Deep Learning-Based Robustness Verification Of Neural Networks

In recent years,with the rapid development of artificial intelligence,the breakthrough of deep neural networks has promoted their applications in safety-related fields such as autopilot,aircraft control,biometrics and so on.However,the robustness of the deep neural network is not verified,which leads to the potential security threat in the practical application.Therefore,it is an urgent issue to verify whether the deep neural network is robust or not.Currently,the research on robustness verification of deep neural networks is still challenging.On the one hand,the complexity of network robustness verification rises with the increasingly size of the network,and the verification process takes a long time.On the other hand,different network structures put forward higher requirements for the scalability of verification methods.This is becoming more difficult today when facing with large-scale neural networks.Therefore,the robustness verification of deep neural networks needs to take into account the efficiency and scalability.This paper aims at the problems and challenges of robustness verification of deep neural networks,and proposes a verification method based on deep learning,which improves the verification efficiency on the premise of ensuring the verification accuracy,and can adapt to multiple network structures.The main work of this paper is as follows:· The paper first proposes a deep learning-based bounding verification method to tighten the bounds of the robustness verification problem,which constructs a graph convolution network(GCN)according to the structure of the network to be verified and the corresponding MILP problem.GCN-based bounding method learns and optimizes the bounding parameters from the data through multiple rounds of forward and backward propagation to obtain tightening bounding results,so as the robustness of the network can be verified more efficiently.Experiments show that GCN-based bounding method can tighten the bounds and improves the efficiency of robustness verification.In addition,the GCN-based bounding method is also available to networks of different size,showing good scalability.· Furthermore,a branching verification method based on deep learning is proposed to gradually repair the deviation caused by relaxation of bounding.The method adopts an iterative scheme,utilizes the characteristics of recurrent neural network(RNN),and formulates a branching strategy to guide the selection of neurons through datadriven manner,so as to iteratively reduce the deviation of relaxation until the property for network robustness verification is satisfied.RNN-based branching method improves the efficiency of verification on the premise of ensuring the verification accuracy.The experimental results show that the RNN-based branching method is scalable and has better performance on robustness verification compared with other branching methods.· Based on the above two methods,the paper further extends and proposes a branchand-bound verification framework based on deep learning.The framework combines GCN-based bounding method and RNN-based branching method,that is,it using GCN for bounding and RNN for branching to form a complete branch-andbound method based on deep learning for robustness verification of neural network.The method of GCN+RNN further improving the effectiveness and efficiency of the robustness verification of the deep neural network.Through experiments and evaluations,the method of GCN+RNN is ahead of other mainstream verification methods in most cases with high performance.