Share Config for Cifar10 eps != 1

[RoyRin]

Hi,

In the paper https://arxiv.org/pdf/2204.13650.pdf, the following plot is presented: . However, the configs associated https://github.com/deepmind/jax_privacy/tree/main/experiments/image_classification/configs only have CIFAR10 for stop_training_at_epsilon = 1.

I am curious what the config that acheived SOTA looks like. Is the only change the "stop_training_at_epsilon" was set to a larger number, in order to generated the plots in the paper (I presume 8). Or are other values are different, such as std_relative=10.

• If possible, could you share how std_relative=10. was selected?
• and if this config was the config used for the SOTA CIFAR-10 results.

[lberrada]

Hi, all the hyper-parameters are reported in appendix C.5 of the paper, their values should be straightforward to plug into the existing config files.