Could you please provide a download link to the cifar10_ssba_train_b1.npy for running issba attack? about backdoorbench HOT 2 CLOSED

https://drive.google.com/drive/folders/1QU771F2_1mKgfNQZm3OMCyegu2ONJiU2?usp=sharing

You should notice that since SSBA uses an autoencoder-like network to generate the poisoned samples, the hyper-parameters used in the training process and poison process may influence the final performance.

Here, we only use 1-bit information to do the poisoning process such that it can be as imperceptible as possible for low-resolution datasets. (For 224 * 224 pictures, SSBA uses more information to do the poisoning as its standard since larger resolution allows higher information embedded.)

We have not tried the poisoning process on low-resolution datasets with a more significant amount of information. And it is possible to improve the attack performance with a more significant amount of information embedded.

from backdoorbench.

Thanks very much for your kindly sharing!

from backdoorbench.