Awesome

SecretGen: Privacy Recovery on Pre-trained Models via Distribution Discrimination

Requirements

Python 3.8 or higher PyTorch 1.8 or higher

$ pip install requirements.txt

Performing Attack

stage1.py: Train the generation backbone on public data.

$ python stage1.py --name <taret_model_arch> --mask <type_of_mask>

Set bb to True if it's blackbox case, which will use a public model instead of the target model for diversity loss.

stage2.py: Perform attack.

$ python stage2.py --name <taret_model_arch> --mask <type_of_mask> --target <method>

For the target parameter:

• pii: PII (whitebox)
• pii-bb: PII (blackbox)
• gmi: GMI
• init-bb: SecretGen (blackbox)
• full-bb: SecretGen (blackbox + ground truth label)
• init-wb: SecretGen (whitebox)
• full: SecretGen (white + ground truth label)

Set save to True if you want to run evaluation protocol 2, which requires a completely recovered dataset.

Pre-trained Checkpoints

We release the checkpoints for our VGG16 target model and the corresponding generation backbones at this link:

https://drive.google.com/drive/folders/149LMfBEmhcFr1S2y6PLXf3WqqPA8-We0?usp=sharing