Boosting targeted adversarial transferability via fine-grained feature mixup perturbation and reference-based gradient refinement

Xining Gao, Sen Huangfu, Ying Zhao, Zijiao Zhang · 2026

Read Boosting targeted adversarial transferability via fine-grained feature mixup perturbation and reference-based gradient refinement by Xining Gao, Sen Huangfu, Ying Zhao, Zijiao Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13400606.txt.

Description

Xining Gao, Sen Huangfu, Ying Zhao and Zijiao Zhang propose, in 2026, a stronger way to craft adversarial images that fool deep neural networks into a chosen wrong answer. Such transferable attacks are a serious security concern because they work even against models the attacker cannot see inside. Prior methods mixed clean features too crudely.

Their method splits feature maps into blocks and mixes each with clean features shuffled from other positions, creating richer perturbations. A second step refines gradients by amplifying how they differ from a reference, steering away from quirks of the stand-in model. On two standard image datasets, the approach beat leading rivals against both convolutional and transformer models.

Catalog record

CREATORXining Gao, Sen Huangfu, Ying Zhao, Zijiao Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57157
LICENSEhttps://creativecommons.org/licenses/by/4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/article/MED/PMC13400606

Browse the full library · About the Tower of Babel