English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The wide application of deep neural networks (DNNs) demands an increasingamount of attention to their real-world robustness, i.e., whether a DNN resistsblack-box adversarial attacks, among which score-based query attacks (SQAs) aremost threatening since they can effectively hurt a victim network with the onlyaccess to model outputs. Defending against SQAs requires a slight but artfulvariation of outputs due to the service purpose for users, who share the sameoutput information with SQAs. In this paper, we propose a real-world defense byUnifying Gradients (UniG) of different data so that SQAs could only probe amuch weaker attack direction that is similar for different samples. Since suchuniversal attack perturbations have been validated as less aggressive than theinput-specific perturbations, UniG protects real-world DNNs by indicatingattackers a twisted and less informative attack direction. We implement UniGefficiently by a Hadamard product module which is plug-and-play. According toextensive experiments on 5 SQAs, 2 adaptive attacks and 7 defense baselines,UniG significantly improves real-world robustness without hurting cleanaccuracy on CIFAR10 and ImageNet. For instance, UniG maintains a model of77.80% accuracy under 2500-query Square attack while the state-of-the-artadversarially-trained model only has 67.34% on CIFAR10. Simultaneously, UniGoutperforms all compared baselines in terms of clean accuracy and achieves thesmallest modification of the model output. The code is released athttps://github.com/snowien/UniG-pytorch.

Unifying Gradients to Improve Real-world Robustness for Deep Networks