Example scripts and configurations for image classification experiments are collected in scripts/exp-img and configs/exp-img respectively.
Here we present a brief tutorial on how to run experiments with these scripts and configurations (Please check comments in these scripts to see how to set encoder, downstream task, etc.).
Step 0: Enter the folder:
cd ./scripts/exp-img
Step 1: Build the embedding dataset, e.g., rn34-hf + cifar10:
bash embed.sh ../../
Step 2: Train the downstream classifier on the built embedding dataset e.g., rn34-hf + cifar10, with the configuration configs/exp-img/train.yaml:
bash train.sh ../../
Step 3: Synthesize PEI attack images for a candidate encoder (e.g., rn34-hf) with the configuration configs/exp-img/atk.yaml and objective images in scripts/obj-img:
bash atk.sh ../../
Step 4: Calculate the PEI score of a candidate encoder (e.g., rn50-hf) on a targeted downstream service (e.g., rn34-hf + cifar10) with the configuration configs/exp-img/eval.yaml:
bash eval.sh ../../
Obtained PEI scores can then be used to calculate PEI z-scores to conduct the PEI attack (see the paper for details).