Xiaosen Wang

Zhijin Ge, Hongying Liu, Xiaosen Wang et al.

Transfer-based attack adopts the adversarial examples generated on the surrogate model to attack various models, making it applicable in the physical world and attracting increasing interest. Recently, various adversarial attacks have emerged to boost adversarial transferability from different perspectives. In this work, inspired by the observation that flat local minima are correlated with good generalization, we assume and empirically validate that adversarial examples at a flat local region tend to have good transferability by introducing a penalized gradient norm to the original loss function. Since directly optimizing the gradient regularization norm is computationally expensive and intractable for generating adversarial examples, we propose an approximation optimization method to simplify the gradient update of the objective function. Specifically, we randomly sample an example and adopt a first-order procedure to approximate the curvature of Hessian/vector product, which makes computing more efficient by interpolating two neighboring gradients. Meanwhile, in order to obtain a more stable gradient direction, we randomly sample multiple examples and average the gradients of these examples to reduce the variance due to random sampling during the iterative process. Extensive experimental results on the ImageNet-compatible dataset show that the proposed method can generate adversarial examples at flat local regions, and significantly improve the adversarial transferability on either normally trained models or adversarially trained models than the state-of-the-art attacks. Our codes are available at: https://github.com/Trustworthy-AI-Group/PGN.

Zhijin Ge, Fanhua Shang, Hongying Liu et al.

Deep neural networks are vulnerable to adversarial examples crafted by applying human-imperceptible perturbations on clean inputs. Although many attack methods can achieve high success rates in the white-box setting, they also exhibit weak transferability in the black-box setting. Recently, various methods have been proposed to improve adversarial transferability, in which the input transformation is one of the most effective methods. In this work, we notice that existing input transformation-based works mainly adopt the transformed data in the same domain for augmentation. Inspired by domain generalization, we aim to further improve the transferability using the data augmented from different domains. Specifically, a style transfer network can alter the distribution of low-level visual features in an image while preserving semantic content for humans. Hence, we propose a novel attack method named Style Transfer Method (STM) that utilizes a proposed arbitrary style transfer network to transform the images into different domains. To avoid inconsistent semantic information of stylized images for the classification network, we fine-tune the style transfer network and mix up the generated images added by random noise with the original images to maintain semantic consistency and boost input diversity. Extensive experimental results on the ImageNet-compatible dataset show that our proposed method can significantly improve the adversarial transferability on either normally trained models or adversarially trained models than state-of-the-art input transformation-based attacks. Code is available at: https://github.com/Zhijin-Ge/STM.

Xiaosen Wang, Kangheng Tong, Kun He

Transfer-based attacks generate adversarial examples on the surrogate model, which can mislead other black-box models without access, making it promising to attack real-world applications. Recently, several works have been proposed to boost adversarial transferability, in which the surrogate model is usually overlooked. In this work, we identify that non-linear layers (e.g., ReLU, max-pooling, etc.) truncate the gradient during backward propagation, making the gradient w.r.t. input image imprecise to the loss function. We hypothesize and empirically validate that such truncation undermines the transferability of adversarial examples. Based on these findings, we propose a novel method called Backward Propagation Attack (BPA) to increase the relevance between the gradient w.r.t. input image and loss function so as to generate adversarial examples with higher transferability. Specifically, BPA adopts a non-monotonic function as the derivative of ReLU and incorporates softmax with temperature to smooth the derivative of max-pooling, thereby mitigating the information loss during the backward propagation of gradients. Empirical results on the ImageNet dataset demonstrate that not only does our method substantially boost the adversarial transferability, but it is also general to existing transfer-based attacks. Code is available at https://github.com/Trustworthy-AI-Group/RPA.

Zhiyuan Wang, Zeliang Zhang, Siyuan Liang et al.

Given the great threat of adversarial attacks against Deep Neural Networks (DNNs), numerous works have been proposed to boost transferability to attack real-world applications. However, existing attacks often utilize advanced gradient calculation or input transformation but ignore the white-box model. Inspired by the fact that DNNs are over-parameterized for superior performance, we propose diversifying the high-level features (DHF) for more transferable adversarial examples. In particular, DHF perturbs the high-level features by randomly transforming the high-level features and mixing them with the feature of benign samples when calculating the gradient at each iteration. Due to the redundancy of parameters, such transformation does not affect the classification performance but helps identify the invariant features across different models, leading to much better transferability. Empirical evaluations on ImageNet dataset show that DHF could effectively improve the transferability of existing momentum-based attacks. Incorporated into the input transformation-based attacks, DHF generates more transferable adversarial examples and outperforms the baselines with a clear margin when attacking several defense models, showing its generalization to various attacks and high effectiveness for boosting transferability. Code is available at https://github.com/Trustworthy-AI-Group/DHF.

Xiaosen Wang, Zeliang Zhang, Jianping Zhang

Given the severe vulnerability of Deep Neural Networks (DNNs) against adversarial examples, there is an urgent need for an effective adversarial attack to identify the deficiencies of DNNs in security-sensitive applications. As one of the prevalent black-box adversarial attacks, the existing transfer-based attacks still cannot achieve comparable performance with the white-box attacks. Among these, input transformation based attacks have shown remarkable effectiveness in boosting transferability. In this work, we find that the existing input transformation based attacks transform the input image globally, resulting in limited diversity of the transformed images. We postulate that the more diverse transformed images result in better transferability. Thus, we investigate how to locally apply various transformations onto the input image to improve such diversity while preserving the structure of image. To this end, we propose a novel input transformation based attack, called Structure Invariant Attack (SIA), which applies a random image transformation onto each image block to craft a set of diverse images for gradient calculation. Extensive experiments on the standard ImageNet dataset demonstrate that SIA exhibits much better transferability than the existing SOTA input transformation based attacks on CNN-based and transformer-based models, showing its generality and superiority in boosting transferability. Code is available at https://github.com/xiaosen-wang/SIT.

Yijun Yang, Ruiyuan Gao, Xiaosen Wang et al.

In recent years, Text-to-Image (T2I) models have seen remarkable advancements, gaining widespread adoption. However, this progress has inadvertently opened avenues for potential misuse, particularly in generating inappropriate or Not-Safe-For-Work (NSFW) content. Our work introduces MMA-Diffusion, a framework that presents a significant and realistic threat to the security of T2I models by effectively circumventing current defensive measures in both open-source models and commercial online services. Unlike previous approaches, MMA-Diffusion leverages both textual and visual modalities to bypass safeguards like prompt filters and post-hoc safety checkers, thus exposing and highlighting the vulnerabilities in existing defense mechanisms.

Kunyu Wang, Xuanran He, Wenxuan Wang et al.

Adversarial examples mislead deep neural networks with imperceptible perturbations and have brought significant threats to deep learning. An important aspect is their transferability, which refers to their ability to deceive other models, thus enabling attacks in the black-box setting. Though various methods have been proposed to boost transferability, the performance still falls short compared with white-box attacks. In this work, we observe that existing input transformation based attacks, one of the mainstream transfer-based attacks, result in different attention heatmaps on various models, which might limit the transferability. We also find that breaking the intrinsic relation of the image can disrupt the attention heatmap of the original image. Based on this finding, we propose a novel input transformation based attack called block shuffle and rotation (BSR). Specifically, BSR splits the input image into several blocks, then randomly shuffles and rotates these blocks to construct a set of new images for gradient calculation. Empirical evaluations on the ImageNet dataset demonstrate that BSR could achieve significantly better transferability than the existing input transformation based methods under single-model and ensemble-model settings. Combining BSR with the current input transformation method can further improve the transferability, which significantly outperforms the state-of-the-art methods. Code is available at https://github.com/Trustworthy-AI-Group/BSR

Jianping Zhang, Jen-tse Huang, Wenxuan Wang et al. · pku, tencent-ai

Deep neural networks have achieved unprecedented success on diverse vision tasks. However, they are vulnerable to adversarial noise that is imperceptible to humans. This phenomenon negatively affects their deployment in real-world scenarios, especially security-related ones. To evaluate the robustness of a target model in practice, transfer-based attacks craft adversarial samples with a local model and have attracted increasing attention from researchers due to their high efficiency. The state-of-the-art transfer-based attacks are generally based on data augmentation, which typically augments multiple training images from a linear path when learning adversarial samples. However, such methods selected the image augmentation path heuristically and may augment images that are semantics-inconsistent with the target images, which harms the transferability of the generated adversarial samples. To overcome the pitfall, we propose the Path-Augmented Method (PAM). Specifically, PAM first constructs a candidate augmentation path pool. It then settles the employed augmentation paths during adversarial sample generation with greedy search. Furthermore, to avoid augmenting semantics-inconsistent images, we train a Semantics Predictor (SP) to constrain the length of the augmentation path. Extensive experiments confirm that PAM can achieve an improvement of over 4.8% on average compared with the state-of-the-art baselines in terms of the attack success rates.

Zeliang Zhang, Peihan Liu, Xiaosen Wang et al.

Deep neural networks are widely known to be vulnerable to adversarial examples, especially showing significantly poor performance on adversarial examples generated under the white-box setting. However, most white-box attack methods rely heavily on the target model and quickly get stuck in local optima, resulting in poor adversarial transferability. The momentum-based methods and their variants are proposed to escape the local optima for better transferability. In this work, we notice that the transferability of adversarial examples generated by the iterative fast gradient sign method (I-FGSM) exhibits a decreasing trend when increasing the number of iterations. Motivated by this finding, we argue that the information of adversarial perturbations near the benign sample, especially the direction, benefits more on the transferability. Thus, we propose a novel strategy, which uses the Scheduled step size and the Dual example (SD), to fully utilize the adversarial information near the benign sample. Our proposed strategy can be easily integrated with existing adversarial attack methods for better adversarial transferability. Empirical evaluations on the standard ImageNet dataset demonstrate that our proposed method can significantly enhance the transferability of existing adversarial attacks.

Zheng Fang, Xiaosen Wang, Shenyi Zhang et al.

Jailbreak attacks on audio language models (ALMs) optimize audio perturbations to elicit unsafe generations, and they typically update the entire waveform densely throughout optimization. In this work, we investigate the necessity of such dense optimization by analyzing the structure of token-aligned gradients in ALMs. We find that gradient energy is highly non-uniform across audio tokens, indicating that only a small subset of token-aligned audio regions dominates the optimization signal. Motivated by this observation, we propose Token-Aware Gradient Optimization (TAGO), which enables sparse jailbreak optimization by retaining only waveform gradients aligned with audio tokens that have high gradient energy, while masking the remaining gradients at each iteration. Across three ALMs, TAGO outperforms baselines, and substantial sparsification preserves strong attack success rates (e.g. on Qwen3-Omni, $\mathrm{ASR}_{l}$ remains at 86% with a token retention ratio of 0.25, compared to 87% with full token retention). These results demonstrate that dense waveform updates are largely redundant, and we advocate that future audio jailbreak and safety alignment research should further leverage this heterogeneous token-level gradient structure.

Xiaosen Wang, Zeyuan Yin

Mixup augmentation has been widely integrated to generate adversarial examples with superior adversarial transferability when immigrating from a surrogate model to other models. However, the underlying mechanism influencing the mixup's effect on transferability remains unexplored. In this work, we posit that the adversarial examples located at the convergence of decision boundaries across various categories exhibit better transferability and identify that Admix tends to steer the adversarial examples towards such regions. However, we find the constraint on the added image in Admix decays its capability, resulting in limited transferability. To address such an issue, we propose a new input transformation-based attack called Mixing the Image but Separating the gradienT (MIST). Specifically, MIST randomly mixes the input image with a randomly shifted image and separates the gradient of each loss item for each mixed image. To counteract the imprecise gradient, MIST calculates the gradient on several mixed images for each input sample. Extensive experimental results on the ImageNet dataset demonstrate that MIST outperforms existing SOTA input transformation-based attacks with a clear margin on both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) w/wo defense mechanisms, supporting MIST's high effectiveness and generality.

Xiaosen Wang, Zhijin Ge, Bohan Liu et al.

Adversarial transferability refers to the capacity of adversarial examples generated on the surrogate model to deceive alternate, unexposed victim models. This property eliminates the need for direct access to the victim model during an attack, thereby raising considerable security concerns in practical applications and attracting substantial research attention recently. In this work, we discern a lack of a standardized framework and criteria for evaluating transfer-based attacks, leading to potentially biased assessments of existing approaches. To rectify this gap, we have conducted an exhaustive review of hundreds of related works, organizing various transfer-based attacks into six distinct categories. Subsequently, we propose a comprehensive framework designed to serve as a benchmark for evaluating these attacks. In addition, we delineate common strategies that enhance adversarial transferability and highlight prevalent issues that could lead to unfair comparisons. Finally, we provide a brief review of transfer-based attacks beyond image classification.

Hanwen Cao, Haobo Lu, Xiaosen Wang et al.

Ensemble-based attacks have been proven to be effective in enhancing adversarial transferability by aggregating the outputs of models with various architectures. However, existing research primarily focuses on refining ensemble weights or optimizing the ensemble path, overlooking the exploration of ensemble models to enhance the transferability of adversarial attacks. To address this gap, we propose applying adversarial augmentation to the surrogate models, aiming to boost overall generalization of ensemble models and reduce the risk of adversarial overfitting. Meanwhile, observing that ensemble Vision Transformers (ViTs) gain less attention, we propose ViT-EnsembleAttack based on the idea of model adversarial augmentation, the first ensemble-based attack method tailored for ViTs to the best of our knowledge. Our approach generates augmented models for each surrogate ViT using three strategies: Multi-head dropping, Attention score scaling, and MLP feature mixing, with the associated parameters optimized by Bayesian optimization. These adversarially augmented models are ensembled to generate adversarial examples. Furthermore, we introduce Automatic Reweighting and Step Size Enlargement modules to boost transferability. Extensive experiments demonstrate that ViT-EnsembleAttack significantly enhances the adversarial transferability of ensemble-based attacks on ViTs, outperforming existing methods by a substantial margin. Code is available at https://github.com/Trustworthy-AI-Group/TransferAttack.

Xiaosen Wang, Kun He

Deep neural networks are vulnerable to adversarial examples that mislead the models with imperceptible perturbations. Though adversarial attacks have achieved incredible success rates in the white-box setting, most existing adversaries often exhibit weak transferability in the black-box setting, especially under the scenario of attacking models with defense mechanisms. In this work, we propose a new method called variance tuning to enhance the class of iterative gradient based attack methods and improve their attack transferability. Specifically, at each iteration for the gradient calculation, instead of directly using the current gradient for the momentum accumulation, we further consider the gradient variance of the previous iteration to tune the current gradient so as to stabilize the update direction and escape from poor local optima. Empirical results on the standard ImageNet dataset demonstrate that our method could significantly improve the transferability of gradient-based adversarial attacks. Besides, our method could be used to attack ensemble models or be integrated with various input transformations. Incorporating variance tuning with input transformations on iterative gradient-based attacks in the multi-model setting, the integrated method could achieve an average success rate of 90.1% against nine advanced defense methods, improving the current best attack performance significantly by 85.1% . Code is available at https://github.com/JHL-HUST/VT.

Xiaosen Wang, Xuanran He, Jingdong Wang et al.

Deep neural networks are known to be extremely vulnerable to adversarial examples under white-box setting. Moreover, the malicious adversaries crafted on the surrogate (source) model often exhibit black-box transferability on other models with the same learning task but having different architectures. Recently, various methods are proposed to boost the adversarial transferability, among which the input transformation is one of the most effective approaches. We investigate in this direction and observe that existing transformations are all applied on a single image, which might limit the adversarial transferability. To this end, we propose a new input transformation based attack method called Admix that considers the input image and a set of images randomly sampled from other categories. Instead of directly calculating the gradient on the original input, Admix calculates the gradient on the input image admixed with a small portion of each add-in image while using the original label of the input to craft more transferable adversaries. Empirical evaluations on standard ImageNet dataset demonstrate that Admix could achieve significantly better transferability than existing input transformation methods under both single model setting and ensemble-model setting. By incorporating with existing input transformations, our method could further improve the transferability and outperforms the state-of-the-art combination of input transformations by a clear margin when attacking nine advanced defense models under ensemble-model setting. Code is available at https://github.com/JHL-HUST/Admix.

Xiaosen Wang, Zhijin Ge, Shaokang Wang

Despite the notable advancements and versatility of multi-modal diffusion models, such as text-to-image models, their susceptibility to adversarial inputs remains underexplored. Contrary to expectations, our investigations reveal that the alignment between textual and Image modalities in existing diffusion models is inadequate. This misalignment presents significant risks, especially in the generation of inappropriate or Not-Safe-For-Work (NSFW) content. To this end, we propose a novel attack called Prompt-Restricted Multi-modal Attack (PReMA) to manipulate the generated content by modifying the input image in conjunction with any specified prompt, without altering the prompt itself. PReMA is the first attack that manipulates model outputs by solely creating adversarial images, distinguishing itself from prior methods that primarily generate adversarial prompts to produce NSFW content. Consequently, PReMA poses a novel threat to the integrity of multi-modal diffusion models, particularly in image-editing applications that operate with fixed prompts. Comprehensive evaluations conducted on image inpainting and style transfer tasks across various models confirm the potent efficacy of PReMA.

Ruofan Wang, Juncheng Li, Yixu Wang et al.

As large Vision-Language Models (VLMs) gain prominence, ensuring their safe deployment has become critical. Recent studies have explored VLM robustness against jailbreak attacks-techniques that exploit model vulnerabilities to elicit harmful outputs. However, the limited availability of diverse multimodal data has constrained current approaches to rely heavily on adversarial or manually crafted images derived from harmful text datasets, which often lack effectiveness and diversity across different contexts. In this paper, we propose IDEATOR, a novel jailbreak method that autonomously generates malicious image-text pairs for black-box jailbreak attacks. IDEATOR is grounded in the insight that VLMs themselves could serve as powerful red team models for generating multimodal jailbreak prompts. Specifically, IDEATOR leverages a VLM to create targeted jailbreak texts and pairs them with jailbreak images generated by a state-of-the-art diffusion model. Extensive experiments demonstrate IDEATOR's high effectiveness and transferability, achieving a 94% attack success rate (ASR) in jailbreaking MiniGPT-4 with an average of only 5.34 queries, and high ASRs of 82%, 88%, and 75% when transferred to LLaVA, InstructBLIP, and Chameleon, respectively. Building on IDEATOR's strong transferability and automated process, we introduce the VLJailbreakBench, a safety benchmark comprising 3,654 multimodal jailbreak samples. Our benchmark results on 11 recently released VLMs reveal significant gaps in safety alignment. For instance, our challenge set achieves ASRs of 46.31% on GPT-4o and 19.65% on Claude-3.5-Sonnet, underscoring the urgent need for stronger defenses. VLJailbreakBench is publicly available at https://roywang021.github.io/VLJailbreakBench.

Zilong Wang, Xiang Zheng, Xiaosen Wang et al.

Text-to-image (T2I) models such as Stable Diffusion have advanced rapidly and are now widely used in content creation. However, these models can be misused to generate harmful content, including nudity or violence, posing significant safety risks. While most platforms employ content moderation systems, underlying vulnerabilities can still be exploited by determined adversaries. Recent research on red-teaming and adversarial attacks against T2I models has notable limitations: some studies successfully generate highly toxic images but use adversarial prompts that are easily detected and blocked by safety filters, while others focus on bypassing safety mechanisms but fail to produce genuinely harmful outputs, neglecting the discovery of truly high-risk prompts. Consequently, there remains a lack of reliable tools for evaluating the safety of defended T2I models. To address this gap, we propose GenBreak, a framework that fine-tunes a red-team large language model (LLM) to systematically explore underlying vulnerabilities in T2I generators. Our approach combines supervised fine-tuning on curated datasets with reinforcement learning via interaction with a surrogate T2I model. By integrating multiple reward signals, we guide the LLM to craft adversarial prompts that enhance both evasion capability and image toxicity, while maintaining semantic coherence and diversity. These prompts demonstrate strong effectiveness in black-box attacks against commercial T2I generators, revealing practical and concerning safety weaknesses.

Xiaosen Wang, Kunyu Wang

Deep neural networks (DNNs) are vulnerable to various types of adversarial examples, bringing huge threats to security-critical applications. Among these, adversarial patches have drawn increasing attention due to their good applicability to fool DNNs in the physical world. However, existing works often generate patches with meaningless noise or patterns, making it conspicuous to humans. To address this issue, we explore how to generate visually realistic adversarial patches to fool DNNs. Firstly, we analyze that a high-quality adversarial patch should be realistic, position irrelevant, and printable to be deployed in the physical world. Based on this analysis, we propose an effective attack called VRAP, to generate visually realistic adversarial patches. Specifically, VRAP constrains the patch in the neighborhood of a real image to ensure the visual reality, optimizes the patch at the poorest position for position irrelevance, and adopts Total Variance loss as well as gamma transformation to make the generated patch printable without losing information. Empirical evaluations on the ImageNet dataset demonstrate that the proposed VRAP exhibits outstanding attack performance in the digital world. Moreover, the generated adversarial patches can be disguised as the scrawl or logo in the physical world to fool the deep models without being detected, bringing significant threats to DNNs-enabled applications.

Yuyang Luo, Xiaosen Wang, Zhijin Ge et al.

Adversarial examples pose significant threats to deep neural networks (DNNs), and their property of transferability in the black-box setting has led to the emergence of transfer-based attacks, making it feasible to target real-world applications employing DNNs. Among them, feature-level attacks, where intermediate features are perturbed based on feature importance weight matrix computed from transformed images, have gained popularity. In this work, we find that existing feature-level attacks primarily manipulate the semantic information to derive the weight matrix. Inspired by several works that find CNNs tend to focus more on high-frequency components (a.k.a. abstract features, e.g., texture, edge, etc.), we validate that transforming images in the high-frequency space also improves transferability. Based on this finding, we propose a balanced approach called Semantic and Abstract FEatures disRuption (SAFER). Specifically, SAFER conducts BLOCKMIX on the input image and SELF-MIX on the frequency spectrum when computing the weight matrix to highlight crucial features. By using such a weight matrix, we can direct the attacker to disrupt both semantic and abstract features, leading to improved transferability. Extensive experiments on the ImageNet dataset also demonstrate the effectiveness of our method in boosting adversarial transferability.

Ruofan Wang, Xiang Zheng, Xiaosen Wang et al.

Vision-Language Models (VLMs) are vulnerable to jailbreak attacks, where adversaries bypass safety mechanisms to elicit harmful outputs. In this work, we examine an insidious variant of this threat: toxic continuation. Unlike standard jailbreaks that rely solely on malicious instructions, toxic continuation arises when the model is given a malicious input alongside a partial toxic output, resulting in harmful completions. This vulnerability poses a unique challenge in multimodal settings, where even subtle image variations can disproportionately affect the model's response. To this end, we propose RedDiffuser (RedDiff), the first red teaming framework that uses reinforcement learning to fine-tune diffusion models into generating natural-looking adversarial images that induce toxic continuations. RedDiffuser integrates a greedy search procedure for selecting candidate image prompts with reinforcement fine-tuning that jointly promotes toxic output and semantic coherence. Experiments demonstrate that RedDiffuser significantly increases the toxicity rate in LLaVA outputs by 10.69% and 8.91% on the original and hold-out sets, respectively. It also exhibits strong transferability, increasing toxicity rates on Gemini by 5.1% and on LLaMA-Vision by 26.83%. These findings uncover a cross-modal toxicity amplification vulnerability in current VLM alignment, highlighting the need for robust multimodal red teaming. We will release the RedDiffuser codebase to support future research.

Bohan Liu, Xiaosen Wang

Transfer-based attacks pose a significant threat to real-world applications by directly targeting victim models with adversarial examples generated on surrogate models. While numerous approaches have been proposed to enhance adversarial transferability, existing works often overlook the intrinsic relationship between adversarial perturbations and input images. In this work, we find that adversarial perturbation often exhibits poor translation invariance for a given clean image and model, which is attributed to local invariance. Through empirical analysis, we demonstrate that there is a positive correlation between the local invariance of adversarial perturbations w.r.t. the input image and their transferability across different models. Based on this finding, we propose a general adversarial transferability boosting technique called Local Invariance Boosting approach (LI-Boost). Extensive experiments on the standard ImageNet dataset demonstrate that LI-Boost could significantly boost various types of transfer-based attacks (e.g., gradient-based, input transformation-based, model-related, advanced objective function, ensemble, etc.) on CNNs, ViTs, and defense mechanisms. Our approach presents a promising direction for future research in improving adversarial transferability across different models.

Zeliang Zhang, Wei Yao, Xiaosen Wang

Deep neural networks are widely known to be vulnerable to adversarial examples. However, vanilla adversarial examples generated under the white-box setting often exhibit low transferability across different models. Since adversarial transferability poses more severe threats to practical applications, various approaches have been proposed for better transferability, including gradient-based, input transformation-based, and model-related attacks, \etc. In this work, we find that several tiny changes in the existing adversarial attacks can significantly affect the attack performance, \eg, the number of iterations and step size. Based on careful studies of existing adversarial attacks, we propose a bag of tricks to enhance adversarial transferability, including momentum initialization, scheduled step size, dual example, spectral-based input transformation, and several ensemble strategies. Extensive experiments on the ImageNet dataset validate the high effectiveness of our proposed tricks and show that combining them can further boost adversarial transferability. Our work provides practical insights and techniques to enhance adversarial transferability, and offers guidance to improve the attack performance on the real-world application through simple adjustments.

Yichen Yang, Xiaosen Wang, Kun He

We attribute the vulnerability of natural language processing models to the fact that similar inputs are converted to dissimilar representations in the embedding space, leading to inconsistent outputs, and we propose a novel robust training method, termed Fast Triplet Metric Learning (FTML). Specifically, we argue that the original sample should have similar representation with its adversarial counterparts and distinguish its representation from other samples for better robustness. To this end, we adopt the triplet metric learning into the standard training to pull words closer to their positive samples (i.e., synonyms) and push away their negative samples (i.e., non-synonyms) in the embedding space. Extensive experiments demonstrate that FTML can significantly promote the model robustness against various advanced adversarial attacks while keeping competitive classification accuracy on original samples. Besides, our method is efficient as it only needs to adjust the embedding and introduces very little overhead on the standard training. Our work shows great potential of improving the textual robustness through robust word embedding.

Zhen Yu, Xiaosen Wang, Wanxiang Che et al.

Existing textual adversarial attacks usually utilize the gradient or prediction confidence to generate adversarial examples, making it hard to be deployed in real-world applications. To this end, we consider a rarely investigated but more rigorous setting, namely hard-label attack, in which the attacker can only access the prediction label. In particular, we find we can learn the importance of different words via the change on prediction label caused by word substitutions on the adversarial examples. Based on this observation, we propose a novel adversarial attack, termed Text Hard-label attacker (TextHacker). TextHacker randomly perturbs lots of words to craft an adversarial example. Then, TextHacker adopts a hybrid local search algorithm with the estimation of word importance from the attack history to minimize the adversarial perturbation. Extensive evaluations for text classification and textual entailment show that TextHacker significantly outperforms existing hard-label attacks regarding the attack performance as well as adversary quality.

Xiaosen Wang, Zeliang Zhang, Kangheng Tong et al.

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.

Xiaosen Wang, Yifeng Xiong, Kun He

A line of work has shown that natural text processing models are vulnerable to adversarial examples. Correspondingly, various defense methods are proposed to mitigate the threat of textual adversarial examples, eg, adversarial training, input transformations, detection, etc. In this work, we treat the optimization process for synonym substitution based textual adversarial attacks as a specific sequence of word replacement, in which each word mutually influences other words. We identify that we could destroy such mutual interaction and eliminate the adversarial perturbation by randomly substituting a word with its synonyms. Based on this observation, we propose a novel textual adversarial example detection method, termed Randomized Substitution and Vote (RS&V), which votes the prediction label by accumulating the logits of k samples generated by randomly substituting the words in the input text with synonyms. The proposed RS&V is generally applicable to any existing neural networks without modification on the architecture or extra training, and it is orthogonal to prior work on making the classification network itself more robust. Empirical evaluations on three benchmark datasets demonstrate that our RS&V could detect the textual adversarial examples more successfully than the existing detection methods while maintaining the high classification accuracy on benign samples.

Xiaosen Wang, Chuanbiao Song, Liwei Wang et al.

In the field of adversarial robustness, there is a common practice that adopts the single-step adversarial training for quickly developing adversarially robust models. However, the single-step adversarial training is most likely to cause catastrophic overfitting, as after a few training epochs it will be hard to generate strong adversarial examples to continuously boost the adversarial robustness. In this work, we aim to avoid the catastrophic overfitting by introducing multi-step adversarial examples during the single-step adversarial training. Then, to balance the large training overhead of generating multi-step adversarial examples, we propose a Multi-stage Optimization based Adversarial Training (MOAT) method that periodically trains the model on mixed benign examples, single-step adversarial examples, and multi-step adversarial examples stage by stage. In this way, the overall training overhead is reduced significantly, meanwhile, the model could avoid catastrophic overfitting. Extensive experiments on CIFAR-10 and CIFAR-100 datasets demonstrate that under similar amount of training overhead, the proposed MOAT exhibits better robustness than either single-step or multi-step adversarial training methods.

Xiaosen Wang, Jiadong Lin, Han Hu et al.

Deep learning models are known to be vulnerable to adversarial examples crafted by adding human-imperceptible perturbations on benign images. Many existing adversarial attack methods have achieved great white-box attack performance, but exhibit low transferability when attacking other models. Various momentum iterative gradient-based methods are shown to be effective to improve the adversarial transferability. In what follows, we propose an enhanced momentum iterative gradient-based method to further enhance the adversarial transferability. Specifically, instead of only accumulating the gradient during the iterative process, we additionally accumulate the average gradient of the data points sampled in the gradient direction of the previous iteration so as to stabilize the update direction and escape from poor local maxima. Extensive experiments on the standard ImageNet dataset demonstrate that our method could improve the adversarial transferability of momentum-based methods by a large margin of 11.1% on average. Moreover, by incorporating with various input transformation methods, the adversarial transferability could be further improved significantly. We also attack several extra advanced defense models under the ensemble-model setting, and the enhancements are remarkable with at least 7.8% on average.

Xiaosen Wang, Yichen Yang, Yihe Deng et al.

Adversarial training is the most empirically successful approach in improving the robustness of deep neural networks for image classification.For text classification, however, existing synonym substitution based adversarial attacks are effective but not efficient to be incorporated into practical text adversarial training. Gradient-based attacks, which are very efficient for images, are hard to be implemented for synonym substitution based text attacks due to the lexical, grammatical and semantic constraints and the discrete text input space. Thereby, we propose a fast text adversarial attack method called Fast Gradient Projection Method (FGPM) based on synonym substitution, which is about 20 times faster than existing text attack methods and could achieve similar attack performance. We then incorporate FGPM with adversarial training and propose a text defense method called Adversarial Training with FGPM enhanced by Logit pairing (ATFL). Experiments show that ATFL could significantly improve the model robustness and block the transferability of adversarial examples.

Xiaosen Wang, Hao Jin, Yichen Yang et al.

In the area of natural language processing, deep learning models are recently known to be vulnerable to various types of adversarial perturbations, but relatively few works are done on the defense side. Especially, there exists few effective defense method against the successful synonym substitution based attacks that preserve the syntactic structure and semantic information of the original text while fooling the deep learning models. We contribute in this direction and propose a novel adversarial defense method called Synonym Encoding Method (SEM). Specifically, SEM inserts an encoder before the input layer of the target model to map each cluster of synonyms to a unique encoding and trains the model to eliminate possible adversarial perturbations without modifying the network architecture or adding extra data. Extensive experiments demonstrate that SEM can effectively defend the current synonym substitution based attacks and block the transferability of adversarial examples. SEM is also easy and efficient to scale to large models and big datasets.

Kun He, Wu Wang, Xiaosen Wang et al.

Separable Non-negative Matrix Factorization (SNMF) is an important method for topic modeling, where "separable" assumes every topic contains at least one anchor word, defined as a word that has non-zero probability only on that topic. SNMF focuses on the word co-occurrence patterns to reveal topics by two steps: anchor word selection and topic recovery. The quality of the anchor words strongly influences the quality of the extracted topics. Existing anchor word selection algorithm is to greedily find an approximate convex hull in a high-dimensional word co-occurrence space. In this work, we propose a new method for the anchor word selection by associating the word co-occurrence probability with the words similarity and assuming that the most different words on semantic are potential candidates for the anchor words. Therefore, if the similarity of a word-pair is very low, then the two words are very likely to be the anchor words. According to the statistical information of text corpora, we can get the similarity of all word-pairs. We build the word similarity graph where the nodes correspond to words and weights on edges stand for the word-pair similarity. Following this way, we design a greedy method to find a minimum edge-weight anchor clique of a given size in the graph for the anchor word selection. Extensive experiments on real-world corpus demonstrate the effectiveness of the proposed anchor word selection method that outperforms the common convex hull-based methods on the revealed topic quality. Meanwhile, our method is much faster than typical SNMF based method.

Xiaosen Wang, Kun He, Chuanbiao Song et al.

Despite the rapid development of adversarial machine learning, most adversarial attack and defense researches mainly focus on the perturbation-based adversarial examples, which is constrained by the input images. In comparison with existing works, we propose non-constrained adversarial examples, which are generated entirely from scratch without any constraint on the input. Unlike perturbation-based attacks, or the so-called unrestricted adversarial attack which is still constrained by the input noise, we aim to learn the distribution of adversarial examples to generate non-constrained but semantically meaningful adversarial examples. Following this spirit, we propose a novel attack framework called AT-GAN (Adversarial Transfer on Generative Adversarial Net). Specifically, we first develop a normal GAN model to learn the distribution of benign data, and then transfer the pre-trained GAN model to estimate the distribution of adversarial examples for the target model. In this way, AT-GAN can learn the distribution of adversarial examples that is very close to the distribution of real data. To our knowledge, this is the first work of building an adversarial generator model that could produce adversarial examples directly from any input noise. Extensive experiments and visualizations show that the proposed AT-GAN can very efficiently generate diverse adversarial examples that are more realistic to human perception. In addition, AT-GAN yields higher attack success rates against adversarially trained models under white-box attack setting and exhibits moderate transferability against black-box models.