This paper investigates recently proposed approaches for defending against adversarial examples and evaluating adversarial robustness. We motivate adversarial risk as an objective for achieving models robust to worst-case inputs. We then frame commonly used attacks and evaluation metrics as defining a tractable surrogate objective to the true adversarial risk. This suggests that models may optimize this surrogate rather than the true adversarial risk. We formalize this notion as obscurity to an adversary, and develop tools and heuristics for identifying obscured models and designing transparent models. We demonstrate that this is a significant problem in practice by repurposing gradient-free optimization techniques into adversarial attacks, which we use to decrease the accuracy of several recently proposed defenses to near zero. Our hope is that our formulations and results will help researchers to develop more powerful defenses.

Adaptive attacks have (rightfully) become the de facto standard for evaluating defenses to adversarial examples. We find, however, that typical adaptive evaluations are incomplete. We demonstrate that thirteen defenses recently published at ICLR, ICML and NeurIPS-and which illustrate a diverse set of defense strategies-can be circumvented despite attempting to perform evaluations using adaptive attacks. While prior evaluation papers focused mainly on the end result-showing that a defense was ineffective-this paper focuses on laying out the methodology and the approach necessary to perform an adaptive attack. Some of our attack strategies are generalizable, but no single strategy would have been sufficient for all defenses. This underlines our key message that adaptive attacks cannot be automated and always require careful and appropriate tuning to a given defense. We hope that these analyses will serve as guidance on how to properly perform adaptive attacks against defenses to adversarial examples, and thus will allow the community to make further progress in building more robust models.

We identify obfuscated gradients, a kind of gradient masking, as a phenomenon that leads to a false sense of security in defenses against adversarial examples. While defenses that cause obfuscated gradients appear to defeat iterative optimizationbased attacks, we find defenses relying on this effect can be circumvented. We describe characteristic behaviors of defenses exhibiting the effect, and for each of the three types of obfuscated gradients we discover, we develop attack techniques to overcome it. In a case study, examining noncertified white-box-secure defenses at ICLR 2018, we find obfuscated gradients are a common occurrence, with 7 of 9 defenses relying on obfuscated gradients. Our new attacks successfully circumvent 6 completely, and 1 partially, in the original threat model each paper considers.

Recent work has demonstrated that deep neural networks are vulnerable to adversarial examples-inputs that are almost indistinguishable from natural data and yet classified incorrectly by the network. In fact, some of the latest findings suggest that the existence of adversarial attacks may be an inherent weakness of deep learning models. To address this problem, we study the adversarial robustness of neural networks through the lens of robust optimization. This approach provides us with a broad and unifying view on much of the prior work on this topic. Its principled nature also enables us to identify methods for both training and attacking neural networks that are reliable and, in a certain sense, universal. In particular, they specify a concrete security guarantee that would protect against any adversary. These methods let us train networks with significantly improved resistance to a wide range of adversarial attacks. They also suggest the notion of security against a first-order adversary as a natural and broad security guarantee. We believe that robustness against such well-defined classes of adversaries is an important stepping stone towards fully resistant deep learning models. 1

Adversarial examples are perturbed inputs designed to fool machine learning models. Adversarial training injects such examples into training data to increase robustness. To scale this technique to large datasets, perturbations are crafted using fast single-step methods that maximize a linear approximation of the model's loss. We show that this form of adversarial training converges to a degenerate global minimum, wherein small curvature artifacts near the data points obfuscate a linear approximation of the loss. The model thus learns to generate weak perturbations, rather than defend against strong ones. As a result, we find that adversarial training remains vulnerable to black-box attacks, where we transfer perturbations computed on undefended models, as well as to a powerful novel single-step attack that escapes the non-smooth vicinity of the input data via a small random step. We further introduce Ensemble Adversarial Training, a technique that augments training data with perturbations transferred from other models. On ImageNet, Ensemble Adversarial Training yields models with stronger robustness to blackbox attacks. In particular, our most robust model won the first round of the NIPS 2017 competition on Defenses against Adversarial Attacks (Kurakin et al., 2017c). However, subsequent work found that more elaborate black-box attacks could significantly enhance transferability and reduce the accuracy of our models.

The field of defense strategies against adversarial attacks has significantly grown over the last years, but progress is hampered as the evaluation of adversarial defenses is often insufficient and thus gives a wrong impression of robustness. Many promising defenses could be broken later on, making it difficult to identify the state-of-the-art. Frequent pitfalls in the evaluation are improper tuning of hyperparameters of the attacks, gradient obfuscation or masking. In this paper we first propose two extensions of the PGD-attack overcoming failures due to suboptimal step size and problems of the objective function. We then combine our novel attacks with two complementary existing ones to form a parameter-free, computationally affordable and user-independent ensemble of attacks to test adversarial robustness. We apply our ensemble to over 50 models from papers published at recent top machine learning and computer vision venues. In all except one of the cases we achieve lower robust test accuracy than reported in these papers, often by more than 10%, identifying several broken defenses.

在测试时间进行优化的自适应防御能力有望改善对抗性鲁棒性。我们对这种自适应测试时间防御措施进行分类，解释其潜在的好处和缺点，并评估图像分类的最新自适应防御能力的代表性。不幸的是，经过我们仔细的案例研究评估时，没有任何显着改善静态防御。有些甚至削弱了基本静态模型，同时增加了推理计算。尽管这些结果令人失望，但我们仍然认为自适应测试时间防御措施是一项有希望的研究途径，因此，我们为他们的彻底评估提供了建议。我们扩展了Carlini等人的清单。（2019年）通过提供针对自适应防御的具体步骤。

Neural networks provide state-of-the-art results for most machine learning tasks. Unfortunately, neural networks are vulnerable to adversarial examples: given an input x and any target classification t, it is possible to find a new input x that is similar to x but classified as t. This makes it difficult to apply neural networks in security-critical areas. Defensive distillation is a recently proposed approach that can take an arbitrary neural network, and increase its robustness, reducing the success rate of current attacks' ability to find adversarial examples from 95% to 0.5%.In this paper, we demonstrate that defensive distillation does not significantly increase the robustness of neural networks by introducing three new attack algorithms that are successful on both distilled and undistilled neural networks with 100% probability. Our attacks are tailored to three distance metrics used previously in the literature, and when compared to previous adversarial example generation algorithms, our attacks are often much more effective (and never worse). Furthermore, we propose using high-confidence adversarial examples in a simple transferability test we show can also be used to break defensive distillation. We hope our attacks will be used as a benchmark in future defense attempts to create neural networks that resist adversarial examples.

在本讨论文件中，我们调查了有关机器学习模型鲁棒性的最新研究。随着学习算法在数据驱动的控制系统中越来越流行，必须确保它们对数据不确定性的稳健性，以维持可靠的安全至关重要的操作。我们首先回顾了这种鲁棒性的共同形式主义，然后继续讨论训练健壮的机器学习模型的流行和最新技术，以及可证明这种鲁棒性的方法。从强大的机器学习的这种统一中，我们识别并讨论了该地区未来研究的迫切方向。

The authors thank Nicholas Carlini (UC Berkeley) and Dimitris Tsipras (MIT) for feedback to improve the survey quality. We also acknowledge X. Huang (Uni. Liverpool), K. R. Reddy (IISC), E. Valle (UNICAMP), Y. Yoo (CLAIR) and others for providing pointers to make the survey more comprehensive.

虽然深度神经网络在各种任务中表现出前所未有的性能，但对对抗性示例的脆弱性阻碍了他们在安全关键系统中的部署。许多研究表明，即使在黑盒设置中也可能攻击，其中攻击者无法访问目标模型的内部信息。大多数黑匣子攻击基于查询，每个都可以获得目标模型的输入输出，并且许多研究侧重于减少所需查询的数量。在本文中，我们注意了目标模型的输出完全对应于查询输入的隐含假设。如果将某些随机性引入模型中，它可以打破假设，因此，基于查询的攻击可能在梯度估计和本地搜索中具有巨大的困难，这是其攻击过程的核心。从这种动机来看，我们甚至观察到一个小的添加剂输入噪声可以中和大多数基于查询的攻击和名称这个简单但有效的方法小噪声防御（SND）。我们分析了SND如何防御基于查询的黑匣子攻击，并展示其与CIFAR-10和ImageNet数据集的八种最先进的攻击有效性。即使具有强大的防御能力，SND几乎保持了原始的分类准确性和计算速度。通过在推断下仅添加一行代码，SND很容易适用于预先训练的模型。

Neural networks are known to be vulnerable to adversarial examples: inputs that are close to natural inputs but classified incorrectly. In order to better understand the space of adversarial examples, we survey ten recent proposals that are designed for detection and compare their efficacy. We show that all can be defeated by constructing new loss functions. We conclude that adversarial examples are significantly harder to detect than previously appreciated, and the properties believed to be intrinsic to adversarial examples are in fact not. Finally, we propose several simple guidelines for evaluating future proposed defenses.

Although deep neural networks (DNNs) have achieved great success in many tasks, they can often be fooled by adversarial examples that are generated by adding small but purposeful distortions to natural examples. Previous studies to defend against adversarial examples mostly focused on refining the DNN models, but have either shown limited success or required expensive computation. We propose a new strategy, feature squeezing, that can be used to harden DNN models by detecting adversarial examples. Feature squeezing reduces the search space available to an adversary by coalescing samples that correspond to many different feature vectors in the original space into a single sample. By comparing a DNN model's prediction on the original input with that on squeezed inputs, feature squeezing detects adversarial examples with high accuracy and few false positives.This paper explores two feature squeezing methods: reducing the color bit depth of each pixel and spatial smoothing. These simple strategies are inexpensive and complementary to other defenses, and can be combined in a joint detection framework to achieve high detection rates against state-of-the-art attacks.

深度神经网络已成为现代图像识别系统的驱动力。然而，神经网络对抗对抗性攻击的脆弱性对受这些系统影响的人构成严重威胁。在本文中，我们专注于一个真实的威胁模型，中间对手恶意拦截和erturbs网页用户上传在线。这种类型的攻击可以在简单的性能下降之上提高严重的道德问题。为了防止这种攻击，我们设计了一种新的双层优化算法，该算法在对抗对抗扰动的自然图像附近找到点。CiFar-10和Imagenet的实验表明我们的方法可以有效地强制在给定的修改预算范围内的自然图像。我们还显示所提出的方法可以在共同使用随机平滑时提高鲁棒性。

Designing powerful adversarial attacks is of paramount importance for the evaluation of $\ell_p$-bounded adversarial defenses. Projected Gradient Descent (PGD) is one of the most effective and conceptually simple algorithms to generate such adversaries. The search space of PGD is dictated by the steepest ascent directions of an objective. Despite the plethora of objective function choices, there is no universally superior option and robustness overestimation may arise from ill-suited objective selection. Driven by this observation, we postulate that the combination of different objectives through a simple loss alternating scheme renders PGD more robust towards design choices. We experimentally verify this assertion on a synthetic-data example and by evaluating our proposed method across 25 different $\ell_{\infty}$-robust models and 3 datasets. The performance improvement is consistent, when compared to the single loss counterparts. In the CIFAR-10 dataset, our strongest adversarial attack outperforms all of the white-box components of AutoAttack (AA) ensemble, as well as the most powerful attacks existing on the literature, achieving state-of-the-art results in the computational budget of our study ($T=100$, no restarts).

作为研究界，我们仍然缺乏对对抗性稳健性的进展的系统理解，这通常使得难以识别训练强大模型中最有前途的想法。基准稳健性的关键挑战是，其评估往往是出错的导致鲁棒性高估。我们的目标是建立对抗性稳健性的标准化基准，尽可能准确地反映出考虑在合理的计算预算范围内所考虑的模型的稳健性。为此，我们首先考虑图像分类任务并在允许的型号上引入限制（可能在将来宽松）。我们评估了与AutoAtrack的对抗鲁棒性，白和黑箱攻击的集合，最近在大规模研究中显示，与原始出版物相比，改善了几乎所有稳健性评估。为防止对自动攻击进行新防御的过度适应，我们欢迎基于自适应攻击的外部评估，特别是在自动攻击稳健性潜在高估的地方。我们的排行榜，托管在https://robustbench.github.io/，包含120多个模型的评估，并旨在反映在$ \ ell_ \ infty $的一套明确的任务上的图像分类中的当前状态 - 和$ \ ell_2 $ -Threat模型和共同腐败，未来可能的扩展。此外，我们开源源是图书馆https://github.com/robustbench/robustbench，可以提供对80多个强大模型的统一访问，以方便他们的下游应用程序。最后，根据收集的模型，我们分析了稳健性对分布换档，校准，分配检测，公平性，隐私泄漏，平滑度和可转移性的影响。

对抗性的鲁棒性已经成为深度学习的核心目标，无论是在理论和实践中。然而，成功的方法来改善对抗的鲁棒性（如逆势训练）在不受干扰的数据上大大伤害了泛化性能。这可能会对对抗性鲁棒性如何影响现实世界系统的影响（即，如果它可以提高未受干扰的数据的准确性），许多人可能选择放弃鲁棒性）。我们提出内插对抗培训，该培训最近雇用了在对抗培训框架内基于插值的基于插值的培训方法。在CiFar -10上，对抗性训练增加了标准测试错误（当没有对手时）从4.43％到12.32％，而我们的内插对抗培训我们保留了对抗性的鲁棒性，同时实现了仅6.45％的标准测试误差。通过我们的技术，强大模型标准误差的相对增加从178.1％降至仅为45.5％。此外，我们提供内插对抗性培训的数学分析，以确认其效率，并在鲁棒性和泛化方面展示其优势。

神经网络对攻击的缺乏鲁棒性引起了对安全敏感环境（例如自动驾驶汽车）的担忧。虽然许多对策看起来可能很有希望，但只有少数能够承受严格的评估。使用随机变换（RT）的防御能力显示出令人印象深刻的结果，尤其是Imagenet上的Bart（Raff等，2019）。但是，这种防御尚未经过严格评估，使其稳健性的理解不足。它们的随机特性使评估更具挑战性，并使对确定性模型的许多拟议攻击不可应用。首先，我们表明BART评估中使用的BPDA攻击（Athalye等，2018a）无效，可能高估了其稳健性。然后，我们尝试通过明智的转换和贝叶斯优化来调整其参数来构建最强的RT防御。此外，我们创造了最强烈的攻击来评估我们的RT防御。我们的新攻击极大地胜过基线，与常用的EOT攻击减少19％相比，将准确性降低了83％（$ 4.3 \ times $改善）。我们的结果表明，在Imagenette数据集上的RT防御（ImageNet的十级子集）在对抗性示例上并不强大。进一步扩展研究，我们使用新的攻击来对抗RT防御（称为Advrt），从而获得了巨大的稳健性增长。代码可从https://github.com/wagner-group/demystify-random-transform获得。

对对抗性攻击的鲁棒性通常以对抗精度评估。但是，该指标太粗糙，无法正确捕获机器学习模型的所有鲁棒性。当对强烈的攻击进行评估时，许多防御能力并不能提供准确的改进，同时仍会部分贡献对抗性鲁棒性。流行的认证方法遇到了同一问题，因为它们提供了准确性的下限。为了捕获更精细的鲁棒性属性，我们提出了一个针对L2鲁棒性，对抗角稀疏性的新指标，该指标部分回答了“输入周围有多少个对抗性示例”的问题。我们通过评估“强”和“弱”的防御能力来证明其有用性。我们表明，一些最先进的防御能力具有非常相似的精度，在它们不强大的输入上可能具有截然不同的稀疏性。我们还表明，一些弱防御能力实际上会降低鲁棒性，而另一些防御能力则以无法捕获的准确性来加强它。这些差异可以预测这种防御与对抗性训练相结合时的实用性。

评估机器学习模型对对抗性示例的鲁棒性是一个具有挑战性的问题。已经证明，许多防御能力通过导致基于梯度的攻击失败，从而提供了一种错误的鲁棒感，并且在更严格的评估下它们已被打破。尽管已经提出了指南和最佳实践来改善当前的对抗性鲁棒性评估，但缺乏自动测试和调试工具，使以系统的方式应用这些建议变得困难。在这项工作中，我们通过以下方式克服了这些局限性：（i）根据它们如何影响基于梯度的攻击的优化对攻击失败进行分类，同时还揭示了两种影响许多流行攻击实施和过去评估的新型故障； （ii）提出了六个新的失败指标，以自动检测到攻击优化过程中这种失败的存在； （iii）建议采用系统协议来应用相应的修复程序。我们广泛的实验分析涉及3个不同的应用域中的15多个模型，表明我们的失败指标可用于调试和改善当前的对抗性鲁棒性评估，从而为自动化和系统化它们提供了第一步。我们的开源代码可在以下网址获得：https：//github.com/pralab/indicatorsofattackfailure。