A Robustness-Oriented Data Augmentation Method for DNN
Meixi Liu, Weijiang Hong, Weiyu Pan, Chendong Feng · 2021 IEEE 21st International Conference on Software Quality, Reliability and Security Companion (QRS-C) · 2021
The application of deep neural networks (DNN) is extensive and has achieved great success in many fields. However, the security and reliability are two great challenges for the system based on DNN. The robustness is an important attribute to measure the security and reliability of DNN models. We observe that further training the model through the small variation of the training dataset can improve the robustness while maintaining the accuracy on the testing dataset. Based on this observation, we propose a general DNN training framework, STYX,conducting on the alternative training dataset. We have carried out experiments on several commonly used datasets, whose results show that STYX can improve the model's robustness and keep its accuracy on the testing dataset as much as possible.