Congzheng Song

If the output is not confidence value then the attack reduces to using one-hot encoding of the predicted label for determine membership inference and the attack performance will likely drop...

Yes if you flatten the image in to 1D array

I skimmed this paper and I think they basically extends our work with some relaxed assumptions, but the core attack scheme is the same.

Hi, You can contact the first author Tyler ([email protected]) about code for Chiron. In the case of MI with few number of classes, we currently do not have a good...

Hi, On cifar10 you will need a bigger neural nets than the one in the demo code. CNN is a good idea :)

You will have to fit each dataset with its own best hyper-parameters.

You may want to pass --save_data=1 for saving target_data.npz

Because he needs to pad sentence with zero vectors. See this function: get_idx_from_sent

Each sentence has different number of words and thus the inputs to the CNN would have different sizes. Therefore, padding is needed to ensure all inputs have the same size.

Hi, can you log the error here? So I can know what is the problem exactly