【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks

核心思想:

①Curls迭代:

梯度下降/上升方向 + 二分法优化 + varience-reduced优化
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第1张图片
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第2张图片
其中,(12)记录先前的对抗以对当前方向产生影响:
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第3张图片
(15)为二分法,比较简单:
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第4张图片

②Whey优化:

将Curls找出的最优对抗样本中的元素按大小排列,逐一减半观察效果,不变保留操作否则放弃操作:
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第5张图片

③实验结果

【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第6张图片
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第7张图片
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第8张图片
【迁移攻击笔记】Curls & Whey: Boosting Black-Box Adversarial Attacks_第9张图片

你可能感兴趣的:(像素攻击,算法,python,对抗攻击)