Survey on Acceleration Techniques for Complete Neural Network Verification
Author:
Affiliation:

Clc Number:

Fund Project:

  • Article
  • |
  • Figures
  • |
  • Metrics
  • |
  • Reference
  • |
  • Related
  • |
  • Cited by
  • |
  • Materials
  • |
  • Comments
    Abstract:

    Artificial intelligence (AI) has been widely applied to various aspects of lives. However, as the primary technique of realizing AI, neural networks can exhibit undefined behavior when faced with inputs outside of their training data or under adversarial attacks. In safety-critical fields such as autonomous driving and intelligent healthcare, undefined behavior can pose significant threats to human safety. Therefore, it is particularly crucial to employ complete verification methods to verify the properties of neural networks and ensure the correctness of the behavior. To enhance the verification efficiency, various complete neural network verification tools have proposed their optimization methods. However, the true influence of these methods has not been thoroughly explored, making it challenging for researchers to identify the most effective optimization directions. This paper introduces the common techniques in neural network verification and presents a universal framework for complete neural network verification. Within this framework, it focuses on discussing the optimization methods employed by state-of-the-art tools for constraint solving, branch selection, and boundary computation. Meanwhile, a series of experiments are designed to investigate the contributions of various acceleration techniques to the performance of each tool and to explore the most effective acceleration strategies and more promising optimization directions. Finally, valuable references can be provided for researchers in this field.

    Reference
    Related
    Cited by
Get Citation

刘宗鑫,杨鹏飞,张立军,吴志林,黄小炜.完备神经网络验证加速技术综述.软件学报,2024,35(9):4038-4068

Copy
Share
Article Metrics
  • Abstract:
  • PDF:
  • HTML:
  • Cited by:
History
  • Received:September 10,2023
  • Revised:October 30,2023
  • Adopted:
  • Online: January 05,2024
  • Published: September 06,2024
You are the firstVisitors
Copyright: Institute of Software, Chinese Academy of Sciences Beijing ICP No. 05046678-4
Address:4# South Fourth Street, Zhong Guan Cun, Beijing 100190,Postal Code:100190
Phone:010-62562563 Fax:010-62562533 Email:jos@iscas.ac.cn
Technical Support:Beijing Qinyun Technology Development Co., Ltd.

Beijing Public Network Security No. 11040202500063