Common Knowledge Learning for Generating Transferable Adversarial Examples

TL;DR

This paper introduces a common knowledge learning framework that improves the transferability of black-box adversarial examples by distilling knowledge from multiple teacher models and constraining gradients to reduce output inconsistency.

Contribution

It proposes a multi-teacher knowledge distillation approach with gradient constraints to enhance adversarial transferability across different DNN architectures.

Findings

Significantly improves transferability of adversarial examples

Reduces output inconsistency between models

Enhances black-box attack effectiveness

Abstract

This paper focuses on an important type of black-box attacks, i.e., transfer-based adversarial attacks, where the adversary generates adversarial examples by a substitute (source) model and utilize them to attack an unseen target model, without knowing its information. Existing methods tend to give unsatisfactory adversarial transferability when the source and target models are from different types of DNN architectures (e.g. ResNet-18 and Swin Transformer). In this paper, we observe that the above phenomenon is induced by the output inconsistency problem. To alleviate this problem while effectively utilizing the existing DNN models, we propose a common knowledge learning (CKL) framework to learn better network weights to generate adversarial examples with better transferability, under fixed network architectures. Specifically, to reduce the model-specific features and obtain better output distributions, we construct a multi-teacher framework, where the knowledge is distilled from different teacher architectures into one student network. By considering that the gradient of input is usually utilized to generated adversarial examples, we impose constraints on the gradients between the student and teacher models, to further alleviate the output inconsistency problem and enhance the adversarial transferability. Extensive experiments demonstrate that our proposed work can significantly improve the adversarial transferability.