引言
using a less perceptible adversarial perturbation
Adversarial examples are synthetic patterns carefully crafted by adding a peculiar noise to legitimate examples
neural networks are susceptible to adversarial examples, carefully perturbed inputs that cause net-works to misbehave in arbitrarily chosen ways.
词汇
differentiable surrogate loss function 替代损失函数
asymptotically 渐进的
In a nutshell 简言之
精心制作 elaborate
Attacks are increasingly elaborate,
as demonstrated by the variety of strategies available to evade malware detection built with ML
identity theft（身份盗窃）、phishing（网络钓鱼）、smishing（短信钓鱼）、stoozing（信用卡套现投资）
句型
Machine learning algorithms are usually designed under
the assumption that models are trained on samples drawn
from a distribution that is representative of test samples for which they will later make predictions
实验
we evaluate our approach on multiple adversarial ex-
ample crafting methods (including the fast gradient sign and Jacobian-based saliency map methods) with several
datasets.
