastic) also performs poorly, never matching the UAR score of training against Elastic at moderate distortion size (ε = 2).
D The Frank-Wolfe Algorithm
We chose to use the Frank-Wolfe algorithm for optimizing the L1 attack, as Projected Gradient Descent would require projecting onto a truncated L1 ball, which is a complicated operation. In contrast, Frank-Wolfe only requires optimizing linear functions gTx over a truncated L1 ball; this can be done by sorting coordinates by the magnitude of g and moving the top k coordinates to the boundary of their range (with k chosen by binary search). This is detailed in Algorithm 1.
E Full evaluation results
E.1 Full evaluation results and analysis for ImageNet-100
We show the full results of all adversarial attacks against all adversarial defenses for ImageNet-100 in Figure 6. These results also include L1-JPEG and L2-JPEG attacks, which are modifications of the JPEG attack applying Lp-constraints in the compressed JPEG space instead of L constraints. Full UAR scores are provided for ImageNet-100 in Figure 7.
E.2 Full evaluation results and analysis for CIFAR-10
We show the results of adversarial attacks and defenses for CIFAR-10 in Figure 8. We experienced difficulty training the L2 and L1 attacks at distortion sizes greater than those shown and have omitted those runs, which we believe may be related to the small size of CIFAR-10 images. Full UAR values for CIFAR-10 are shown in Figure 9.
F Robustness of our results