Rethinking the optimization objective for transferable adversarial examples from a fuzzy perspective

Open Webpage

Xiangyuan Yang, Jie Lin, Hanlin Zhang, Peng Zhao

Published: 01 Apr 2025, Last Modified: 12 Nov 2025Neural NetworksEveryoneRevisionsCC BY-SA 4.0
Abstract: Highlights•Fuzzy domain is proposed to characterize untransferable adversarial examples.•FOTA enhances the transferability of adversarial example by evading the fuzzy domain.•Adaptive FOTA further enhance the transferability with dynamic update iterations.•The increased transferability is as high as 16.1% on ImageNet.