Improving Adversarial Transferability on Vision Transformers via Dual-Flow Adversarial Attack
TL;DR
This paper introduces a dual-flow adversarial attack method to enhance transferability on Vision Transformers by addressing limitations of fixed-decay EMA in smoothing token embeddings.
Improving Adversarial Transferability on Vision Transformers via Dual-Flow Adversarial Attack
Bohai Zhou; Anlu Shi; Wei Wu
https://doi.org/10.1109/ACCESS.2025.3649234
Volume 14
State-of-the-art transferable adversarial attack methods for Vision Transformers (ViTs) have introduced token-embedding momentum inside the Transformer blocks, but they typically rely on fixed-decay exponential moving averages (EMA) to smooth historical embeddings. While EMA can suppress noise, it ignores the varying usefulness of historical information across time steps; moreover, cross-architect...