一种约束验证神经网络的方法

Constraint-based Verification Method for Neural Networks

神经网络的验证一直是人工智能领域的主要挑战之一。文中基于DeepZ方法,提出一种通过约束提升深度神经网络的局部鲁棒性验证精度的方法。在传播过程中加入约束来缩小抽象域,通过线性规划求解一个更小的神经网络输出范围,以此推断出神经网络输出节点的新的边界。应用新的边界,可以得出更精确的验证结果。基于此方法,实现了DeepZero工具,并在MNIST数据集上进行了充分的实验。实验结果表明,所提方法能有效提升DeepZ方法的验证成功率。在实验中,验证成功率平均可提升49%,说明了所提方法的有效性。

domain,then a tighter neural network output range is solved by linear programming,hence deducing new bounds of neural network output node.With the new bounds,more accurate verification results can be obtained.Based on this method,the DeepZero tool is implemented in this paper,and comprehensive evaluation is carried out on the MNIST dataset.Experimental results show that our method effectively improves the verification success rate of DeepZ.Specifically,the verification success rate of DeepZ increases by 49%in average,indicating the effectiveness of the proposed method.