计算机科学 ›› 2023, Vol. 50 ›› Issue (11A): 221000045-5.doi: 10.11896/jsjkx.221000045
郜玉钊, 邢云汉, 刘嘉祥
GAO Yuzhao, XING Yunhan, LIU Jiaxiang
摘要： 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。
中图分类号:
| [1]HINTON G,DENG L,YU D,et al.Deep neural networks foracoustic modeling in speech recognition:The shared views of four research groups[J].IEEE Signal Processing Magazine,2012,29(6):82-97. [2]KRIZHEVSKY A,SUTSKEVER I,HINTON G E.Imagenetclassification with deep convolutional neural networks[J].Communications of the ACM,2017,60(6):84-90. [3]LUBOSCH M,KUNATH M,WINKLER H.Industrial sche-duling with Monte Carlo tree search and machine learning[C]//Procedia CIRP.2018:1283-1287. [4]HUANG X,KROENING D,RUAN W,et al.A survey of safetyand trustworthiness of deep neural networks:Verification,testing,adversarial attack and defence,and interpretability[J].Computer Science Review,2020,37:100270. [5]ZHANG M H,DU D H,ZHANG M Z,et al.Spatio-temporaltrajectory data-driven autonomous driving scenario meta-mode-ling approach[J].Ruan Jian Xue Bao/Journal of Software,2021,32(4):973-987. [6]KALRA N,PADDOCK S M.Driving to safety:How many miles of driving would it take to demonstrate autonomous vehicle reliab-ility?[J].Transportation Research Part A:Policy and Practice,2016,94:182-193. [7]WANG Z,YAN M,LIU S,et al.Survey on testing of deep neural networks[J].Ruan Jian Xue Bao/Journal of Software,2020,31(5):1255-1275. [8]WANG J,ZHAN N J,FENG X Y,et al.Overview of formalmethods[J].Ruan Jian Xue Bao/Journal of Software,2019,30(1):33-61. [9]LIU C,ARNON T,LAZARUS C,et al.Algorithms for verifying deep neural networks[J].Foundations and Trends© in Optimization,2021,4(3/4):244-404. [10]PULINA L,TACCHELLA A.An abstraction-refinement ap-proach to verification of artificial neural networks[C]//International Conference on Computer Aided Verification.Berlin:Springer,2010:243-257. [11]XIANG W,TRAN H D,JOHNSON T T.Reachable set computation and safety verification for neural networks with relu activations[J].arXiv:1712.08163,2017. [12]XIANG W,TRAN H D,JOHNSON T T.Output reachable set estimation and verification for multilayer neural networks[J].IEEE Transactions on Neural Networks and Learning Systems,2018,29(11):5777-5783. [13]SINGH G,GEHR T,MIRMAN M,et al.Fast and effective robustness certification[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems(NIPS’18).2018:10825-10836. [14]SINGH G,GEHR T,PÜSCHEL M,et al.An abstract domain for certifying neural networks [C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems(NIPS ’18).2018:10825-10836. [15]DVIJOTHAM K,STANFORTH R,GOWAL S,et al.A Dual Approach to Scalable Verification of Deep Networks[C]//UAI.2018:3. [16]RAGHUNATHAN A,STEINHARDT J,LIANG P.Certified defenses against adversarial examples[J].arXiv:1801.09344,2018. [17]ASHOK P,HASHEMI V,KŘETÍNSKÝ J,et al.Deepab-stract:Neural network abstraction for accelerating verification[C]//International Symposium on Automated Technology for Verification and Analysis.Cham:Springer,2020:92-107. [18]GEHR T,MIRMAN M,DRACHSLER-COHEN D,et al.Ai2:Safety and robustness certification of neural networks with abstract interpretation[C]//2018 IEEE Symposium on Security and Privacy(SP).IEEE,2018:3-18. [19]TRAN H D,MANZANAS LOPEZ D,MUSAU P,et al.Star-based reachability analysis of deep neural networks[C]//International Symposium on Formal Methods.Cham:Springer,2019:670-686. [20]ZHANG H,WENG T W,CHEN P Y,et al.Efficient neural network robustness certification with general activation functions[J/OL].Advances in Neural Information Processing Systems,2018,31.https://dl.acm.org/doi/abs/10.5555/3327345.3327402. [21]XU K,ZHANG H,WANG S,et al.Fast and complete:Enabling complete neural network verification with rapid and massively parallel incomplete verifiers[J].arXiv:2011.13824,2020. [22]WANG S,ZHANG H,XU K,et al.Beta-crown:Efficient bound propagation with per-neuron split constraints for neural network robustness verification[J].Advances in Neural Information Processing Systems,2021,34:29909-29921. [23]DE PALMA A,BUNEL R,DESMAISON A,et al.Improvedbranch and bound for neural network verification via lagrangian decomposition[J].arXiv:2104.06718,2021. [24]BUNEL R,MUDIGONDA P,TURKASLAN I,et al.Branchand bound for piecewise linear neural network verification[J/OL].Journal of Machine Learning Research,2020,21(2020).https://dl.acm.org/doi/10.5555/3455716.3455758. [25]LI J,LIU J,YANG P,et al.Analyzing deep neural networks with symbolic propagation:Towards higher precision and faster verification[C]//International Static Analysis Symposium.Cham:Springer,2019:296-319. [26]LIN W,YANG Z,CHEN X,et al.Robustness verification ofclassification deep neural networks via linear programming[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2019:11418-11427. [27]YANG P,LI R,LI J,et al.Improving neural network verification through spurious region guided refinement[C]//International Conference on Tools and Algorithms for the Construction and Analysis of Systems.Cham:Springer,2021:389-408. |
|
||
首页