Back to Search Start Over

Distribution-restrained Softmax Loss for the Model Robustness

Authors :
Wang, Hao
Li, Chen
Jiang, Jinzhe
Zhang, Xin
Zhao, Yaqian
Gong, Weifeng
Publication Year :
2023

Abstract

Recently, the robustness of deep learning models has received widespread attention, and various methods for improving model robustness have been proposed, including adversarial training, model architecture modification, design of loss functions, certified defenses, and so on. However, the principle of the robustness to attacks is still not fully understood, also the related research is still not sufficient. Here, we have identified a significant factor that affects the robustness of models: the distribution characteristics of softmax values for non-real label samples. We found that the results after an attack are highly correlated with the distribution characteristics, and thus we proposed a loss function to suppress the distribution diversity of softmax. A large number of experiments have shown that our method can improve robustness without significant time consumption.

Details

Database :
arXiv
Publication Type :
Report
Accession number :
edsarx.2303.12363
Document Type :
Working Paper