Li Jiasen's readme:
python inp.py
run FMNIST or MNIST
MNIST 99.5
FMNIST 90
This is ResNet-6
Our task
train mnist2/3/deeper (Li Jiasen) acc > 99.2
train a fooler for specific model(mnist2/3) (Li Jiasen) can fool 10 % of all data
see if the fooling image looks like the origin image (low change)
use fooler to fool model_mnist2 see if it can fool the model_mnist3 (the same model with different initial weights)
if it can fool mnist3 try to fool mnist_deep (another model)(ResNet-7)
if it cannot fool mnist3 then we come up with a method to output: class or uncertain
train mnist_defooling with fooling images run about
白盒fooling
1.具体的参数的值是否影响fooling(同样的网络,不同的初始值)(此实验需要重复多次)
2.如果有效,那么对于不同结构和相同数据是否有效
3.如果无效,那么是否对于不同结构,同时被fooling的概率相当小?
4.用自己的fooling image去训练