Skip to content
InfoResearchPreprint

SmoothShift: Disrupting Visual Adversarial Attacks on LVLMs at Test Time

Published
Record updated
View JSON

Summary

SmoothShift is a lightweight, training-free defense against visual adversarial attacks on large vision-language models (LVLMs). It combines local smoothing, which weakens fine-grained perturbation patterns, with spatial transformations that disrupt their alignment with the encoder's patch grid. The authors report improved robustness across three LVLM families, four datasets and eight attack methods, with clean performance largely preserved.

Mitigation

SmoothShift: a training-free defense combining local smoothing and spatial transformations. It requires no model updates, auxiliary networks, or additional model inference.