Towards sustainable adversarial training with successive perturbation generation
Back to article page
Regular Papers|Updated:2024-04-29
|
Towards sustainable adversarial training with successive perturbation generation
Enhanced Publication
基于连续扰动生成方法的可持续对抗训练
“2024年4月28日,来自College of Computer Science and Mathematics的LIN Wei团队和College of Economics and Management的LIAO Lichuan团队,在知名期刊《Frontiers of Information Technology & Electronic Engineering》上发表了一篇关于提升神经网络模型鲁棒性的研究论文。他们发现,通过逐步增强训练过程中对抗样本的攻击力,可以有效提高模型的鲁棒性,同时,适当的模型迁移可以在几乎不增加计算成本的情况下保持模型的泛化性能。为了验证这一观点,他们提出了一种名为SPGAT的连续扰动生成对抗训练方法,该方法通过在前一个周期转移的对抗样本上添加扰动来逐步增强对抗样本的攻击力,并在周期之间迁移模型以提高对抗训练的效率。实验结果表明,SPGAT方法在计算效率上远超标准对抗训练方法,同时,在模型对抗精度和清洁精度上的性能提升分别超过7%和3%。他们在小规模的MNIST、中等规模的CIFAR-10以及大规模的CIFAR-100数据集上都进行了广泛的评估,结果显示SPGAT方法在效率和性能上均优于当前最先进的方法,为神经网络模型的鲁棒性提升提供了新的解决方案。”
”
Frontiers of Information Technology & Electronic EngineeringVol. 25, Issue 4, Pages: 527-539(2024)
Affiliations:
1.College of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350118, China
2.College of Economics and Management, Xi’an University of Technology, Xi’an 710048, China
3.Fujian Provincial Key Laboratory of Big Data Mining and Applications, Fujian University of Technology, Fuzhou 350118, China
Wei LIN, Lichuan LIAO. Towards sustainable adversarial training with successive perturbation generation. [J]. Frontiers of Information Technology & Electronic Engineering 25(4):527-539(2024)
DOI:
Wei LIN, Lichuan LIAO. Towards sustainable adversarial training with successive perturbation generation. [J]. Frontiers of Information Technology & Electronic Engineering 25(4):527-539(2024) DOI: 10.1631/FITEE.2300474.
Towards sustainable adversarial training with successive perturbation generationEnhanced Publication