Transfer-based adversarial example is one of the most important classes of black-box attacks. However, there is a trade-off between transferability and imperceptibility of the adversarial perturbation. Prior work in this direction often requires a fixed but large $\ell_p$-norm perturbation budget to reach a good transfer success rate, leading to perceptible adversarial perturbations. On the other hand, most of the current unrestricted adversarial attacks that aim to generate semantic-preserving perturbations suffer from weaker transferability to the target model. In this work, we propose a geometry-aware framework to generate transferable adversarial examples with minimum changes. Analogous to model selection in statistical machine learning, we leverage a validation model to select the optimal perturbation budget for each image under both the $\ell_{\infty}$-norm and unrestricted threat models. Extensive experiments verify the effectiveness of our framework on balancing imperceptibility and transferability of the crafted adversarial examples. The methodology is the foundation of our entry to the CVPR'21 Security AI Challenger: Unrestricted Adversarial Attacks on ImageNet, in which we ranked 1st place out of 1,559 teams and surpassed the runner-up submissions by 4.59% and 23.91% in terms of final score and average image quality level, respectively. Code is available at https://github.com/Equationliu/GA-Attack.
Many works have investigated the adversarial attacks or defenses under the settings where a bounded and imperceptible perturbation can be added to the input. However in the real-world, the attacker does not need to comply with this restriction. In fact, more threats to the deep model come from unrestricted adversarial examples, that is, the attacker makes large and visible modifications on the image, which causes the model classifying mistakenly, but does not affect the normal observation in human perspective. Unrestricted adversarial attack is a popular and practical direction but has not been studied thoroughly. We organize this competition with the purpose of exploring more effective unrestricted adversarial attack algorithm, so as to accelerate the academical research on the model robustness under stronger unbounded attacks. The competition is held on the TianChi platform (\url{https://tianchi.aliyun.com/competition/entrance/531853/introduction}) as one of the series of AI Security Challengers Program.
We propose self-adaptive training -- a unified training algorithm that dynamically calibrates and enhances training process by model predictions without incurring extra computational cost -- to advance both supervised and self-supervised learning of deep neural networks. We analyze the training dynamics of deep networks on training data that are corrupted by, e.g., random noise and adversarial examples. Our analysis shows that model predictions are able to magnify useful underlying information in data and this phenomenon occurs broadly even in the absence of \emph{any} label information, highlighting that model predictions could substantially benefit the training process: self-adaptive training improves the generalization of deep networks under noise and enhances the self-supervised representation learning. The analysis also sheds light on understanding deep learning, e.g., a potential explanation of the recently-discovered double-descent phenomenon in empirical risk minimization and the collapsing issue of the state-of-the-art self-supervised learning algorithms. Experiments on the CIFAR, STL and ImageNet datasets verify the effectiveness of our approach in three applications: classification with label noise, selective classification and linear evaluation. To facilitate future research, the code has been made public available at https://github.com/LayneH/self-adaptive-training.
We prove that classifiers with the ability to abstain are provably more powerful than those that cannot against an adversary that can perturb datapoints by arbitrary amounts in random directions. Specifically, we show that no matter how well-behaved the natural data is, any classifier that cannot abstain will be defeated by such an adversary. However, by allowing abstention, we give a parameterized algorithm with provably good performance against such an adversary when classes are reasonably well-separated and the data dimension is high. We further use a data-driven method to set our algorithm parameters to optimize over the accuracy vs. abstention trade-off with strong theoretical guarantees. Our theory has direct applications to the technique of contrastive learning, where we empirically demonstrate the ability of our algorithms to obtain high robust accuracy with only small amounts of abstention in both supervised and self-supervised settings.
In this paper we introduce a provably stable architecture for Neural Ordinary Differential Equations (ODEs) which achieves non-trivial adversarial robustness under white-box adversarial attacks even when the network is trained naturally. For most existing defense methods withstanding strong white-box attacks, to improve robustness of neural networks, they need to be trained adversarially, hence have to strike a trade-off between natural accuracy and adversarial robustness. Inspired by dynamical system theory, we design a stabilized neural ODE network named SONet whose ODE blocks are skew-symmetric and proved to be input-output stable. With natural training, SONet can achieve comparable robustness with the state-of-the-art adversarial defense methods, without sacrificing natural accuracy. Even replacing only the first layer of a ResNet by such a ODE block can exhibit further improvement in robustness, e.g., under PGD-20 ($\ell_\infty=0.031$) attack on CIFAR-10 dataset, it achieves 91.57\% and natural accuracy and 62.35\% robust accuracy, while a counterpart architecture of ResNet trained with TRADES achieves natural and robust accuracy 76.29\% and 45.24\%, respectively. To understand possible reasons behind this surprisingly good result, we further explore the possible mechanism underlying such an adversarial robustness. We show that the adaptive stepsize numerical ODE solver, DOPRI5, has a gradient masking effect that fails the PGD attacks which are sensitive to gradient information of training loss; on the other hand, it cannot fool the CW attack of robust gradients and the SPSA attack that is gradient-free. This provides a new explanation that the adversarial robustness of ODE-based networks mainly comes from the obfuscated gradients in numerical ODE solvers.
A standard method for improving the robustness of neural networks is adversarial training, where the network is trained on adversarial examples that are close to the training inputs. This produces classifiers that are robust, but it often decreases clean accuracy. Prior work even posits that the tradeoff between robustness and accuracy may be inevitable. We investigate this tradeoff in more depth through the lens of local Lipschitzness. In many image datasets, the classes are separated in the sense that images with different labels are not extremely close in $\ell_\infty$ distance. Using this separation as a starting point, we argue that it is possible to achieve both accuracy and robustness by encouraging the classifier to be locally smooth around the data. More precisely, we consider classifiers that are obtained by rounding locally Lipschitz functions. Theoretically, we show that such classifiers exist for any dataset such that there is a positive distance between the support of different classes. Empirically, we compare the local Lipschitzness of classifiers trained by several methods. Our results show that having a small Lipschitz constant correlates with achieving high clean and robust accuracy, and therefore, the smoothness of the classifier is an important property to consider in the context of adversarial examples. Code available at https://github.com/yangarbiter/robust-local-lipschitz .
We show a hardness result for random smoothing to achieve certified adversarial robustness against attacks in the $\ell_p$ ball of radius $\epsilon$ when $p>2$. Although random smoothing has been well understood for the $\ell_2$ case using the Gaussian distribution, much remains unknown concerning the existence of a noise distribution that works for the case of $p>2$. This has been posed as an open problem by Cohen et al. (2019) and includes many significant paradigms such as the $\ell_\infty$ threat model. In this work, we show that any noise distribution $\mathcal{D}$ over $\mathbb{R}^d$ that provides $\ell_p$ robustness for all base classifiers with $p>2$ must satisfy $\mathbb{E}\eta_i^2=\Omega(d^{1-2/p}\epsilon^2(1-\delta)/\delta^2)$ for 99% of the features (pixels) of vector $\eta\sim\mathcal{D}$, where $\epsilon$ is the robust radius and $\delta$ is the score gap between the highest-scored class and the runner-up. Therefore, for high-dimensional images with pixel values bounded in $[0,255]$, the required noise will eventually dominate the useful information in the images, leading to trivial smoothed classifiers.
We propose self-adaptive training---a new training algorithm that dynamically corrects problematic training labels by model predictions without incurring extra computational cost---to improve generalization of deep learning for potentially corrupted training data. This problem is crucial towards robustly learning from data that are corrupted by, e.g., label noises and out-of-distribution samples. The standard empirical risk minimization (ERM) for such data, however, may easily overfit noises and thus suffers from sub-optimal performance. In this paper, we observe that model predictions can substantially benefit the training process: self-adaptive training significantly improves generalization over ERM under various levels of noises, and mitigates the overfitting issue in both natural and adversarial training. We evaluate the error-capacity curve of self-adaptive training: the test error is monotonously decreasing w.r.t. model capacity. This is in sharp contrast to the recently-discovered double-descent phenomenon in ERM which might be a result of overfitting of noises. Experiments on CIFAR and ImageNet datasets verify the effectiveness of our approach in two applications: classification with label noise and selective classification. We release our code at \url{https://github.com/LayneH/self-adaptive-training}.