使用单步引导扩散模型进行对抗性净化
Yanchun Li1, Zemin Li2, Long Huang2
1School of Computer Science, Xiangtan University, Xiangtan, 411105, China; Hunan International Scientific and Technological Cooperation Base of Intelligent network, Xiangtan, 411105, China; Key Laboratory of Hunan Province for Internet of Things and Information Security, Xiangtan, 411105, China.
概括
我们推出了一步引导扩散模型 (OSGD),以实现高效的对抗性净化. 这种方法加快了采样速度,并提高了防御对抗性示例的准确性.
科学领域:
- 计算机视觉 计算机视觉
- 机器学习 机器学习
- 人工智能的人工智能
背景情况:
- 扩散模型显示,由于它们的概括能力,对抗性净化很有希望.
- 目前的扩散模型需要多步采样,导致大量的时间和资源消耗.
- 有效的防御机制对于强大的机器学习模型对抗敌对攻击至关重要.
研究的目的:
- 提出一种新的一步引导扩散模型 (OSGD),以实现高效的对抗性净化.
- 为了加快对抗性净化扩散模型的采样速度.
- 为了实现卓越的净化效果,并加强对抗对手的防御.
主要方法:
- 开发了一种单步指导扩散模型 (OSGD),采用单步排毒过程.
- 实施了一项指导策略,使用第一轮撤销的初步撤销图像.
- 应用了两轮单步解,第二轮以第一轮的输出为指导.
主要成果:
- 与传统的多步扩散模型相比,OSGD显著加快了采样速度.
- 建议的指导策略有效地消除了对抗性干扰,改善了净化结果.
- 在对抗PGD,AutoAttack和BPDA+EOT攻击的Cifar10和ImageNet数据集上的精度和效率方面实现了最先进的性能.
结论:
- OSGD为对抗性净化提供了一种高效有效的解决方案.
- 一步指导的无色化方法证明了初步无色化图像作为指导信号的可行性.
- 该方法为各种复杂的对抗性攻击策略提供了强有力的防御.
相关概念视频
Dialysis
812
Dialysis is a diffusion-based purification process that separates analyte molecules from a complex matrix. This is accomplished by allowing molecules in the solution to pass through a semipermeable membrane into a liquid on the other side. The membrane is usually made of cellulose acetate or cellulose nitrate, and the second liquid must be miscible with the solution. Ions (e.g., chloride or sodium) or organic molecules (e.g., glucose) can pass through the membrane pores, which generally have...
812
Affinity Chromatography
1.0K
Affinity chromatography is a powerful technique extensively utilized for separating and purifying specific biomolecules from complex mixtures. It capitalizes on the highly selective binding between an analyte and its counterpart, such as antibody-antigen interactions. The counterpart is immobilized on the stationary phase, forming an affinity column. The stationary phase typically consists of solid support, such as agarose or porous glass beads, immobilizing the affinity ligand. The mobile...
1.0K


