多域主动防御:通过ALL-to-ALL脱训练检测多域后门中毒样本,而无需清洁数据集
Binhao Ma1, Jiahui Wang1, Dejun Wang1
1School of Computer Science, South-Central Min Zu University, Wuhan 430074, China.
概括
这项研究引入了一种针对深度学习后门攻击的新防御方法,生成跨域的清洁数据以检测恶意模型而不需要原始清洁样本. 这种方法提高了多域数据集的安全性.
科学领域:
- 人工智能的人工智能
- 机器学习安全 机器学习安全
- 深度学习 (Deep Learning) 是一种深度学习.
背景情况:
- 深度学习模型容易受到后门中毒攻击,攻击者通过中毒的训练数据插入隐藏的漏洞.
- 现有的检测方法通常依赖于数据可分离性假设,而这些假设被适应性中毒策略所破坏.
- 目前的防御对多域数据集是不切实际的,并且由于需要清洁样本,引发了隐私问题.
研究的目的:
- 在深度学习中开发一个强大的,保护隐私的防御,防止后门中毒攻击.
- 创建一个实用的解决方案来检测多域数据集的后门,而不需要清洁的样本.
- 克服现有的检测方法的局限性,这些检测方法在适应性中毒策略下失败.
主要方法:
- 提出了一种多域主动防御方法,可以从各种领域生成多样化的清洁样本.
- 使用生成的清洁样本实现了神经网络的循环解.
- 分离的特征和标签使后门中毒样本更容易检测,而无需合适的清洁数据.
主要成果:
- 提议的防御有效地检测后门中毒样本.
- 该方法在多个不同的数据集中展示了实用性和有效性,包括CIFAR10,CelebA,MNIST,MNIST-M,USPS,SVHN和Tiny-ImageNet.
- 成功地解决了先前方法在适应性中毒和多领域适用性方面的局限性.
结论:
- 开发的多域主动防御在保护深度学习模型免受复杂的后门攻击方面取得了重大进展.
- 该方法为涉及各种数据的现实应用提供了一个实用的,对隐私有意识的解决方案.
- 未来的工作可以在各种深度学习架构中探索进一步的优化和应用.
相关概念视频
01:04¹³C NMR: ¹H–¹³C Decoupling
1.1K
The probability of having two carbon-13 atoms next to each other is negligible because of the low natural abundance of carbon-13. Consequently, peak splitting due to carbon-carbon spin-spin coupling is not observed in spectra. However, protons up to three sigma bonds away split the carbon signal according to the n+1 rule, resulting in complicated spectra.
A broadband decoupling technique is used to simplify these complex, sometimes overlapping, signals. Broadband decoupling relies on a...
A broadband decoupling technique is used to simplify these complex, sometimes overlapping, signals. Broadband decoupling relies on a...
1.1K
02:26Conservation of Protein Domains Over Different Proteins
10.9K
Protein domains are small structurally independent units that are part of a single amino acid chain. Although these domains are often structurally independent, they may rely on synergistic effects to perform their functions as part of a larger protein. Protein domains may be conserved within the same organism, as well as across different organisms.
A limited set of protein domains often duplicate and recombine during evolution. These domains can be organized in different combinations to...
A limited set of protein domains often duplicate and recombine during evolution. These domains can be organized in different combinations to...
10.9K
01:19Masking and Demasking Agents
2.5K
EDTA titrations may necessitate masking and demasking agents to temporarily protect a particular metal ion in a mixture from the EDTA reaction. These agents facilitate the sequential analysis of the metal ions by forming stable complexes with some—but not all—metal ions during certain steps.
There are many masking agents, such as cyanide, fluoride, triethanolamine, thiourea, and 2,3-bis(sulfanyl)propan-1-ol (formerly 2,3-dimercapto-1-propanol), with the masking agent chosen based on...
There are many masking agents, such as cyanide, fluoride, triethanolamine, thiourea, and 2,3-bis(sulfanyl)propan-1-ol (formerly 2,3-dimercapto-1-propanol), with the masking agent chosen based on...
2.5K
01:18Effects of EDTA on End-Point Detection Methods
286
Different methods, such as visual observance of metal-ion indicators, spectroscopic techniques, and potentiometric methods, can determine the endpoint of an EDTA titration.
In the visual method, metal-ion indicators (metallochromic dyes), which have distinct colors in their free and complex forms, are added to the mixture to signal the titration's end point. They form stable complexes with metal ions, but these complexes are weaker than the corresponding metal–EDTA complexes. As a...
In the visual method, metal-ion indicators (metallochromic dyes), which have distinct colors in their free and complex forms, are added to the mixture to signal the titration's end point. They form stable complexes with metal ions, but these complexes are weaker than the corresponding metal–EDTA complexes. As a...
286
01:20¹³C NMR: Distortionless Enhancement by Polarization Transfer (DEPT)
1.1K
When proton-coupled carbon-13 spectra are simplified by a broadband proton decoupling technique, structural information about the coupled protons is lost. Distortionless enhancement by polarization transfer (DEPT) is a technique that provides information on the number of hydrogens attached to each carbon in a molecule. While the DEPT experiment utilizes complex pulse sequences, the pulse delay and flip angle are specifically manipulated. The resulting signals have different phases depending on...
1.1K
01:05Difference from Background: Limit of Detection
6.4K
The limit of detection (LOD) is the smallest amount of analyte that can be distinguished from the background noise. The LOD value corresponds to the concentration at which the analyte signal is three times larger than the standard deviation of the blank signal. Below this value, the analyte signal cannot be differentiated from the background noise. It is calculated by dividing the calibration slope by 3 times the standard deviation of the blank signals.
The LOD indicates the presence or absence...
The LOD indicates the presence or absence...
6.4K