| With the emergence of deep learning technology,the application of artificial intelligence has experienced explosive growth,and it has been rapidly applied to various industries such as medical care,finance,and autonomous driving.However,existing studies have found that deep learning models are vulnerable to adversarial sample attacks.Adversarial samples attacks refer to the addition of perturbations that are difficult to detect by humans in the input of the model,which makes the model make wrong decisions.This discovery exposed the vulnerability of deep learning models,greatly affected the application of artificial intelligence technology in fields with high security requirements,and hindered the development of artificial intelligence technology.Since the discovery of adversarial examples,many scholars have invested in adversarial example defense methods,hoping to improve the robustness of deep learning models.Most of the existing research focuses on improving the robustness of the model from the aspects of model input and model parameters,such as image transformation and compression,adversarial training,regularization training,etc.The model structure is also a key factor affecting the robustness of the model.At present,the research on the influence of model structure on model robustness is still in the preliminary stage.In view of the above situation,this paper firstly proposes a robust network construction technology based on neural architecture search.The neural architecture search technology is used to automatically build a network structure with high robustness;the problem of expensive evaluation in the search process is solved by the proxy model;the problem of robustness evaluation of the model is solved by the CLEVER score which is attack-agnostic.The experimental results show that the model constructed by the proposed method is better than the artificially designed model,and the robustness of the model is higher than that of other search algorithms when the model accuracy is comparable.Secondly,this paper also compares common model robustness improvement methods,such as total variance minimization defense,adversarial training,Jacobian regularization,etc.Through experimental comparison,the effectiveness of these methods is proved,which provides theoretical support for the construction of deep learning defense system.Finally,based on the above results,this paper designs and implements a deep learning defense system from three aspects:model input,model parameters and model structure.The system can improve the robustness of deep learning models from model input(adversarial samples),model design to model training.The system can be applied to improve the robustness of models and significantly improve the security of current artificial intelligence systems. |
