MagNet 和 "高效抵御对抗攻击" 对抗对抗性样本的鲁棒性不足

Nov, 2017

MagNet 和 "高效抵御对抗攻击" 对抗对抗性样本的鲁棒性不足

MagNet and "Efficient Defenses Against Adversarial Attacks" are Not Robust to Adversarial Examples

Nicholas Carlini, David Wagner

TL;DR研究发现 MagNet 模型和高效防御模型并不能完全对抗有轻微失真的对抗样本攻击。

Abstract

magnet and "efficient defenses..." were recently proposed as a defense to adversarial examples. We find that we can construct

magnet efficient defenses adversarial examples distortion

发现论文，激发创造

MagNet：对抗样本的双重防御

MagNet is proposed as a defense mechanism for neural network classifiers against adversarial examples in deep learning, which learns to differentiate between normal and adversarial examples by approximating the manifold of normal examples and reconstructing adversarial examples by moving them towards the manifold, and it also proposes a mechanism to defend against graybox attack by using diversity to strengthen MagNet.

May, 2017

对抗样本防御：弱防御的集成无法变强

本文探讨多种防御手段组成的强防御系统在防御神经网络对抗性样本方面的效果，研究表明仅靠多种弱防御手段的组合并不能提供强大的对抗性样本防御能力。

Jun, 2017

防御蒸馏不具备对抗性示例的稳健性

防御性蒸馏并不安全，在面对有针对性的错误分类攻击时，其并没有比未经保护的神经网络更具抵抗力。

Jul, 2016

倡导采用多个防御策略对抗对抗性示例

本文通过几何分析和数值实验，探讨防御机制在保护神经网络免受 $\ell_2$ 对抗样本攻击效果不佳的问题，并评估已有的混合防御策略对于对抗样本攻击的现实意义和潜在问题。

Dec, 2020

加强对抗性容错性评估的置信度

该研究提出了一种测试方法以识别弱攻击和防御评估，为了增强透明和信心，将攻击单元测试作为未来强度评估的重要组成部分。