A Sound Abstraction Method Towards Efficient Neural Networks Verification

被引:0
|
作者
Boudardara, Fateh [1 ]
Boussif, Abderraouf [1 ]
Ghazel, Mohamed [1 ,2 ]
机构
[1] Technol Res Inst Railenium, 180 rue Joseph Louis Lagrange, F-59308 Valenciennes, France
[2] Univ Gustave Eiffel, COSYS ESTAS, 20 rue Elisee Reclus, F-59666 Villeneuve Dascq, France
关键词
Neural network abstraction; model reduction; Neural network verification; Output range computation;
D O I
10.1007/978-3-031-49737-7_6
中图分类号
TP301 [理论、方法];
学科分类号
081202 ;
摘要
With the increasing application of neural networks (NN) in safety-critical systems, the (formal) verification of NN is becoming more than essential. Although several NN verification techniques have been developed in recent years, these techniques are often limited to small networks and do not scale well to larger NN. The primarily reason for this limitation is the complexity and non-linearity of neural network models. Abstraction and model reduction approaches, that aim to reduce the size of the neural networks and over-approximate their outcomes, have been seen as a promising research direction to help existing verification methods to handle larger models. In this paper, we introduce a model reduction method for neural networks with non-negative activation functions (e.g., ReLU and Sigmoid). The method relies on merging neurons while ensuring that the obtained model (i.e., the abstract model) over-approximates the original one. Concretely, it consists in merging a set of neurons that have positive outgoing weights and substituting it with a single abstract neuron, while ensuring that if a given property holds on the abstract network, it necessarily holds on the original one. In order to assess the efficiency of the approach, we perform an experimental comparison with two existing model reduction methods on the ACAS Xu benchmark. The obtained results show that our approach outperforms both methods in terms of precision and execution time.
引用
收藏
页码:76 / 89
页数:14
相关论文
共 50 条
  • [41] A Parallel Optimization Method for Robustness Verification of Deep Neural Networks
    Lin, Renhao
    Zhou, Qinglei
    Nan, Xiaofei
    Hu, Tianqing
    MATHEMATICS, 2024, 12 (12)
  • [42] INNAbstract: An INN-Based Abstraction Method for Large-Scale Neural Network Verification
    Boudardara, Fateh
    Boussif, Abderraouf
    Meyer, Pierre-Jean
    Ghazel, Mohamed
    IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, 2024, 35 (12) : 18455 - 18469
  • [43] Sound and Complete Verification of Polynomial Networks
    Rocamora, Elias Abad
    Sahin, Mehmet Fatih
    Liu, Fanghui
    Chrysos, Grigorios G.
    Cevher, Volkan
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 35, NEURIPS 2022, 2022,
  • [44] Computationally Sound Abstraction and Verification of Secure Multi-Party Computations
    Backes, Michael
    Maffei, Matteo
    Mohammadi, Esfandiar
    IARCS ANNUAL CONFERENCE ON FOUNDATIONS OF SOFTWARE TECHNOLOGY AND THEORETICAL COMPUTER SCIENCE (FSTTCS 2010), 2010, 8 : 352 - 363
  • [45] Towards Energy Efficient Architecture for Spaceborne Neural Networks Computation
    Wang, Shiyu
    Zhang, Shengbing
    Wang, Jihe
    Huang, Xiaoping
    ALGORITHMS AND ARCHITECTURES FOR PARALLEL PROCESSING, ICA3PP 2020, PT II, 2020, 12453 : 575 - 586
  • [46] Self-Distillation: Towards Efficient and Compact Neural Networks
    Zhang, Linfeng
    Bao, Chenglong
    Ma, Kaisheng
    IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2022, 44 (08) : 4388 - 4403
  • [47] AutoSNN: Towards Energy-Efficient Spiking Neural Networks
    Na, Byunggook
    Mok, Jisoo
    Park, Seongsik
    Lee, Dongjin
    Choe, Hyeokjun
    Yoon, Sungroh
    INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 162, 2022,
  • [48] Towards Formal Repair and Verification of Industry-scale Deep Neural Networks
    Munakata, Satoshi
    Tokumoto, Susumu
    Yamamoto, Koji
    Munakata, Kazuki
    2023 IEEE/ACM 45TH INTERNATIONAL CONFERENCE ON SOFTWARE ENGINEERING: COMPANION PROCEEDINGS, ICSE-COMPANION, 2023, : 360 - 364
  • [49] Towards Compositional Abstraction of Analog Neuronal Networks
    Tarraf, Ahmad
    Hedrich, Lars
    2021 IEEE 11TH ANNUAL COMPUTING AND COMMUNICATION WORKSHOP AND CONFERENCE (CCWC), 2021, : 34 - 37
  • [50] Interval Weight-Based Abstraction for Neural Network Verification
    Boudardara, Fateh
    Boussif, Abderraouf
    Meyer, Pierre-Jean
    Ghazel, Mohamed
    COMPUTER SAFETY, RELIABILITY, AND SECURITY, SAFECOMP 2022 WORKSHOPS, 2022, 13415 : 330 - 342