I can't get good qulity of adversarial examples by DeepFool.

[Jiang010]

These parameters are setted for DeepFool.
umm, I had also tried others parameters. But, the result is not good.

The left figure is normal example. And the righ figure is adversarial example, it looks so bad.

please, who can tell me, how to improve this question?

[beat-buesser]

Hi @Jiang010 Does your model/classifier predict probabilities or logits?

[Jiang010]

Hi @beat-buesser thx you respond my question.
this is my classifier code.

    Conv2D(filters=36, kernel_size=7, activation='relu', input_shape= im_shape, padding="same"),
    MaxPooling2D(pool_size=2),
    Conv2D(filters=54, kernel_size=5, activation='relu'),
    MaxPooling2D(pool_size=2),
    Flatten(),
    .
    .
    .
    Dense(20, activation='softmax')  

I use the "softmax" as activation. I think that my model output is probabilities.
By the way, I had also tried the "sigmoid" as activation. The DeepFool can work.
It should be caused by this reason.

So, DeepFool can't run in my project.

[beat-buesser]

Hi @Jiang010 Can you try with activation=linear? I think this attack works best with logits output.

[Jiang010]

hi, @beat-buesser,
Yeah, just like you say. Using the activation="linear" can well run the DeepFool.
if the activation="linear" not be used, the data of specific label can't well work. It will get terrible results...

umm.. Using the DeepFool may occur some troble in my project.

activation="linear"

cnn_model= Sequential([
     Conv2D(filters=36, kernel_size=7, activation='linear', input_shape= im_shape, padding="same"),
     MaxPooling2D(pool_size=2),
     Conv2D(filters=54, kernel_size=5, activation='linear'),
     MaxPooling2D(pool_size=2),
     Flatten(), 
     Dense(1024, activation='linear'),
     Dense(512, activation='linear'),
     Dense(20, activation='softmax')  
     ])

activation="sigmoid"

cnn_model= Sequential([
    Conv2D(filters=36, kernel_size=7, 
                activation='relu', 
                input_shape= im_shape, 
                padding="same"),
    MaxPooling2D(pool_size=2),
    Conv2D(filters=54, 
                kernel_size=5, 
                activation='relu'),
    MaxPooling2D(pool_size=2),
    Flatten(),
    Dense(1024, activation='relu'),
    Dense(512, activation='relu'),
    Dense(20, activation='sigmoid')  
])

[beat-buesser]

That's great! I think the reason is the algorithm of DeepFool is expecting logits and not probabilities. The implementation of DeepFool algorithm can run with probabilities, but it will be much less successful.