一种跨域的视频时空攻击方法
Cross-Domain Video Spatiotemporal Attack Method
刘占鹏 1王元斌 2周倜2
作者信息
- 1. 武汉数字工程研究所,武汉 430200;厦门大学,厦门 361100
- 2. 武汉数字工程研究所,武汉 430200
- 折叠
摘要
当前,深度学习在视频识别领域被广泛应用.然而,在实际应用中,深度神经网络面临着对抗性攻击的威胁.目前研究多集中于图像模型的攻击方法,对视频识别模型的攻防研究尚不充分.提出了一种多层时空攻击方法.该方法通过使用图像模型来完成对视频识别模型的攻击,通过多层特征融合与视频帧之间的对抗交互,从空间和时间层面对视频生成对抗性扰动.多层时空攻击方法由空间攻击模块和时间攻击模块组成.此外,为进一步提升攻击的有效性和鲁棒性,引入了自适应噪声和自适应损失权重机制.在UCF101和Kinetics-400数据集上进行了广泛的试验验证,试验结果表明,与现有技术相比,该方法的攻击成功率显著提升.
Abstract
Currently,deep learning is widely applied in the field of video recognition.However,in practical applications,deep neural networks are faced with the threat of adversarial attacks.Existing studies primarily focus on attack methods for image recognition models,but studies on the attack and defense of video recognition models are still insufficient.This paper proposes a multi-layer spatiotemporal attack method,which uses image model to attack video recognition model,and generates adversarial dis-turbance to video from space and time through multi-layer feature fusion and antagonistic interaction between video frames.The multi-layer spatiotemporal attack method is composed of a spatial attack module and a temporal attack module.In addition,in or-der to further improve the effectiveness and robustness of the attack,adaptive noise and adaptive loss weight mechanisms were in-troduced.Extensive experimental verification was conducted on the UCF101 and Kinetics-400 datasets,and the results showed that the success rate of attacks using this method was significantly improved compared to existing technologies.
关键词
黑盒攻击/视频识别/跨域攻击Key words
black-box attack/video recognition/cross-domain attack引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据