1.更新策略选择

①optimization-based和多步方法IGSM:成功率高;转移性低。(可能“过拟合”or陷入局部最优)
②单步方法FGSM: 成功率低;转移性高。
所以用: MI-FGSM
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum

2.集成方式选择

MI-FGSM还不够,得利用集成网络来进一步提高成功率。利用的方式有3种待测试:
①输入softmax的logits
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
②softmax输出的prediciton
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
③计算得到的loss
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
最终通过测试选择输入softmax的logits
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum

3.实验

单网络:
①I-FGSM与MI-FGSM
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
②动量因子μ
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
③扰动范围:
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
④总结果:
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
集成网络:
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum
【迁移攻击论文笔记】动量逻辑集成!MI-FGSM!Boosting Adversarial Attacks with Momentum

希望路过这儿的你可以关注我一下~~我会定期更新一系列阅读笔记和总结,加入自己的见解和思路,希望能对你有用~

相关文章: