Abstract: While the interpretability of deep learning is developing,its security is also facing significant challenges.There is a risk that the interpretation results of the model on input data may be maliciously manipulated and attacked,which seriously affects the application scenarios of interpretability technology and hinders human exploration and cognition of the model.To address this issue,an interpretable robust adversarial training method using model gradients as similarity constraints is proposed.Firstly,adversarial training data is generated by sampling along the interpretation direction.Secondly,multiple similarity metrics between the interpretations of the sampled data are calculated by combining the gradient information of the samples during the training process,which is used to regularize the model and smooth its curvature.Finally,to verify the effectiveness of the proposed interpretable robust adversarial training method,it is validated on multiple datasets and interpretation methods.The experimental results show that the proposed method has a significant effect on defending against adversarial interpretation samples.

Key words: Deep learning, Interpretability, Adversarial attack, Adversarial training, Adversarial samples