A Training Approach Of Robust Neural Network Guided By The Verification Of Misclassified Samples

As deep learning techniques continue to make progress in the field of computer vision,artificial intelligence systems are being widely used in scenarios such as autonomous driving and face recognition.The application of artificial intelligence under these major security domains has created an urgent need to understand the robustness of neural networks.It is well known that deep neural networks are not robust,and even imperceptible perturbations can cause neural networks to make incorrect predictions.However,the robustness improvement of most defense methods cannot be proven,and thus they are often broken by other types of attacks.To overcome these difficulties,many researchers have proposed the use of formal verification techniques for training verifiably robust neural networks.Given a range of perturbations,verifiably robust training methods can compute robust bounds by verification techniques and minimize them to train a verifiably robust model.However,the current training approach leads to a significant decrease in model accuracy.Therefore,how to handle misclassified samples and efficiently train deep neural networks with better accuracy and robustness is a pressing problem.By introducing verification techniques into training,we propose a training approach of robust neural network guided by the verification of misclassified samples.With the training set and maximum perturbation distance provided,we validate the under-trained network on the training set and add the verification results to the loss function to guide parameter updates.Also,we optimize the type of loss function used in training robust models by introducing an accelerated cross-entropy loss function based on increasing the robust bound,which is suitable for training verifiable robust neural network models.In addition,the method in this paper can be used in conjunction with existing robust neural network training tools.In order to evaluate the effectiveness of the method in this paper,we applied the training method in this paper to the latest robust neural network training methods of IBP and Crown-IBP,and compared them experimentally.Experiments show that our training method can effectively improve the accuracy and robustness of the model.