一句话:提出 DeepHyperion,用光照搜索探索输入的可解释特征空间,揭示不同特征对深度学习系统行为的影响。
深度学习已广泛应用于安全攸关领域,近年来出现多种 DL 测试方法,但现有工作均未评估生成输入的不同可解释特征如何影响系统行为。本文提出 DeepHyperion,一种基于光照搜索(illumination search)的测试方法,旨在探索 DL 系统输入的特征空间。它不单纯追求发现失效,而是通过光照算法(如 MAP-Elites)在特征维度上维持多样性,生成覆盖不同可解释特征组合的输入,从而系统性地揭示特征与系统行为之间的关系。摘要未提供具体实验设置与结果,但该方法为 DL 测试提供了从特征空间角度评估系统行为的新视角。
Deep Learning (DL) has been successfully applied to a wide range of application domains, including safety-critical ones. Several DL testing approaches have been recently proposed in the literature but none of them aims to assess how different interpretable features of the generated inputs affect the system's behaviour.
DOI 原文 ·
@inproceedings{paperbot6656,
title = {DeepHyperion: exploring the feature space of deep learning-based systems through illumination search},
author = {Tahereh Zohdinasab and Vincenzo Riccio and Alessio Gambi and Paolo Tonella},
booktitle = {ISSTA},
year = {2021},
doi = {10.1145/3460319.3464811}
}