| 本文已被：浏览 2774次 下载 7607次 |
码上扫一扫！ |
|
|
| 前馈神经网络和循环神经网络的鲁棒性验证综述 |
|
刘颖1,2, 杨鹏飞1,3, 张立军1,2, 吴志林1,2, 冯元4
|
|
1.计算机科学国家重点实验室 (中国科学院 软件研究所), 北京 100190;2.中国科学院大学, 北京 100049;3.琶洲实验室, 广东 广州 510320;4.University of Technology Sydney, Sydney 2007, Australia
|
|
| 摘要: |
| 随着智能时代的到来，部署了深度神经网络的智能系统应用已经渗透到了人类生活的各个方面.然而，由于神经网络具有黑盒特性和规模庞大的特点，其预测结果难以让人完全信服，当应用于自动驾驶等安全攸关的领域时，如何保证其安全性仍然是学术界和工业界面临的巨大挑战.为此，学术界针对神经网络一种特殊的安全性——鲁棒性展开了研究，并提出了很多鲁棒性的分析和验证方法.目前为止，验证前馈神经网络的方法包括精确验证方法和近似验证方法，已经发展得比较繁荣；而对于其他类型的网络，如循环神经网络的鲁棒性验证研究还处于起步阶段.回顾深度神经网络的发展以及部署到日常生活中面临的挑战；详尽地调研前馈神经网络和循环神经网络的鲁棒性验证方法，并对这些验证方法间的内在联系进行分析和比较；调研循环神经网络在现实应用场景中的安全性验证方法；阐明神经网络鲁棒性验证领域未来可以深入研究的方向. |
| 关键词: 神经网络 鲁棒性 人工智能安全 智能系统 形式化方法 |
| DOI：10.13328/j.cnki.jos.006863 |
| 分类号: |
| 基金项目:中国科学院青年科学基金（YSBR-040） |
|
| Survey on Robustness Verification of Feedforward Neural Networks and Recurrent Neural Networks |
|
LIU Ying1,2, YANG Peng-Fei1,3, ZHANG Li-Jun1,2, WU Zhi-Lin1,2, FENG Yuan4
|
|
1.State Key Laboratory of Computer Science (Institute of Software, Chinese Academy of Sciences), Beijing 100190, China;2.University of Chinese Academy of Sciences, Beijing 100049, China;3.Pazhou Lab, Guangzhou 510320, China;4.University of Technology Sydney, Sydney 2007, Australia
|
| Abstract: |
| With the advent of the intelligent age, the applications of intelligent systems equipped with deep neural networks (DNNs) have penetrated into every aspect of our life. However, due to the black-box and large-scale characteristics, the predictions of the neural networks are difficult to be completely convincing. When neural networks are applied to security-critical fields such as autonomous driving, how to ensure their security is still a great challenge for the academia and industry. For this reason, the academia carried out much research on robustness—a kind of special security of neural networks, and proposed many algorithms for robustness analysis and verification. The verification algorithms for feedforward neural networks (FNNs) include precise algorithms and approximate algorithms, which have been developed relatively prosperously. The verification algorithms for other types of networks, such as recurrent neural networks (RNNs), are still in the primary stage. This study reviews the current development of DNNs and the challenges of deploying them into our life. It also exhaustively investigates the robustness verification algorithms of FNNs and RNNs, analyzes and compares the intrinsic connection among these algorithms. The security verification algorithms of RNNs in specific application scenarios are investigated, and the future research directions in the robustness verification field of neural networks are clarified. |
| Key words: neural network robustness artificial intelligence security intelligent system formal method |
