一句话：提出 RANUM，首个既能用失败测试确认 DNN 潜在数值缺陷可行性、又能自动生成修复的方法，在 63 个真实架构上优于现有方法。
深度神经网络（DNN）已广泛部署，而数值缺陷（如 NaN、无穷大、溢出与下溢）是 DNN 中最常见的缺陷之一，可能引发系统失效。本文提出 RANUM 方法，针对三项可靠性保障任务给出新技术：检测架构中潜在的数值缺陷、通过能触发失败的测试确认潜在缺陷的可行性、以及自动给出缺陷修复建议。据作者所知，RANUM 是首个既能确认潜在缺陷可行性、又能自动修复的方法。在 63 个真实 DNN 架构基准上的大量实验表明，RANUM 在三项任务上均优于当前最优方法；将其生成的修复与开源项目开发者的人工修复对比，40 例中有 37 例等价甚至更优。
With the widespread deployment of deep neural networks (DNNs), ensuring the reliability of DNN-based systems is of great importance. Serious reliability issues such as system failures can be caused by numerical defects, one of the most frequent defects in DNNs. To assure high reliability against numerical defects, in this paper, we propose the RANUM approach including novel techniques for three reliability assurance tasks: detection of potential numerical defects, confirmation of potential-defect feasibility, and suggestion of defect fixes. To the best of our knowledge, RANUM is the first approach that confirms potential-defect feasibility with failure-exhibiting tests and suggests fixes automatically. Extensive experiments on the benchmarks of 63 real-world DNN architectures show that RANUM outperforms state-of-the-art approaches across the three reliability assurance tasks. In addition, when the RANUM-generated fixes are compared with developers' fixes on open-source projects, in 37 out of 40 cases, RANUM-generated fixes are equivalent to or even better than human fixes.
DOI 原文 · arXiv · PDF（开放获取） · DBLP
@inproceedings{abs-2302-06086,
title = {Reliability Assurance for Deep Neural Network Architectures Against Numerical Defects},
author = {Linyi Li and Yuhao Zhang and Luyao Ren and Yingfei Xiong and Tao Xie},
booktitle = {ICSE},
year = {2023},
doi = {10.1109/icse48619.2023.00156}
}