Hengwei Zhang

Bo Yang, Hengwei Zhang, Jindong Wang et al.

Transferable adversarial examples cause practical security risks since they can mislead a target model without knowing its internal knowledge. A conventional recipe for maximizing transferability is to keep only the optimal adversarial example from all those obtained in the optimization pipeline. In this paper, for the first time, we revisit this convention and demonstrate that those discarded, sub-optimal adversarial examples can be reused to boost transferability. Specifically, we propose ``Adversarial Example Soups'' (AES), with AES-tune for averaging discarded adversarial examples in hyperparameter tuning and AES-rand for stability testing. In addition, our AES is inspired by ``model soups'', which averages weights of multiple fine-tuned models for improved accuracy without increasing inference time. Extensive experiments validate the global effectiveness of our AES, boosting 10 state-of-the-art transfer attacks and their combinations by up to 13\% against 10 diverse (defensive) target models. We also show the possibility of generalizing AES to other types, \textit{e.g.}, directly averaging multiple in-the-wild adversarial examples that yield comparable success. A promising byproduct of AES is the improved stealthiness of adversarial examples since the perturbation variances are naturally reduced.

Bo Yang, Hengwei Zhang, Jindong Wang et al.

In surrogate ensemble attacks, using more surrogate models yields higher transferability but lower resource efficiency. This practical trade-off between transferability and efficiency has largely limited existing attacks despite many pre-trained models are easily accessible online. In this paper, we argue that such a trade-off is caused by an unnecessary common assumption, i.e., all models should be \textit{identical} across iterations. By lifting this assumption, we can use as many surrogates as we want to unleash transferability without sacrificing efficiency. Concretely, we propose Selective Ensemble Attack (SEA), which dynamically selects diverse models (from easily accessible pre-trained models) across iterations based on our new interpretation of decoupling within-iteration and cross-iteration model diversity. In this way, the number of within-iteration models is fixed for maintaining efficiency, while only cross-iteration model diversity is increased for higher transferability. Experiments on ImageNet demonstrate the superiority of SEA in various scenarios. For example, when dynamically selecting 4 from 20 accessible models, SEA yields 8.5% higher transferability than existing attacks under the same efficiency. The superiority of SEA also generalizes to real-world systems, such as commercial vision APIs and large vision-language models. Overall, SEA opens up the possibility of adaptively balancing transferability and efficiency according to specific resource requirements.

Bo Yang, Hengwei Zhang, Yuchen Zhang et al.

Deep neural networks are vulnerable to adversarial examples, which are crafted by applying small, human-imperceptible perturbations on the original images, so as to mislead deep neural networks to output inaccurate predictions. Adversarial attacks can thus be an important method to evaluate and select robust models in safety-critical applications. However, under the challenging black-box setting, most existing adversarial attacks often achieve relatively low success rates on adversarially trained networks and advanced defense models. In this paper, we propose AdaBelief Iterative Fast Gradient Method (ABI-FGM) and Crop-Invariant attack Method (CIM) to improves the transferability of adversarial examples. ABI-FGM and CIM can be readily integrated to build a strong gradient-based attack to further boost the success rates of adversarial examples for black-box attacks. Moreover, our method can also be naturally combined with other gradient-based attack methods to build a more robust attack to generate more transferable adversarial examples against the defense models. Extensive experiments on the ImageNet dataset demonstrate the method's effectiveness. Whether on adversarially trained networks or advanced defense models, our method has higher success rates than state-of-the-art gradient-based attack methods.

Bo Yang, Kaiyong Xu, Hengjun Wang et al.

Deep neural networks are vulnerable to adversarial examples, which are crafted by adding small, human-imperceptible perturbations to the original images, but make the model output inaccurate predictions. Before deep neural networks are deployed, adversarial attacks can thus be an important method to evaluate and select robust models in safety-critical applications. However, under the challenging black-box setting, the attack success rate, i.e., the transferability of adversarial examples, still needs to be improved. Based on image augmentation methods, we found that random transformation of image brightness can eliminate overfitting in the generation of adversarial examples and improve their transferability. To this end, we propose an adversarial example generation method based on this phenomenon, which can be integrated with Fast Gradient Sign Method (FGSM)-related methods to build a more robust gradient-based attack and generate adversarial examples with better transferability. Extensive experiments on the ImageNet dataset demonstrate the method's effectiveness. Whether on normally or adversarially trained networks, our method has a higher success rate for black-box attacks than other attack methods based on data augmentation. We hope that this method can help to evaluate and improve the robustness of models.