Expected length of time for attack?

[ShailenSmith]

Hi,

My lab and I are planning on including IF-GMI in our latest project on model inversion defenses, but we ran into some computational limitations. We seek to attack a Resnet-152 model on all 530 target classes of the FaceScrub dataset; we see you have a similar experimental setup in your paper.

We're running the attack on 2 A100 GPUs with what I believe is the same configuration as your main FaceScrub experiments. See our full config file here: our_IF_config.txt

For us, this run is taking about 8 minutes per class (2 minutes per optimization batch x 3 batches per class + 2 min for evaluation computations). It appears like this will result in a >60hr run to attack all classes, which feels quite long. How long did your experimental runs with a Resnet-152 on FaceScrub take? If I may ask, what kind of GPUs did you use, and how many?

Thanks for your time, and for making your code publicly available!

-- Shailen

[final-solution]

Thank you for your attention! We used 4 × NVIDIA RTX 4090 GPUs (24GB VRAM each) in our experiments and it took about 3.5 days to run with a Resnet-152 on FaceScrub. Given the A100’s superior VRAM capacity, we recommend increasing the batch size parameter in the attack stage to maximize throughput and reduce runtime.