Parallelizable Reachability Analysis Algorithms for Feed-Forward Neural Networks

被引:30
作者
Tran, Hoang-Dung [1 ]
Musau, Patrick [1 ]
Lopez, Diego Manzanas [1 ]
Yang, Xiaodong [1 ]
Nguyen, Luan Viet [2 ]
Xiang, Weiming [1 ]
Johnson, Taylor T. [1 ]
机构
[1] Vanderbilt Univ, 221 Kirkland Hall, Nashville, TN 37235 USA
[2] Univ Penn, Philadelphia, PA 19104 USA
来源
2019 IEEE/ACM 7TH INTERNATIONAL WORKSHOP ON FORMAL METHODS IN SOFTWARE ENGINEERING (FORMALISE 2019) | 2019年
基金
美国国家科学基金会;
关键词
D O I
10.1109/FormaliSE.2019.00012
中图分类号
TP31 [计算机软件];
学科分类号
081202 ; 0835 ;
摘要
Artificial neural networks (ANN) have displayed considerable utility in a wide range of applications such as image processing, character and pattern recognition, self-driving cars, evolutionary robotics, and non-linear system identification and control. While ANNs are able to carry out complicated tasks efficiently, they are susceptible to unpre-dictable and errant behavior due to irregularities that emanate from their complex non-linear structure. As a result, there have been reservations about incorporating them into safety-critical systems. In this paper, we present a reachability analysis method for feed-forward neural networks (FNN) that employ rectified linear units (ReLUs) as activation functions. The crux of our approach relies on three reachable-set computation algorithms, namely exact schemes, lazy-approximate schemes, and mixing schemes. The exact scheme computes an exact reachable set for FNN, while the lazy-approximate and mixing schemes generate an over-approximation of the exact reachable set. All schemes are designed efficiently to run on parallel platforms to reduce the computation time and enhance the scalability. Our methods are implemented in a MATLAB (R) toolbox called, NNV, and is evaluated using a set of benchmarks that consist of realistic neural networks with sizes that range from tens to a thousand neurons. Notably, NNV successfully computes and visualizes the exact reachable sets of the real world ACAS Xu deep neural networks (DNNs), which are a variant of a family of novel airborne collision detection systems known as the ACAS System X, using a representation of tens to hundreds of polyhedra.
引用
收藏
页码:31 / 40
页数:10
相关论文
共 17 条
  • [1] [Anonymous], 2016, ARXIV PREPRINT ARXIV
  • [2] [Anonymous], 2005, MNIST DATABASE HANDW
  • [3] [Anonymous], ARXIV170909130
  • [4] Bradley P. S., 1998, Proceedings Fourth International Conference on Knowledge Discovery and Data Mining, P9
  • [5] AI2: Safety and Robustness Certification of Neural Networks with Abstract Interpretation
    Gehr, Timon
    Mirman, Matthew
    Drachsler-Cohen, Dana
    Tsankov, Petar
    Chaudhuri, Swarat
    Vechev, Martin
    [J]. 2018 IEEE SYMPOSIUM ON SECURITY AND PRIVACY (SP), 2018, : 3 - 18
  • [6] Deep Neural Networks for Acoustic Modeling in Speech Recognition
    Hinton, Geoffrey
    Deng, Li
    Yu, Dong
    Dahl, George E.
    Mohamed, Abdel-rahman
    Jaitly, Navdeep
    Senior, Andrew
    Vanhoucke, Vincent
    Patrick Nguyen
    Sainath, Tara N.
    Kingsbury, Brian
    [J]. IEEE SIGNAL PROCESSING MAGAZINE, 2012, 29 (06) : 82 - 97
  • [7] Safety Verification of Deep Neural Networks
    Huang, Xiaowei
    Kwiatkowska, Marta
    Wang, Sen
    Wu, Min
    [J]. COMPUTER AIDED VERIFICATION, CAV 2017, PT I, 2017, 10426 : 3 - 29
  • [8] Julian K. D., 2018, J GUID CONTROL DYNAM, DOI DOI 10.2514/1.G003724
  • [9] Reluplex: An Efficient SMT Solver for Verifying Deep Neural Networks
    Katz, Guy
    Barrett, Clark
    Dill, David L.
    Julian, Kyle
    Kochenderfer, Mykel J.
    [J]. COMPUTER AIDED VERIFICATION, CAV 2017, PT I, 2017, 10426 : 97 - 117
  • [10] Kvasnica M, 2004, LECT NOTES COMPUT SC, V2993, P448