基于模型统计特征的深度神经网络对抗鲁棒性评估指标RDI
摘要
深度神经网络(DNN)对对抗样本极其敏感，这引发了对其在 safety-critical 任务中可靠性的关注。当前，对抗鲁棒性评估方法主要归类为 attack-based 和 certified robustness evaluation 两类方法。前者不仅依赖于特定的攻击算法，而且计算耗时，而后者由于其分析性质，通常难以实现于大型复杂模型。少数研究基于模型的 decision boundary 来评估模型鲁棒性，但 suffers from low evaluation accuracy。为此，本文提出了一种 novel 的对抗鲁棒性评估指标Robustness Difference Index (RDI)，其基于 model statistical features。RDI draws inspiration from clustering evaluation by analyzing the intra-class and inter-class distances of feature vectors separated by the decision boundary to quantify model robustness。It is attack-independent and has high computational efficiency。实验表明，RDI demonstrates a stronger correlation with the gold-standard adversarial robustness metric of attack success rate (ASR)。RDI的平均计算时间仅为基于PGD攻击评估方法的1/30。我们的 open-source code is available at: https://github.com/BUPTAIOC/RDI。
引用
@article{arxiv.2504.18556,
title = {RDI: An adversarial robustness evaluation metric for deep neural networks based on model statistical features},
author = {Jialei Song and Xingquan Zuo and Feiyang Wang and Hai Huang and Tianle Zhang},
journal= {arXiv preprint arXiv:2504.18556},
year = {2025}
}