相互GNN-MLP蒸用于强大的图形对抗防御
Bowen Deng1, Jialong Chen2, Yanming Hu2
1School of Computer Science and Engineering, Sun Yat-sen University, No. 132, Outer Ring East Road, Guangzhou, 510006, Guangdong, China; School of Systems Science and Engineering, Sun Yat-sen University, No. 135, Xingang West Road, Guangzhou, 510275, Guangdong, China.
概括
本研究介绍了相互GNN-MLP蒸 (MGMD),以改善图形神经网络 (GNN) 的对抗防御. MGMD增强了对图形异构的适应性,并提供可扩展的推理,克服了当前GNN防御方法的局限性.
科学领域:
- 人工智能的人工智能
- 机器学习 机器学习
- 图形神经网络的神经网络
背景情况:
- 目前对图形神经网络 (GNN) 的对抗防御在适应图形异构性,概括到早期的GNN模型 (如GraphSAGE) 和实现可扩展推理方面存在局限性.
- 这些局限性限制了GNN在资源有限的环境和对复杂的对抗性攻击的实际应用.
研究的目的:
- 提出一种新的框架,即相互GNN-MLP蒸 (MGMD),以解决现有的GNN对抗防御的局限性.
- 增强GNN对图形异构的适应性,并提高对结构和节点特征攻击的稳定性.
- 为GNN实现高推断可扩展性,使其适合资源有限的场景.
主要方法:
- 开发了GNN-MLP相互蒸 (MGMD) 框架,该框架结合了GNN和多层感知子 (MLP) 的优势.
- 集成的GNN和MLP,以提高对图形异构的适应性,并防御对抗性攻击.
- 引入了一种新的学习率调度器,灵感来自于融合分析,以减轻GNN和MLP组件之间的诱导偏差冲突.
- 通过决策边界分析,正式证明了MGMD的对抗性强度和适应性.
主要成果:
- 与以前的方法相比,MGMD显示了对图形异构的增强适应性和更好的对抗性稳定性.
- 蒸的MLP组件实现了显著高的推理可扩展性.
- 对各种同型和异型图的实验验验证了拟议的学习速率调度器的有效性.
- MGMD显示出明显的优势,而不是现有的对抗性防御方法GNNs.
结论:
- 拟议的互惠GNN-MLP蒸 (MGMD) 框架有效地解决了目前GNN对抗防御的关键局限性.
- MGMD为GNN提供了可扩展和强大的解决方案,特别是在异性图设置和资源受限制的环境中.
- 新型学习率调度器有助于稳定的培训和改进混合GNN-MLP模型的性能.
相关概念视频
01:23Protecting Groups for Aldehydes and Ketones: Introduction
6.5K
Protecting groups are compounds that can bind to a specific functional group in the presence of other functional groups to protect them from undesired chemical reactions. These compounds can selectively bind to particular functional groups and advance chemoselective reactions in polyfunctional systems (Figure 1). After the functional group has served its purpose, it is removed by reacting it with specific compounds.
6.5K
01:25Radical Anti-Markovnikov Addition to Alkenes: Overview
3.2K
The addition of hydrogen bromide to alkenes in the presence of hydroperoxides or peroxides proceeds via an anti-Markovnikov pathway and yields alkyl bromides.
3.2K
02:27Predicting Molecular Geometry
34.0K
VSEPR Theory for Determination of Electron Pair Geometries
34.0K
02:18Network Covalent Solids
13.3K
Network covalent solids contain a three-dimensional network of covalently bonded atoms as found in the crystal structures of nonmetals like diamond, graphite, silicon, and some covalent compounds, such as silicon dioxide (sand) and silicon carbide (carborundum, the abrasive on sandpaper). Many minerals have networks of covalent bonds.
To break or to melt a covalent network solid, covalent bonds must be broken. Because covalent bonds are relatively strong, covalent network solids are typically...
To break or to melt a covalent network solid, covalent bonds must be broken. Because covalent bonds are relatively strong, covalent network solids are typically...
13.3K
01:32Radical Anti-Markovnikov Addition to Alkenes: Thermodynamics
1.8K
The anti-Markovnikov addition of hydrogen halides to an alkene is thermodynamically feasible only with HBr. The radical addition reaction with other hydrogen halides like HCl and HI is thermodynamically unfavorable.
1.8K
01:13¹H NMR: Complex Splitting
1.2K
A proton M that is coupled to a proton X results in doublet signals for M. However, NMR-active nuclei can be simultaneously coupled to more than one nonequivalent nucleus. When M is coupled to a second proton A, such as in styrene oxide, each peak in the doublet is split into another doublet.
Splitting diagrams or splitting tree diagrams are routinely used to depict such complex couplings. While drawing splitting diagrams, the splitting with the larger coupling constant is usually applied...
Splitting diagrams or splitting tree diagrams are routinely used to depict such complex couplings. While drawing splitting diagrams, the splitting with the larger coupling constant is usually applied...
1.2K