一句话：提出协同增强框架 CoopHance，用净化器与检测器协作提升深度学习系统抗对抗样本的鲁棒性。
对抗攻击通过对良性输入施加人眼不可感知的扰动，使深度学习系统产生错误行为，因此提升其鲁棒性十分迫切。本文提出 CoopHance，协同利用两个定制组件：Regulator 负责净化低或中等失真的对抗样本，Inspector 则通过捕捉深度学习系统的异常状态来检测高失真对抗样本。作者在 CIFAR-10 和 SVHN 数据集上使用多种攻击进行评估。结果表明，CoopHance 平均可成功抵御 90.62% 和 96.56% 针对未保护系统生成的对抗样本，比 Feature Squeeze、LID、SOAP、Adversarial Training 和 MagNet 五种先进技术有效 188.14%；当攻击者在增强系统上生成新对抗样本时，CoopHance 可拒绝 78.06% 的攻击，平均优于最佳增强技术 82.71%。
Adversarial attacks have been a threat to Deep Learning (DL) systems to be reckoned with. By adding human-imperceptible perturbation to benign inputs, adversarial attacks can cause the incorrect behavior of DL systems. Considering the popularity of DL systems in the industry, it is critical and urgent for developers to enhance the robustness of DL systems against adversarial attacks. In this study, we propose a novel enhancement technique for DL systems, namely CoopHance. CoopHance leverages two specifically customized components, Regulator and Inspector, to cooperatively enhance the DL systems' robustness against adversarial examples with different distortions. Regulator can purify adversarial examples with low or moderate distortions, while Inspector is responsible for detecting these adversarial examples with high distortion by capturing the abnormal status of DL systems. Our evaluation using various attacks shows that, on average, CoopHance can successfully resist 90.62% and 96.56% of the adversarial examples that are generated for the unprotected systems on CIFAR-10 and SVHN datasets separately, which is 188.14% more effective than five state-of-the-art enhancement techniques, including Feature Squeeze, LID, SOAP, Adversarial Training, and MagNet. Meanwhile, when attackers generate new adversarial examples on the enhanced systems, CoopHance can reject 78.06% of attacks, which outperforms the best of five enhancement techniques by 82.71% on average.
@inproceedings{Zhang0DLS0S23,
title = {CoopHance: Cooperative Enhancement for Robustness of Deep Learning Systems},
author = {Quan Zhang and Yongqiang Tian and Yifeng Ding and Shanshan Li and Chengnian Sun and Yu Jiang and Jiaguang Sun},
booktitle = {ISSTA},
year = {2023},
doi = {10.1145/3597926.3598093}
}