推荐katmannthree · 2021-05
“This study[0] discusses adversarial images which fool both ML classifiers and time limited humans, a rather interesting development in the field.”
推荐它,对抗图像能同时骗过 ML 分类器和限时人类,进展有趣
迁移性强的对抗样本也会影响限时观察的人类判断
机器学习模型容易被对抗样本欺骗,例如把校车识别为鸵鸟,但人类是否会犯类似错误尚不清楚。论文借助把对抗样本从已知结构与参数的模型迁移到未知模型的新技术,并模拟人类视觉系统的初始处理,发现在多个计算机视觉模型之间迁移性强的对抗样本,也会影响限时观察者的分类判断。
HN 上的人引用它说明人类也会被针对 AI 构造的扰动欺骗,是对抗样本研究里一个有趣的进展。
“This study[0] discusses adversarial images which fool both ML classifiers and time limited humans, a rather interesting development in the field.”
“humans are vulnerable to perturbations created to fool a multitude of existing AI: Adversarial Examples that Fool both Computer Vision and Time-Limited Humans”
“Looks like we are starting to find examples.”