Recently, a series of papers have shown that deep neural networks (DNN) are vulnerable to various types of attacks (Szegedy, 2014),(Nguyen et al., 2015),(Moosavi-Dezfooli et al., 2017),(Brown et al., 2017),(Su et al., 2017), (Moosavi-Dezfooli et al., 2016).(Carlini & Wagner, 2017c),(Kurakin et al., 2016),(Sharif et al., 2016), (Athalye & Sutskever, 2018) However, the reasons underlying these vulnerabilities are still largely unknown. We argue that one of the most important facets of adversarial machine learning resides in its investigative nature. In other words, adversarial machine learning provides us with key tools to understand current DNNs. As much as attacks tells us about the security behind DNNs, they also tells us to what extent DNNs can reason over data and what do they understand to be, for example, the concept of a ”car” or ”horse”.
In this paper, inspired by recent attacks and defenses we propose a technique called propagation maps that would be able to explain most of them. Here, giving the constrained space, we focus on one attack which is puzzling and largely unexplained, the one-pixel attack. Propagation maps enable us to show how one pixel perturbation may grow in influence over the layers and spread over many pixels to cause a final change in class. Moreover, statistical properties of the propagation reveal many properties of the attacks as well as their distribution (Figure 1).
Additionally, to further understand the one-pixel attack a locality analysis is performed. The locality analysis consists of executing the attack in nearby pixels of a successful one-pixel attack, i.e., using the same pixel perturbation but different pixel position. Indeed, the success rate of nearby pixel is effective and equal among different neural networks (Figure 1), showing that rather than pixels or neurons, the vulnerability lies in some of the receptive fields. This reveals an interesting property shared among DNNs which is independent of the model or attack success rate.
2 Related Work
In the following subsections an overview of different types of attacks and a brief overview of the related works will be covered.
2.1 Adversarial Samples and Different Types of Attacks
The samples that can make machine learning algorithms misclassify received the name of adversarial samples. Let be the output of a machine learning algorithm in which is the input of the algorithm for input and output of sizes and respectively. It is possible to define adversarial samples x’ explicitly as follows:
in which is a small perturbation added to the input.
In adversarial machine learning one want to search for adversarial samples. For that it is possible to use the knowledge of the DNN in question to craft samples such as using back-propagation for obtaining gradient information and subsequently using gradient descent as done by the “fast gradient sign” proposed by I.J. Goodfellow et al. (Goodfellow et al., 2014a). There is also the greedy perturbation searching method proposed by S.M. Moosavi-Dezfooli et al. (Moosavi-Dezfooli et al., 2016) and N. Papernot et al. utilize Jacobian matrix of the function learned during training to create a saliency map which will guide the search for adversarial samples (Papernot et al., 2016a).
However, it is possible to search for adversarial samples without taking into account the internal characteristics of DNNs. This type of model agnostic search is also called black-box attack. To search for adversarial samples in this scenario, it is common to use perturbations that increase a given soft label (targetted attack) in which is the index of the target class. In other words, they can be defined as the following optimization problem:
Regarding untargeted attacks, the objective function can be defined, for example, as the minimization of the soft label for the outputted class . See the complete equation below.
The minimization of the difference between the highest soft-label index and the second highest one is also one of the other possibilities of objective function for untargeted attacks. There are many black-box attacks in the literature. To cite some (Narodytska & Kasiviswanathan, 2017),(Papernot et al., 2017),(Dang et al., 2017).
It is important to note that should be small enough to not allow an image to become a different class. Such a transformation would invalidate the adversarial sample creation because it is not a misclassification. Since most attacks use perturbations which comprise of the whole image,
is a good optimization constraint. However, it is also possible to look the other way around and deal with few perturbed dimensions. In this case, the constraint changes to a L0 norm which actually counts the dimensions of the the perturbation, i.e., the total number of non-zero elements in the perturbation vector. The complete equation is as follows:
where is a small number of dimensions ( for the one-pixel attack).
2.2 Recent Advances in Attacks and Defenses
The question of if machine learning is secure was asked some time ago (Barreno et al., 2006),(Barreno et al., 2010). However, it was only in 2013 that Deep Neural Networks’ (DNN) security was completely put into question (Szegedy, 2014). C. Szegedy et al. demonstrated that by adding noise to an image it is possible to produce a visually identical image which can make DNNs misclassify. This was conterintuitive, since the DNNs that misclassified had a very high accuracy in the tests rivaling even the accuracy of human beings.
Recently, the vulnerabilities of neural networks were shown to be even more aggravating. For example, A. Nguyen et al. did a series of experiments in which images were continuously developed to fool a machine learning algorithm. They found out that DNNs give high confidence results to random noise. DNNs also tend to classify patterns or textures, such as the stripes of a bus as a bus, with high confidency(Nguyen et al., 2015). Universal adversarial perturbations in which a single crafted perturbation is able to make a DNN misclassify multiple samples was also shown to be possible (Moosavi-Dezfooli et al., 2017). In (Brown et al., 2017), the authors showed that it is possible to make DNNs misclassify by adding a patch to an image. Interestingly, each patch is a targetted attack, i.e., a banana patch would make a DNN misclassify the image as a banana when added. Moreover, in (Su et al., 2017) it was shown that even one pixel could make DNNs’ misclassify. Indicating that although DNNs have a high accuracy in recognition tasks, their ”understanding” of what is a ”dog” or ”cat” is still very different from human beings. In fact, adversarial samples can be used to evaluate the robustness of a DNN (Moosavi-Dezfooli et al., 2016).(Carlini & Wagner, 2017c).
Although much of the research in adversarial machine learning is conducted under ideal conditions in a laboratory, the same techniques are not difficult to apply to real world scenarios because printed out adversarial samples still work, i.e., many adversarial samples are robust against different light conditions (Kurakin et al., 2016). In (Sharif et al., 2016), it was show that it is possible to build glasses that could fool DNNs to believe a given person is somebody else. The glasses were custom made and had some orientation in mind when designed. In fact, in (Athalye & Sutskever, 2018) the authors go a step further and verify the existence of 3d adversarial objects which can fool DNNs even when viewpoint, noise, and different light conditions are taking into consideration.
Defensive systems and detection systems were soon proposed to mitigate some of the problems. Regarding defensive systems, defensive distillation in which a smaller neural network squeezes the content learned by the original one was proposed as a defense(Papernot et al., 2016b) however it was soon shown to not be robust enough (Carlini & Wagner, 2017c). Adversarial training was also proposed in wchi adversarial images are added to the training dataset in such a way that the DNN will be able to correctly classify them, increasing its robustness (Goodfellow et al., 2014b),(Huang et al., 2015), (Madry et al., 2018). However this technique is still vulnerable to black-box attacks (Tramèr et al., 2018). There are many recent variations proposed recently (Ma et al., 2018), (Buckman et al., ), (Guo et al., ), (Dhillon et al., ), (Xie et al., ), (Song et al., ), (Samangouei et al., ), it is out of the scope of this article to review all of them. Most of these defenses create some sort of gradient masking, called obfuscated gradients. They are carefully analyzed and many of their shortcommings are explained in (Athalye et al., 2018)
Regarding detection systems, if it is hard to defend perhaps it would be possible to detect attacks. A study from (Grosse et al., 2017) revealed that indeed some adversarial samples have different statistical properties. Moreover, such a detection system would increase the cost of attacks. In (Xu et al., 2017), the authors propose an interesting method in which the classifier compares its prediction with a prediction of the same input but ”squeezed” (either color or spatial smoothing). This allow classifiers to detect with high accuracy adversarial samples. Having said that, many detection systems are subject to fail when adversarial samples differ from test conditions (Carlini & Wagner, 2017a),(Carlini & Wagner, 2017b). Thus, the increase in cost in part of the attacker seems certain but the clear benefits of detection systems remains unconclusive.
There are many works in attacks and defenses but the reason behind such lack of robustness for accurate classifiers is still largely unknown. In (Goodfellow et al., 2014b) it is argued that DNNs’ linearity are one of the main reasons. If this is the case, perhaps hybrid systems that can leverage the non-linearity that arise from complex models by using evolutionary based optimization techniques such as self-organizing classifiers (Vargas et al., 2013) and neuroevolution with unified neuron models (Vargas & Murata, 2017) would make for a promising investigation.
3 One-Pixel Attack
One-Pixel Attack investigated the opposite extreme of most attacks to date. Instead of searching for small spreaded perturbation, it focus on perturbing just one pixel. This vulnerability to one-pixel is a totally different scenario, i.e., neural networks that are vulnerable to usual attacks may not be vulnerable to one-pixel attack and vice-versa.
To achieve such an attack in a black-box scenario the authors used differential evolution which is a simple yet effective evolutionary (DE) algorithm (Storn & Price, 1997). A candidate solution is coded as a pixel position and its related perturbation. The DE search for promising candidate solutions by minimizing the output label of the correct class (Equation 3). In this paper, we use the same differential evolution settings as the original paper (Su et al., 2017). However, here we define a successful attack as an adversarial attack made over a correctly classified sample. As a consequence, adversarial attacks over already missclassified samples will be ignored.
4 Propagation Maps
Perturbation on the input image propagates throughout the neural network to change its class in adversarial samples. However, much of this process is unknown. In other words, how does this perturbation causes a change in the class label? What are the internal differences between adversarial attacks and failed attacks if any?
Here, to walk towards an answer to the questions above we propose a technique called propagation maps which can reveal the perturbation throughout the layers. Propagation maps consists of comparing the feature maps of both adversarial and original samples. Specifically, by calculating the difference between the feature maps and averaging them (or getting their maximum value) for each convolutional layer, the perturbation’s influence can be estimated. Consider an element-wise maximum of a three dimensional arrayfor indices , and to be described as:
where is the resulting two dimensinal array.
Therefore, for a layer , its respective propagation map can be obtained by:
where and are respectively the feature maps for layer and kernel of the natural (original) and adversarial samples.
Alternatively, one may wish to see the average over the filters which exposes a slightly different influence diluted over the kernels in the same layer. It can be computed as follows:
where is the number of filters.
5 Propagation Maps for the One-Pixel Attack
To investigate how a single pixel perturbation can cause changes in class, we will make use of the proposed propagation maps. This will allow us to vizualize the perturbations in each of the layers of the neural network.
For the experiments, Resnet, which is one of the most accurate types of neural networks, is used. Each of the subsections below investigate a specific scenario.
5.1 Single Pixel Perturbations that Change Class
In Figure 2, the propagation map (PMmax) of a successful one-pixel attack is shown. The perturbation is shown to start small and localized and then spread in deeper layers. In the last layer, the perturbation spread enough to influence strongly more than a quarter of the feature map. This is the element-wise maximum behavior which allows us to identify how strong is the maximum difference in feature maps.
The propagation map based on averaged differences (PMavg) show that the difference is concentraded in some feature maps (Figure 3). Moreover this average difference is kept more or less the same throughout the layers. In the case of PMmax, the difference had a sort of wave behavior, sometimes growing in strength, sometimes slightly fading away. All observed adversarial samples shared similar features of propagation maps. This is to be expected, since they need to influence enough in order to change the class.
Surprisingly, one pixel change can cause influences that spread over the entire feature map, specially in deeper layers. This also contradicts to some extent the expectation that high level features will be processed in deeper layers.
5.2 Single Pixel Perturbations that Do Not Change Class
Successful one-pixel attacks were shown to grow its influence throughout the layers, culminating in a strong and spread influence in the last layers. Here we change the position of the pixel to unable the attack to succeed. Figure 4 shows that in such a case the influence’s intensity decreases. In fact, in the last layer it is almost imperceptible the influence. However, this is not the rule. A counterexample is shown in Figure 5 in which a pixel is changed without changing the class label. This time however, the perturbation propagates strongly, being as strong if not stronger than the successful one-pixel attack observed in Figure 2. One might argue that the influence has caused the confidence to decrease but not enough to cause the change. For this case, indeed the confidence decreases from to . However, Figure 6 has a similar behavior although the confidence decreases only one percentage (from to ).
Thus, qualitatively speaking, unsuccessful one-pixel attacks not necessarily fail to achieve a high influence in the last layer. It depends strongly on the pixel position and sample. This is accordance with saliency maps which show that different parts of the image have different importance in the recognition process (Simonyan et al., 2013).
6 Statistical Evaluation of Propagation Maps
In previous sections, single attacks were analyzed in the light of examples and counterexamples. These experiments were important to investigate what happens in detail for each of these attacks. However, they do not tell much about the distribution of attacks. This section aims to fill this gap by investigating the attacks’ distribution and other statistically relevant data.
By evaluating the average of PMmean over successful attacks, we can observe the attack distribution over the entire feature map as well as their average spreadness over the layers (Figure 7). The successful attacks seem to concetrate mostly close to the center of the image. In deeper layers, the influence expands and increase in intensity, specially at its center. This reveals that the behavior observed in Figure 2 is usual for most of the attacks.
Given this distribution for successful attacks, it would be interesting to contrast them with failed attempts. This would enable us to further clarify the characteristics of a successful attack. However, as shown in Figure 8, most of the previous features are shared. The only observed difference is the initial spreadness and spreadness. In other words, failed attacks tend to be farther away from the center of the image and spread more equally than their successful counterparts. Notice that both Figure 7 and 8 are not scaled with the maximum feature map, since we are interested here in their distribution rather than intensity.
To further clarify if there is any explicit difference between successful and failed attacks, we explictly calculated the mean over all the feature maps for the previous successful and unsuccessful attacks. The plot is shown in Figure 9. The average difference is also unable to distinghish between successful and failed attacks, with both having very similar behavior.
7 Position Sensitivity and Locality
|Original One-pixel Attack||59%||33%|
|One-Pixel Attack on Random Pixels||4.9%||3.1%|
|One-pixel Attack on Nearby Pixels||33.1%||31.3%|
The one-pixel attack works by searching for a one-pixel perturbation where the class can be modified (misclassified). This search process is costly but to what extent is the success of the attack dependent on its position?
Here the position sentitivity of the attack will be analyzed. First, we consider an attack in which a pixel is randomly chosen to be perturbated by the same ammount that could create an adversarial sample. The results, which are shown in Table 1, demonstrate that random pixel attacks have a very low success rate. This suggests that position is important and by disregarding it is almost impossible to achieve a successful attack. Having said that, the attack on nearby pixels (i.e., the eight adjacent pixels) shows a positive result. In fact, if we consider that this attack is not conducting any search at all but only taking one random pixel. The results may be considered extremely positive.
The extremely positive results present in Table1, however, are in accordance with the receptive fields of convolutional layers. In other words, every neuron in convolutional layers calculates a convolution of the kernel with a part of the input image which is of the same size. The convolution itself is a linear function in which the change in one input would cause the whole convolution to be affected. Thus, the result of the convolution will be the same for nearby neurons in the same receptive field. Consequently, this shows that the vulnerable part of DNNs were neither neurons nor pixels but some receptive fields.
Interestingly, completely different networks have a very similar success rate for both nearby attacks. This further demonstrate that the receptive field is the vulnerable part. Since neural networks with similar architectures share similar receptive field relationships, nearby attacks on similar networks should have similar success rate.
8 The Conflicting Salience Hypothesis
Propagation maps demonstrated that some pixels’ influence failed to reach the last layer (Figure 4) while others influenced the last layer enough to cause a change in class (Figure 2). This analysis share a close resemblance to saliency maps in which one wishes to discover which pixels are responsible for a class. In fact, since propagation maps measure the ammount of influence from perturbations, it would be reasonable to assume that they may have a close relationship with disturbance in saliency maps. Consequently, adversarial samples would cause enough disturbance in saliency maps to cause a change in class. Thus, we raise here the hypothesis of a conflicting saliency from adversarial samples.
If this is true, then what adversarial machine learning is doing is not fooling DNNs but rather taking away his attention. It is like a magician that calls attention to his right hand while his left hand pushes the magic ball. Or like the blinking light on the street that calls attention of the driver which suddenly drive through a red traffic light.
Having said that, propagation maps is a feedfoward based technique to vizualize and measure the influence of a perturbation while saliency maps aim to investigate the salient pixels for a given class with backpropagated gradients. Therefore, the methods differ in many ways and their relationship, which might be more complex than what is stated here, goes beyond the scope of this paper. We leave this as an open question that should be worthwhile to investigate.
This paper proposed a novel technique called propagation maps and used it to analyze one of the most puzzling attacks, the one-pixel attack. The analysis showed how a pixel modification causes an influence throughout the layers, culminating in the change of the class. Moreover, a locality analysis revealed that receptive fields are the vulnerable parts of DNNs and therefore nearby attacks to successful one-pixel attack have a high success rate. Lastly, a new hypothesis was proposed that could together with the proposed propagation maps help explain the reason behind adversarial attacks in DNNs.
Regarding this paper achievements, we highlight the following:
Propagation Maps - A novel technique to estimate the influence a perturbation can exert over a layer was proposed. Propagation maps make it easy to understand both the ammount as well as the spreadness of influence given a perturbation in each layer of a DNN.
Vulnerability of Receptive Fields - A locality analysis revealed that success rates on nearby pixels of successful one-pixel attacks are equally vulnerable. This demonstrate that pixels or neurons are not the vulnerable parts but receptive fields.
The Influence of One-pixel Perturbation - Propagation maps reveal for the first time how one-pixel perturbation can influence layers of DNNs. Tests conducted on Resnet show that the perturbation inside neural networks grow and spread. In fact, the difference in feature maps was shown to reach values comparable with the maximum output of the original feature map.
Similarity between Successful and Failed Attacks - In this paper, many tests over successful and failed attacks were made. However, both types of perturbation shows surprisingly similar behavior. This was shown to not be necessarily related with a decrease in the confidence with the class label either, i.e., failed attacks that do not change the confidence may also have a high influence in all layers of the DNN and behave similarly to successful attacks.
The Conflicting Salience Hypothesis - Inspired by how propagation maps show the influence of pixels throughout the layers, we raised the hypothesis that adversarial samples are disturbances in saliency maps. Saliency maps share a close resemblance with propagation maps. Some propagation maps’ results also point to conclusions that are in accordance with saliency maps. Both methods, however, come from quasi contrary perspectives. For example, saliency maps is a backpropagated signal while propagation maps is a feedfowarded one. To further verify the hypothesis more tests in their relationship are necessary.
Thus, by shedding light into the influence of perturbations inside any type of DNN, we expect that propagation maps shall aid the understanding of other attacks as well as the development of new defenses. They can also be extended to work together with saliency maps providing yet more clues to uncover the reasoning behind adversarial samples as well as how deep neural networks understand the world.
This work was supported by JST, ACT-I Grant Number JP-50166, Japan.
- Athalye & Sutskever (2018) Athalye, A. and Sutskever, I. Synthesizing robust adversarial examples. In ICML, 2018.
- Athalye et al. (2018) Athalye, A., Carlini, N., and Wagner, D. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. 2018.
- Barreno et al. (2006) Barreno, M., Nelson, B., Sears, R., Joseph, A. D., and Tygar, J. D. Can machine learning be secure? In Proceedings of the 2006 ACM Symposium on Information, computer and communications security, pp. 16–25. ACM, 2006.
- Barreno et al. (2010) Barreno, M., Nelson, B., Joseph, A. D., and Tygar, J. D. The security of machine learning. Machine Learning, 81(2):121–148, 2010.
- Brown et al. (2017) Brown, T. B., Mané, D., Roy, A., Abadi, M., and Gilmer, J. Adversarial patch. arXiv preprint arXiv:1712.09665, 2017.
- (6) Buckman, J., Roy, A., Raffel, C., and Goodfellow, I. Thermometer encoding: One hot way to resist adversarial examples.
Carlini & Wagner (2017a)
Carlini, N. and Wagner, D.
Adversarial examples are not easily detected: Bypassing ten detection
Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pp. 3–14. ACM, 2017a.
- Carlini & Wagner (2017b) Carlini, N. and Wagner, D. Magnet and” efficient defenses against adversarial attacks” are not robust to adversarial examples. arXiv preprint arXiv:1711.08478, 2017b.
- Carlini & Wagner (2017c) Carlini, N. and Wagner, D. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pp. 39–57. IEEE, 2017c.
- Dang et al. (2017) Dang, H., Huang, Y., and Chang, E.-C. Evading classifiers by morphing in the dark. 2017.
- (11) Dhillon, G. S., Azizzadenesheli, K., Lipton, Z. C., Bernstein, J., Kossaifi, J., Khanna, A., and Anandkumar, A. Stochastic activation pruning for robust adversarial defense.
- Goodfellow et al. (2014a) Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014a.
- Goodfellow et al. (2014b) Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014b.
- Grosse et al. (2017) Grosse, K., Manoharan, P., Papernot, N., Backes, M., and McDaniel, P. On the (statistical) detection of adversarial examples. arXiv preprint arXiv:1702.06280, 2017.
- (15) Guo, C., Rana, M., Cisse, M., and van der Maaten, L. Countering adversarial images using input transformations. In ICLR.
- Huang et al. (2015) Huang, R., Xu, B., Schuurmans, D., and Szepesvári, C. Learning with a strong adversary. arXiv preprint arXiv:1511.03034, 2015.
- Kurakin et al. (2016) Kurakin, A., Goodfellow, I., and Bengio, S. Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533, 2016.
- Ma et al. (2018) Ma, X., Li, B., Wang, Y., Erfani, S. M., Wijewickrema, S., Schoenebeck, G., Song, D., Houle, M. E., and Bailey, J. Characterizing adversarial subspaces using local intrinsic dimensionality. arXiv preprint arXiv:1801.02613, 2018.
Madry et al. (2018)
Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A.
Towards deep learning models resistant to adversarial attacks.In ICLR, 2018.
Moosavi-Dezfooli et al. (2016)
Moosavi-Dezfooli, S.-M., Fawzi, A., and Frossard, P.
Deepfool: a simple and accurate method to fool deep neural networks.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2574–2582, 2016.
- Moosavi-Dezfooli et al. (2017) Moosavi-Dezfooli, S.-M., Fawzi, A., Fawzi, O., and Frossard, P. Universal adversarial perturbations. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 86–94. IEEE, 2017.
- Narodytska & Kasiviswanathan (2017) Narodytska, N. and Kasiviswanathan, S. Simple black-box adversarial attacks on deep neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 1310–1318. IEEE, 2017.
- Nguyen et al. (2015) Nguyen, A., Yosinski, J., and Clune, J. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 427–436, 2015.
- Papernot et al. (2016a) Papernot, N., McDaniel, P., Jha, S., Fredrikson, M., Celik, Z. B., and Swami, A. The limitations of deep learning in adversarial settings. In Security and Privacy (EuroS&P), 2016 IEEE European Symposium on, pp. 372–387. IEEE, 2016a.
- Papernot et al. (2016b) Papernot, N., McDaniel, P., Wu, X., Jha, S., and Swami, A. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE Symposium on Security and Privacy (SP), pp. 582–597. IEEE, 2016b.
- Papernot et al. (2017) Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z. B., and Swami, A. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia Conference on Computer and Communications Security, pp. 506–519. ACM, 2017.
- (27) Samangouei, P., Kabkab, M., and Chellappa, R. Defense-gan: Protecting classifiers against adversarial attacks using generative models.
Sharif et al. (2016)
Sharif, M., Bhagavatula, S., Bauer, L., and Reiter, M. K.
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, pp. 1528–1540. ACM, 2016.
- Simonyan et al. (2013) Simonyan, K., Vedaldi, A., and Zisserman, A. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013.
- (30) Song, Y., Kim, T., Nowozin, S., Ermon, S., and Kushman, N. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples.
Storn & Price (1997)
Storn, R. and Price, K.
Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces.Journal of global optimization, 11(4):341–359, 1997.
- Su et al. (2017) Su, J., Vargas, D. V., and Kouichi, S. One pixel attack for fooling deep neural networks. arXiv preprint arXiv:1710.08864, 2017.
- Szegedy (2014) Szegedy, C. e. a. Intriguing properties of neural networks. In In ICLR. Citeseer, 2014.
- Tramèr et al. (2018) Tramèr, F., Kurakin, A., Papernot, N., Goodfellow, I., Boneh, D., and McDaniel, P. Ensemble adversarial training: Attacks and defenses. In ICLR, 2018.
- Vargas & Murata (2017) Vargas, D. V. and Murata, J. Spectrum-diverse neuroevolution with unified neural models. IEEE transactions on neural networks and learning systems, 28(8):1759–1773, 2017.
- Vargas et al. (2013) Vargas, D. V., Takano, H., and Murata, J. Self organizing classifiers: first steps in structured evolutionary machine learning. Evolutionary Intelligence, 6(2):57–72, 2013.
- (37) Xie, C., Wang, J., Zhang, Z., Ren, Z., and Yuille, A. Mitigating adversarial effects through randomization.
- Xu et al. (2017) Xu, W., Evans, D., and Qi, Y. Feature squeezing: Detecting adversarial examples in deep neural networks. arXiv preprint arXiv:1704.01155, 2017.