Trojan attacks on deep neural networks are both dangerous and surreptitious. Over the past few years, Trojan attacks have advanced from using only a single input-agnostic trigger and targeting only one class to using multiple, input-specific triggers and targeting multiple classes. However, Trojan defenses have not caught up with this development. Most defense methods still make out-of-date assumptions about Trojan triggers and target classes, thus, can be easily circumvented by modern Trojan attacks. To deal with this problem, we propose two novel ""filtering"" defenses called Variational Input Filtering (VIF) and Adversarial Input Filtering (AIF) which leverage lossy data compression and adversarial learning respectively to effectively purify all potential Trojan triggers in the input at run time without making assumptions about the number of triggers/target classes or the input dependence property of triggers. In addition, we introduce a new defense mechanism called ""Filtering-then-Contrasting"" (FtC) which helps avoid the drop in classification accuracy on clean data caused by ""filtering"", and combine it with VIF/AIF to derive new defenses of this kind. Extensive experimental results and ablation studies show that our proposed defenses significantly outperform well-known baseline defenses in mitigating five advanced Trojan attacks including two recent state-of-the-art while being quite robust to small amounts of training data and large-norm triggers.
论文检索
输入标题、作者或关键词，从 1,645 篇学术成果中精准定位
Zero-shot fine-grained recognition is an important classification task, whose goal is to recognize visually very similar classes, including the ones without training images. Despite recent advances on the development of zero-shot fine-grained recognition methods, the robustness of such models to adversarial attacks is not well understood. On the other hand, adversarial attacks have been widely studied for conventional classification with visually distinct classes. Such attacks, in particular, universal perturbations that are class-agnostic and ideally should generalize to unseen classes, however, cannot leverage or capture small distinctions among fine-grained classes. Therefore, we propose a compositional attribute-based framework for generating adversarial attacks on zero-shot fine-grained recognition models. To generate attacks that capture small differences between fine-grained classes, generalize well to previously unseen classes and can be applied in real-time, we propose to learn and compose multiple attribute-based universal perturbations (AUPs). Each AUP corresponds to an image-agnostic perturbation on a specific attribute. To build our attack, we compose AUPs with weights obtained by learning a class-attribute compatibility function. To learn the AUPs and the parameters of our model, we minimize a loss, consisting of a ranking loss and a novel utility loss, which ensures AUPs are effectively learned and utilized. By extensive experiments on three datasets for zero-shot fine-grained recognition, we show that our attacks outperform conventional universal classification attacks and transfer well between different recognition architectures.
Despite the fundamental distinction between adversarial and natural training (AT and NT), AT methods generally adopt momentum SGD (MSGD) for the outer optimization. This paper aims to analyze this choice by investigating the overlooked role of outer optimization in AT. Our exploratory evaluations reveal that AT induces higher gradient norm and variance compared to NT. This phenomenon hinders the outer optimization in AT since the convergence rate of MSGD is highly dependent on the variance of the gradients. To this end, we propose an optimization method called ENGM which regularizes the contribution of each input example to the average mini-batch gradients. We prove that the convergence rate of ENGM is independent of the variance of the gradients, and thus, it is suitable for AT. We introduce a trick to reduce the computational cost of ENGM using empirical observations on the correlation between the norm of gradients w.r.t. the network parameters and input examples. Our extensive evaluations and ablation studies on CIFAR-10, CIFAR-100, and TinyImageNet demonstrate that ENGM and its variants consistently improve the performance of a wide range of AT methods. Furthermore, ENGM alleviates major shortcomings of AT including robust overfitting and high sensitivity to hyperparameter settings.
Graph neural networks (GNNs) have achieved outstanding performance in semi-supervised learning tasks with partially labeled graph structured data. However, labeling graph data for training is a challenging task, and inaccurate labels may mislead the training process to erroneous GNN models for node classification. In this paper, we consider label poisoning attacks on training data, where the labels of input data are modified by an adversary before training, to understand to what extent the state-of-the-art GNN models are resistant/vulnerable to such attacks. Specifically, we propose a label poisoning attack framework for graph convolutional networks (GCNs), inspired by the equivalence between label propagation and decoupled GCNs that separate message passing from neural networks. Instead of attacking the entire GCN models, we propose to attack solely label propagation for message passing. It turns out that a gradient-based attack on label propagation is effective and efficient towards the misleading of GCN training. More remarkably, such label attack can be topology-agnostic in the sense that the labels to be attacked can be efficiently chosen without knowing graph structures. Extensive experimental results demonstrate the effectiveness of the proposed method against state-of-the-art GCN-like models.
We study a new approach to learning energy-based models (EBMs) based on adversarial training (AT). We show that (binary) AT learns a special kind of energy function that models the support of the data distribution, and the learning process is closely related to MCMC-based maximum likelihood learning of EBMs. We further propose improved techniques for generative modeling with AT, and demonstrate that this new approach is capable of generating diverse and realistic images. Aside from having competitive image generation performance to explicit EBMs, the studied approach is stable to train, is well-suited for image translation tasks, and exhibits strong out-of-distribution adversarial robustness. Our results demonstrate the viability of the AT approach to generative modeling, suggesting that AT is a competitive alternative approach to learning EBMs.
Previous studies have verified that the functionality of black-box models can be stolen with full probability outputs. However, under the more practical hard-label setting, we observe that existing methods suffer from catastrophic performance degradation. We argue this is due to the lack of rich information in the probability prediction and the overfitting caused by hard labels. To this end, we propose a novel hard-label model stealing method termed black-box dissector, which consists of two erasing-based modules. One is a CAM-driven erasing strategy that is designed to increase the information capacity hidden in hard labels from the victim model. The other is a random-erasing-based self-knowledge distillation module that utilizes soft labels from the substitute model to mitigate overfitting. Extensive experiments on four widely-used datasets consistently demonstrate that our method outperforms state-of-the-art methods, with an improvement of at most 8.27%. We also validate the effectiveness and practical potential of our method on real-world APIs and defense methods. Furthermore, our method promotes other downstream tasks, i.e., transfer adversarial attacks.
Recent studies have shown that Deep Neural Networks (DNNs) are vulnerable to the backdoor attacks, which leads to malicious behaviors of DNNs when specific triggers are attached to the input images. It was further demonstrated that the infected DNNs possess a collection of channels, which are more sensitive to the backdoor triggers compared with normal channels. Pruning these channels was then shown to be effective in mitigating the backdoor behaviors. To locate those channels, it is natural to consider their Lipschitzness, which measures their sensitivity against worst-case perturbations on the inputs. In this work, we introduce a novel concept called Channel Lipschitz Constant (CLC), which is defined as the Lipschitz constant of the mapping from the input images to the output of each channel. Then we provide empirical evidences to show the strong correlation between an Upper bound of the CLC (UCLC) and the trigger-activated change on the channel activation. Since UCLC can be directly calculated from the weight matrices, we can detect the potential backdoor channels in a data-free manner, and do simple pruning on the infected DNN to repair the model. The proposed Channel Lipschitzness based Pruning (CLP) method is super fast, simple, data-free and robust to the choice of the pruning threshold. Extensive experiments are conducted to evaluate the efficiency and effectiveness of CLP, which achieves state-of-the-art results among the mainstream defense methods even without any data. Source codes are available at https://github.com/rkteddy/channel-Lipschitzness-based-pruning.
Decision-based attack poses a severe threat to real-world applications since it regards the target model as a black box and only accesses the hard prediction label. Great efforts have been made recently to decrease the number of queries; however, existing decision-based attacks still require thousands of queries in order to generate good quality adversarial examples. In this work, we find that a benign sample, the current and the next adversarial examples can naturally construct a triangle in a subspace for any iterative attacks. Based on the law of sines, we propose a novel Triangle Attack (TA) to optimize the perturbation by utilizing the geometric information that the longer side is always opposite the larger angle in any triangle. However, directly applying such information on the input image is ineffective because it cannot thoroughly explore the neighborhood of the input sample in the high dimensional space. To address this issue, TA optimizes the perturbation in the low frequency space for effective dimensionality reduction owing to the generality of such geometric property. Extensive evaluations on ImageNet dataset show that TA achieves a much higher attack success rate within 1,000 queries and needs a much less number of queries to achieve the same attack success rate under various perturbation budgets than existing decision-based attacks. With such high efficiency, we further validate the applicability of TA on real-world API, i.e., Tencent Cloud API.
Transfer learning through the use of pre-trained models has become a growing trend for the machine learning community. Consequently, numerous pre-trained models are released online to facilitate further research. However, it raises extensive concerns on whether these pre-trained models would leak privacy-sensitive information of their training data. Thus, in this work, we aim to answer the following questions: ""Can we effectively recover private information from these pre-trained models? What are the sufficient conditions to retrieve such sensitive information?"" We first explore different statistical information which can discriminate the private training distribution from other distributions. Based on our observations, we propose a novel private data reconstruction framework, SecretGen, to effectively recover private information. Compared with previous methods which can recover private data with the ground true prediction of the targeted recovery instance, SecretGen does not require such prior knowledge, making it more practical. We conduct extensive experiments on different datasets under diverse scenarios to compare SecretGen with other baselines and provide a systematic benchmark to better understand the impact of different auxiliary information and optimization operations. We show that without prior knowledge about true class prediction, SecretGen is able to recover private data with similar performance compared with the ones that leverage such prior knowledge. If the prior knowledge is given, SecretGen will significantly outperform baseline methods. We also propose several quantitative metrics to further quantify the privacy vulnerability of pre-trained models, which will help the model selection for privacy-sensitive applications. Our code is available at: https://github.com/AI-secure/SecretGen.
The vulnerability of DNNs severely limits the application of it in the security-sensitive domains. Most of the existing methods improve the robustness of models from weight optimization, such as adversarial training and regularization. However, the architecture is also a key factor to robustness, which is often neglected or underestimated. We propose Robust Network Architecture Search (RNAS) to obtain a robust network against adversarial attacks. We observe that an adversarial perturbation distorting the non-robust features in latent feature space can further aggravate misclassification. Based on this observation, we search the robust architecture through restricting feature distortion in the search process. Specifically, we define a network vulnerability metric based on feature distortion as a constraint in the search process. This process is modeled as a multi-objective bilevel optimization problem and an effective algorithm is proposed to solve this optimization. Extensive experiments conducted on CIFAR-10/100, SVHN and Tiny-ImageNet show that RNAS achieves the best robustness under various adversarial attacks compared with extensive baselines and state-of-the-art methods.
The security of deep neural networks (DNNs) has attracted increasing attention due to their widespread use in various applications. Recently, the deployed DNNs have been demonstrated to be vulnerable to Trojan attacks, which manipulate model parameters with bit flips to inject a hidden behavior and activate it by a specific trigger pattern. However, all existing Trojan attacks adopt noticeable patch-based triggers (e.g., a square pattern), making them perceptible to humans and easy to be spotted by machines. In this paper, we present a novel attack, namely hardly perceptible Trojan attack (HPT). HPT crafts hardly perceptible Trojan images by utilizing the additive noise and per-pixel flow field to tweak the pixel values and positions of the original images, respectively. To achieve superior attack performance, we propose to jointly optimize bit flips, additive noise, and flow field. Since the weight bits of the DNNs are binary, this problem is very hard to be solved. We handle the binary constraint with equivalent replacement and provide an effective optimization algorithm. Extensive experiments on CIFAR-10, SVHN, and ImageNet datasets show that the proposed HPT can generate hardly perceptible Trojan images, while achieving comparable or better attack performance compared to the state-of-the-art methods. The code is available at: https://github.com/jiawangbai/HPT.
We study certified robustness of machine learning classifiers against adversarial perturbations. In particular, we propose the first universally approximated certified robustness (UniCR) framework, which can approximate the robustness certification of \emph{any} input on \emph{any} classifier against \emph{any} $\ell_p$ perturbations with noise generated by \emph{any} continuous probability distribution. Compared with the state-of-the-art certified defenses, UniCR provides many significant benefits: (1) the first universal robustness certification framework for the above 4 “any”s; (2) automatic robustness certification that avoids case-by-case analysis, (3) tightness validation of certified robustness, and (4) optimality validation of noise distributions used by randomized smoothing. We conduct extensive experiments to validate the above benefits of UniCR and the advantages of UniCR over state-of-the-art certified defenses against $\ell_p$ perturbations.
Adversarial robustness is critical for deep learning models to defend against adversarial attacks. Although adversarial training is considered to be one of the most effective ways to improve the model’s adversarial robustness, it usually yields models with lower natural accuracy. In this paper, we argue that, for the attackable examples, traditional adversarial training which utilizes a fixed size perturbation ball can create adversarial examples that deviate far away from the original class towards the target class. Thus, the model’s performance on the natural target class will drop drastically, which leads to the decline of natural accuracy. To this end, we propose the Data-Adaptive Adversarial Training (DAAT) which adaptively adjusts the perturbation ball to a proper size for each of the natural examples with the help of a natural trained calibration network. Besides, a dynamic training strategy empowers the DAAT models with impressive robustness while retaining remarkable natural accuracy. Based on a toy example, we theoretically prove the recession of the natural accuracy caused by adversarial training and show how the data-adaptive perturbation size helps the model resist it. Finally, empirical experiments on benchmark datasets demonstrate the significant improvement of DAAT models on natural accuracy compared with strong baselines.
Contrastive learning (CL) has recently been applied to adversarial learning tasks. Such practice considers adversarial perturbations as additional positive samples of an instance, and by maximizing their agreements with each other, yields better adversarial robustness. However, this mechanism can be potentially flawed, since adversarial perturbations may cause instance-level identity confusion, which can impede CL performance by pulling together different instances with separate identities. To address this issue, we propose to treat adversarial samples unequally when contrasted to positive and negative samples, with an asymmetric InfoNCE objective (A-InfoNCE) that allows discriminating considerations of adversarial samples. Specifically, adversaries are viewed as inferior positives that induce weaker learning signals, or as hard negatives exhibiting higher contrast to other negative samples. In the asymmetric fashion, the adverse impacts of conflicting objectives between CL and adversarial learning can be effectively mitigated. Experiments show that our approach consistently outperforms existing Adversarial CL methods across different finetuning schemes without additional computational cost. The proposed A-InfoNCE is also a generic form that can be readily extended to different CL methods. Code is available at https://github.com/yqy2001/A-InfoNCE.
Trajectory prediction is essential for autonomous vehicles (AVs) to plan correct and safe driving behaviors. While many prior works aim to achieve higher prediction accuracy, few studies the adversarial robustness of their methods. To bridge this gap, we propose to study the adversarial robustness of data-driven trajectory prediction systems. We devise an optimization-based adversarial attack framework that leverages a carefully-designed differentiable dynamic model to generate realistic adversarial trajectories. Empirically, we benchmark the adversarial robustness of state-of-the-art prediction models and show that our attack increases the prediction error for both general metrics and planning-aware metrics by more than 50% and 37%. We also show that our attack can lead an AV to drive off-road or collide into other vehicles in simulation. Finally, we demonstrate how to mitigate the adversarial attacks using an adversarial training scheme.
What is really needed to make an existing 2D GAN 3Daware? To answer this question, we modify a classical GAN, i.e., StyleGANv2, as little as possible. We find that only two modifications are absolutely necessary: 1) a multiplane image style generator branch which produces a set of alpha maps conditioned on their depth; 2) a pose conditioned discriminator. We refer to the generated output as a ‘generative multiplane image’ (GMPI) and emphasize that its renderings are not only high-quality but also guaranteed to be view-consistent, which makes GMPIs different from many prior works. Importantly, the number of alpha maps can be dynamically adjusted and can differ between training and inference, alleviating memory concerns and enabling fast training of GMPIs in less than half a day at a resolution of 1024^2. Our findings are consistent across three challenging and common high-resolution datasets, including FFHQ, AFHQv2 and MetFaces.
Adversarial attacks provide a good way to study the robustness of deep learning models. One category of methods in transfer-based black-box attack utilizes several image transformation operations to improve the transferability of adversarial examples, which is effective, but fails to take the specific characteristic of the input image into consideration. In this work, we propose a novel architecture, called Adaptive Image Transformation Learner (AITL), which incorporates different image transformation operations into a unified framework to further improve the transferability of adversarial examples. Unlike the fixed combinational transformations used in existing works, our elaborately designed transformation learner adaptively selects the most effective combination of image transformations specific to the input image. Extensive experiments on ImageNet demonstrate that our method significantly improves the attack success rates on both normally trained models and defense models under various settings.
Boosting Transferability of Targeted Adversarial Examples via Hierarchical Generative Networks
PDF ↗Transfer-based adversarial attacks can evaluate model robustness in the black-box setting. Several methods have demonstrated impressive untargeted transferability, however, it is still challenging to efficiently produce targeted transferability. To this end, we develop a simple yet effective framework to craft targeted transfer-based adversarial examples, applying a hierarchical generative network. In particular, we contribute to amortized designs that well adapt to multi-class targeted attacks. Extensive experiments on ImageNet show that our method improves the success rates of targeted black-box attacks by a significant margin over the existing methods --- it reaches an average success rate of 29.1% against six diverse models based only on one substitute white-box model, which significantly outperforms the state-of-the-art gradient-based attack methods. Moreover, the proposed method is also more efficient beyond an order of magnitude than gradient-based methods.
Recently backdoor attack has become an emerging threat to the security of deep neural network (DNN) models. To date, most of the existing studies focus on backdoor attack against the uncompressed model; while the vulnerability of compressed DNNs, which are widely used in the practical applications, is little exploited yet. In this paper, we propose to study and develop Robust and Imperceptible Backdoor Attack against Compact DNN models (RIBAC). By performing systematic analysis and exploration on the important design knobs, we propose a framework that can learn the proper trigger patterns, model parameters and pruning masks in an efficient way. Thereby achieving high trigger stealthiness, high attack success rate and high model efficiency simultaneously. Extensive evaluations across different datasets, including the test against the state-of-the-art defense mechanisms, demonstrate the high robustness, stealthiness and model efficiency of RIBAC. Code is available at https://github.com/huyvnphan/ECCV2022-RIBAC
Learning Extremely Lightweight and Robust Model with Differentiable Constraints on Sparsity and Condition Number
PDF ↗Learning lightweight and robust deep learning models is an enormous challenge for safety-critical devices with limited computing and memory resources, owing to robustness against adversarial attacks being proportional to network capacity. The community has extensively explored the integration of adversarial training and model compression, such as weight pruning. However, lightweight models generated by highly pruned over-parameterized models lead to sharp drops in both robust and natural accuracy. It has been observed that the parameters of these models lie in ill-conditioned weight space, i.e., the condition number of weight matrices tend to be large enough that the model is not robust. In this work, we propose a framework for building extremely lightweight models, which combines tensor product with the differentiable constraints for reducing condition number and promoting sparsity. Moreover, the proposed framework is incorporated into adversarial training with the min-max optimization scheme. We evaluate the proposed approach on VGG-16 and Visual Transformer. Experimental results on datasets such as ImageNet, SVHN, and CIFAR-10 show that we can achieve an overwhelming advantage at a high compression ratio, e.g., 200 times.