DENAS: An Input-independent DL Model Interpretation Framework for Security Applications
In our recent project, we propose an input-independent deep learning interpretation framework -DENAS. We find an intrinsic property of the neural networks and this property could model the decision boundary of the neural networks without a specific input.
A Study for Testing Oracle for DL Models
We study the performance of different testing oracles for deep learning models, the study results are shown on our website. Based on our study results, we propose an approach to reduce model redundancy for different testing.
Defend against the Adversarial Samples by Attacking the Adversarial Samples