Feature Smoothing for Generating Transferable Adversarial Examples

Junlin Liu

doi:10.1109/tetci.2026.3690052

Feature Smoothing for Generating Transferable Adversarial Examples

2026-05-12

0

PRE

AI

DOI：10.1109/tetci.2026.3690052

Original

Original request for help

Save

Abstract

En 中文

Transfer-based adversarial attacks generate transferable adversarial examples from a surrogate model to deceive unknown target models. Input diversity is an effective technique to enhance the adversarial transferability by shifting the attention regions of inputs. However, current existing input-diversity-based attacks primarily focus on the input space, neglecting the impact of high-frequency components in the feature space. This limits the adversarial transferability because adversarial perturbations primarily concentrate on high-frequency components. In this paper, we propose a feature orchestration (FO) layer to smooth the extracted embeddings via trained low-pass filters. FO filters out the high-frequency components in the feature space, alleviating the overfitting of adversarial noise to the surrogate model. Based on the proposed FO, we propose a novel feature orchestration attack (FOA). Specifically, FOA apply various transformations and mixups to generate diverse variants in the input space, and smooths the extracted embeddings by FO in the feature space. Evaluations on ImageNet and Caltech101 underscore the superior attack performance of the proposed FOA, compared to the state-of-the-art (SOTA) baselines. In particular, the targeted success rates of FOA outperform the SOTA baselines by about 7%$\sim$65% on ImageNet and about 3%$\sim$33% on Caltech101. Even against defended models, FOA achieves an average targeted success rate of over 53%, exceeding baselines by more than 40%.

Keywords:

Adversarial examples

transferability

input diversity

feature diversity

Feature Smoothing for Generating Transferable Adversarial Examples

Abstract

Journal

Organization

Cited Papers