从频域角度防御隐形后门攻击
Defending Against Invisible Backdoor Attacks from the Aspect of Frequency Domain
马俊智 1丁建军 2苏通 1朱勇杰 1孙超2
作者信息
- 1. 江汉大学智能制造学院 武汉 430056
- 2. 江汉大学智能制造学院 武汉 430056;江汉大学精细爆破国家重点实验室 武汉 430056
- 折叠
摘要
使用第三方模型进行图像识别时,存在受到隐形后门攻击的威胁,而目前的防御研究集中于像素域与训练阶段.针对这种情况,提出基于高斯滤波的后门防御方法.该方法从频域角度切入并在预处理阶段进行防御,首先通过对隐形后门攻击进行频域特性研究,分析出触发器位于图像高频信息区域,然后对于不同种类的攻击,在模型预测阶段引入高斯滤波器进行防御实验.实验结果表明,该方法将隐形后门攻击的攻击成功率降低至10%以内.
Abstract
When using third-party models for image recognition,there is a threat of invisible backdoor attacks,and the cur-rent defense research is focused on the pixel domain and training phase.In view of this,a backdoor defense method based on Gauss-ian filtering is proposed.The method starts from the frequency domain and defends in the pre-processing stage.Firstly,the invisible backdoor attack is studied in the frequency domain,and the trigger is analyzed to be located in the high frequency information re-gion of the image.Experimental results show that the method reduces the attack success rate of invisible backdoor attack to within 10%.
关键词
后门防御/隐形后门攻击/频域/图像识别Key words
backdoor defenses/invisible backdoor attacks/frequency domain/image recognition引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据