Provable Preimage Under-Approximation for Neural Networks

被引:3
|
作者
Zhang, Xiyue [1 ]
Wang, Benjie [1 ]
Kwiatkowska, Marta [1 ]
机构
[1] Univ Oxford, Dept Comp Sci, Oxford, England
来源
TOOLS AND ALGORITHMS FOR THE CONSTRUCTION AND ANALYSIS OF SYSTEMS, PT III, TACAS 2024 | 2024年 / 14572卷
关键词
VERIFICATION;
D O I
10.1007/978-3-031-57256-2_1
中图分类号
TP31 [计算机软件];
学科分类号
081202 ; 0835 ;
摘要
Neural network verification mainly focuses on local robustness properties, which can be checked by bounding the image (set of outputs) of a given input set. However, often it is important to know whether a given property holds globally for the input domain, and if not then for what proportion of the input the property is true. To analyze such properties requires computing preimage abstractions of neural networks. In this work, we propose an efficient anytime algorithm for generating symbolic under-approximations of the preimage of any polyhedron output set for neural networks. Our algorithm combines a novel technique for cheaply computing polytope preimage under-approximations using linear relaxation, with a carefully-designed refinement procedure that iteratively partitions the input region into subregions using input and ReLU splitting in order to improve the approximation. Empirically, we validate the efficacy of our method across a range of domains, including a high-dimensional MNIST classification task beyond the reach of existing preimage computation methods. Finally, as use cases, we showcase the application to quantitative verification and robustness analysis. We present a sound and complete algorithm for the former, which exploits our disjoint union of polytopes representation to provide formal guarantees. For the latter, we find that our method can provide useful quantitative information even when standard verifiers cannot verify a robustness property.
引用
收藏
页码:3 / 23
页数:21
相关论文
共 50 条
  • [31] Safe Over- and Under-Approximation of Reachable Sets for Autonomous Dynamical Systems
    Li, Meilun
    Mosaad, Peter N.
    Franzle, Martin
    She, Zhikun
    Xue, Bai
    FORMAL MODELING AND ANALYSIS OF TIMED SYSTEMS, FORMATS 2018, 2018, 11022 : 252 - 270
  • [32] Reachability Analysis of Deep Neural Networks with Provable Guarantees
    Ruan, Wenjie
    Huang, Xiaowei
    Kwiatkowska, Marta
    PROCEEDINGS OF THE TWENTY-SEVENTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE, 2018, : 2651 - 2659
  • [33] Sensitivity-Informed Provable Pruning of Neural Networks
    Baykal, Cenk
    Liebenwein, Lucas
    Gilitschenski, Igor
    Feldman, Dan
    Rus, Daniela
    SIAM JOURNAL ON MATHEMATICS OF DATA SCIENCE, 2022, 4 (01): : 26 - 45
  • [34] Using Gate Recognition and Random Simulation for Under-Approximation and Optimized Branching in SAT Solvers
    Iser, Markus
    Kutzner, Felix
    Sinz, Carsten
    2017 IEEE 29TH INTERNATIONAL CONFERENCE ON TOOLS WITH ARTIFICIAL INTELLIGENCE (ICTAI 2017), 2017, : 1029 - 1036
  • [35] An under-approximation for the robust uncertain two-level cooperative set covering problem
    Ding, Shuxin
    Zhang, Qi
    Yuan, Zhiming
    2020 59TH IEEE CONFERENCE ON DECISION AND CONTROL (CDC), 2020, : 1152 - 1157
  • [36] Architecture-Preserving Provable Repair of Deep Neural Networks
    Tao, Zhe
    Nawas, Stephanie
    Mitchell, Jacqueline
    Thakur, Aditya V.
    PROCEEDINGS OF THE ACM ON PROGRAMMING LANGUAGES-PACMPL, 2023, 7 (PLDI):
  • [37] Provable Guarantees for Neural Networks via Gradient Feature Learning
    Shi, Zhenmei
    Wei, Junyi
    Liang, Yingyu
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 36 (NEURIPS 2023), 2023,
  • [38] PROVABLE TRANSLATIONAL ROBUSTNESS FOR OBJECT DETECTION WITH CONVOLUTIONAL NEURAL NETWORKS
    Vierling, Axel
    James, Charu
    Berns, Karsten
    Katsaouni, Nikoletta
    2021 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING (ICIP), 2021, : 694 - 698
  • [39] Can SGD Learn Recurrent Neural Networks with Provable Generalization?
    Allen-Zhu, Zeyuan
    Li, Yuanzhi
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 32 (NIPS 2019), 2019, 32
  • [40] Post-training Quantization for Neural Networks with Provable Guarantees*
    Zhang, Jinjie
    Zhou, Yixuan
    Saab, Rayan
    SIAM JOURNAL ON MATHEMATICS OF DATA SCIENCE, 2023, 5 (02): : 373 - 399