Following the success in advancing natural language processing and understanding, transformers are expected to bring revolutionary changes to computer vision. This work provides the first and comprehensive study on the robustness of vision transformers (ViTs) against adversarial perturbations. Tested on various white-box and transfer attack settings, we find that ViTs possess better adversarial robustness when compared with convolutional neural networks (CNNs). This observation also holds for certified robustness. We summarize the following main observations contributing to the improved robustness of ViTs: 1) Features learned by ViTs contain less low-level information and are more generalizable, which contributes to superior robustness against adversarial perturbations. 2) Introducing convolutional or tokens-to-token blocks for learning low-level features in ViTs can improve classification accuracy but at the cost of adversarial robustness. 3) Increasing the proportion of transformers in the model structure (when the model consists of both transformer and CNN blocks) leads to better robustness. But for a pure transformer model, simply increasing the size or adding layers cannot guarantee a similar effect. 4) Pre-training on larger datasets does not significantly improve adversarial robustness though it is critical for training ViTs. 5) Adversarial training is also applicable to ViT for training robust models. Furthermore, feature visualization and frequency analysis are conducted for explanation. The results show that ViTs are less sensitive to high-frequency perturbations than CNNs and there is a high correlation between how well the model learns low-level features and its robustness against different frequency-based perturbations.
相關內容
Deep learning models for graphs have achieved strong performance for the task of node classification. Despite their proliferation, currently there is no study of their robustness to adversarial attacks. Yet, in domains where they are likely to be used, e.g. the web, adversaries are common. Can deep learning models for graphs be easily fooled? In this work, we introduce the first study of adversarial attacks on attributed graphs, specifically focusing on models exploiting ideas of graph convolutions. In addition to attacks at test time, we tackle the more challenging class of poisoning/causative attacks, which focus on the training phase of a machine learning model. We generate adversarial perturbations targeting the node's features and the graph structure, thus, taking the dependencies between instances in account. Moreover, we ensure that the perturbations remain unnoticeable by preserving important data characteristics. To cope with the underlying discrete domain we propose an efficient algorithm Nettack exploiting incremental computations. Our experimental study shows that accuracy of node classification significantly drops even when performing only few perturbations. Even more, our attacks are transferable: the learned attacks generalize to other state-of-the-art node classification models and unsupervised approaches, and likewise are successful even when only limited knowledge about the graph is given.
Recent self-supervision methods have found success in learning feature representations that could rival ones from full supervision, and have been shown to be beneficial to the model in several ways: for example improving models robustness and out-of-distribution detection. In our paper, we conduct an empirical study to understand more precisely in what way can self-supervised learning - as a pre-training technique or part of adversarial training - affects model robustness to $l_2$ and $l_{\infty}$ adversarial perturbations and natural image corruptions. Self-supervision can indeed improve model robustness, however it turns out the devil is in the details. If one simply adds self-supervision loss in tandem with adversarial training, then one sees improvement in accuracy of the model when evaluated with adversarial perturbations smaller or comparable to the value of $\epsilon_{train}$ that the robust model is trained with. However, if one observes the accuracy for $\epsilon_{test} \ge \epsilon_{train}$, the model accuracy drops. In fact, the larger the weight of the supervision loss, the larger the drop in performance, i.e. harming the robustness of the model. We identify primary ways in which self-supervision can be added to adversarial training, and observe that using a self-supervised loss to optimize both network parameters and find adversarial examples leads to the strongest improvement in model robustness, as this can be viewed as a form of ensemble adversarial training. Although self-supervised pre-training yields benefits in improving adversarial training as compared to random weight initialization, we observe no benefit in model robustness or accuracy if self-supervision is incorporated into adversarial training.
Transformers, composed of multiple self-attention layers, hold strong promises toward a generic learning primitive applicable to different data modalities, including the recent breakthroughs in computer vision achieving state-of-the-art (SOTA) standard accuracy. What remains largely unexplored is their robustness evaluation and attribution. In this work, we study the robustness of the Vision Transformer (ViT) against common corruptions and perturbations, distribution shifts, and natural adversarial examples. We use six different diverse ImageNet datasets concerning robust classification to conduct a comprehensive performance comparison of ViT models and SOTA convolutional neural networks (CNNs), Big-Transfer. Through a series of six systematically designed experiments, we then present analyses that provide both quantitative and qualitative indications to explain why ViTs are indeed more robust learners. For example, with fewer parameters and similar dataset and pre-training combinations, ViT gives a top-1 accuracy of 28.10% on ImageNet-A which is 4.3x higher than a comparable variant of BiT. Our analyses on image masking, Fourier spectrum sensitivity, and spread on discrete cosine energy spectrum reveal intriguing properties of ViT attributing to improved robustness. Code for reproducing our experiments is available at //git.io/J3VO0.
Transformers, which are popular for language modeling, have been explored for solving vision tasks recently, \eg, the Vision Transformer (ViT) for image classification. The ViT model splits each image into a sequence of tokens with fixed length and then applies multiple Transformer layers to model their global relation for classification. However, ViT achieves inferior performance to CNNs when trained from scratch on a midsize dataset like ImageNet. We find it is because: 1) the simple tokenization of input images fails to model the important local structure such as edges and lines among neighboring pixels, leading to low training sample efficiency; 2) the redundant attention backbone design of ViT leads to limited feature richness for fixed computation budgets and limited training samples. To overcome such limitations, we propose a new Tokens-To-Token Vision Transformer (T2T-ViT), which incorporates 1) a layer-wise Tokens-to-Token (T2T) transformation to progressively structurize the image to tokens by recursively aggregating neighboring Tokens into one Token (Tokens-to-Token), such that local structure represented by surrounding tokens can be modeled and tokens length can be reduced; 2) an efficient backbone with a deep-narrow structure for vision transformer motivated by CNN architecture design after empirical study. Notably, T2T-ViT reduces the parameter count and MACs of vanilla ViT by half, while achieving more than 3.0\% improvement when trained from scratch on ImageNet. It also outperforms ResNets and achieves comparable performance with MobileNets by directly training on ImageNet. For example, T2T-ViT with comparable size to ResNet50 (21.5M parameters) can achieve 83.3\% top1 accuracy in image resolution 384$\times$384 on ImageNet. (Code: //github.com/yitu-opensource/T2T-ViT)
While existing work in robust deep learning has focused on small pixel-level $\ell_p$ norm-based perturbations, this may not account for perturbations encountered in several real world settings. In many such cases although test data might not be available, broad specifications about the types of perturbations (such as an unknown degree of rotation) may be known. We consider a setup where robustness is expected over an unseen test domain that is not i.i.d. but deviates from the training domain. While this deviation may not be exactly known, its broad characterization is specified a priori, in terms of attributes. We propose an adversarial training approach which learns to generate new samples so as to maximize exposure of the classifier to the attributes-space, without having access to the data from the test domain. Our adversarial training solves a min-max optimization problem, with the inner maximization generating adversarial perturbations, and the outer minimization finding model parameters by optimizing the loss on adversarial perturbations generated from the inner maximization. We demonstrate the applicability of our approach on three types of naturally occurring perturbations -- object-related shifts, geometric transformations, and common image corruptions. Our approach enables deep neural networks to be robust against a wide range of naturally occurring perturbations. We demonstrate the usefulness of the proposed approach by showing the robustness gains of deep neural networks trained using our adversarial training on MNIST, CIFAR-10, and a new variant of the CLEVR dataset.
Person re-identification (re-ID) has attracted much attention recently due to its great importance in video surveillance. In general, distance metrics used to identify two person images are expected to be robust under various appearance changes. However, our work observes the extreme vulnerability of existing distance metrics to adversarial examples, generated by simply adding human-imperceptible perturbations to person images. Hence, the security danger is dramatically increased when deploying commercial re-ID systems in video surveillance, especially considering the highly strict requirement of public safety. Although adversarial examples have been extensively applied for classification analysis, it is rarely studied in metric analysis like person re-identification. The most likely reason is the natural gap between the training and testing of re-ID networks, that is, the predictions of a re-ID network cannot be directly used during testing without an effective metric. In this work, we bridge the gap by proposing Adversarial Metric Attack, a parallel methodology to adversarial classification attacks, which can effectively generate adversarial examples for re-ID. Comprehensive experiments clearly reveal the adversarial effects in re-ID systems. Moreover, by benchmarking various adversarial settings, we expect that our work can facilitate the development of robust feature learning with the experimental conclusions we have drawn.
Adversarial attacks to image classification systems present challenges to convolutional networks and opportunities for understanding them. This study suggests that adversarial perturbations on images lead to noise in the features constructed by these networks. Motivated by this observation, we develop new network architectures that increase adversarial robustness by performing feature denoising. Specifically, our networks contain blocks that denoise the features using non-local means or other filters; the entire networks are trained end-to-end. When combined with adversarial training, our feature denoising networks substantially improve the state-of-the-art in adversarial robustness in both white-box and black-box attack settings. On ImageNet, under 10-iteration PGD white-box attacks where prior art has 27.9% accuracy, our method achieves 55.7%; even under extreme 2000-iteration PGD white-box attacks, our method secures 42.6% accuracy. A network based on our method was ranked first in Competition on Adversarial Attacks and Defenses (CAAD) 2018 --- it achieved 50.6% classification accuracy on a secret, ImageNet-like test dataset against 48 unknown attackers, surpassing the runner-up approach by ~10%. Code and models will be made publicly available.
Meta-learning enables a model to learn from very limited data to undertake a new task. In this paper, we study the general meta-learning with adversarial samples. We present a meta-learning algorithm, ADML (ADversarial Meta-Learner), which leverages clean and adversarial samples to optimize the initialization of a learning model in an adversarial manner. ADML leads to the following desirable properties: 1) it turns out to be very effective even in the cases with only clean samples; 2) it is model-agnostic, i.e., it is compatible with any learning model that can be trained with gradient descent; and most importantly, 3) it is robust to adversarial samples, i.e., unlike other meta-learning methods, it only leads to a minor performance degradation when there are adversarial samples. We show via extensive experiments that ADML delivers the state-of-the-art performance on two widely-used image datasets, MiniImageNet and CIFAR100, in terms of both accuracy and robustness.
Neural networks are known to be vulnerable to adversarial examples. In this note, we evaluate the two white-box defenses that appeared at CVPR 2018 and find they are ineffective: when applying existing techniques, we can reduce the accuracy of the defended models to 0%.
Adversarial attacks are known to succeed on classifiers, but it has been an open question whether more complex vision systems are vulnerable. In this paper, we study adversarial examples for vision and language models, which incorporate natural language understanding and complex structures such as attention, localization, and modular architectures. In particular, we investigate attacks on a dense captioning model and on two visual question answering (VQA) models. Our evaluation shows that we can generate adversarial examples with a high success rate (i.e., > 90%) for these models. Our work sheds new light on understanding adversarial attacks on vision systems which have a language component and shows that attention, bounding box localization, and compositional internal structures are vulnerable to adversarial attacks. These observations will inform future work towards building effective defenses.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191