search

TrustAIResearch / MLHospital

MIT License
42 stars 6 forks source link

Problems about the code #3

Closed chihhsunlin closed 1 year ago

chihhsunlin commented 1 year ago

In MLHospital/mlh/data_preprocessing/data_loader.py , the line 44 and the line 68 are same, is it correct?

In MLHospital/mlh/examples/mia_example.py , why the line 94 and the line 98 use the same data source {args.training_type} ? If I select the training_type=PATE when I train a shadow model, the target model should be as the real data, is it right?

Thanks

TrustAIResearch commented 1 year ago

Hi,

Thanks for reaching out.

For the first question: We have modified the code. Line 68 should be: dataset in configs.SUPPORTED_IMAGE_DATASETS_ATTRIBUTE_INFERENCE

For the second question: The target and shadow models can be trained with different methods, e.g., the target model is trained without defense and the shadow model is trained with PATE.

When conducting the MIA, you can modify the mia_example.py to fit your needs, e.g., training_type=Normal for the target model and training_type=PATE for the shadow model.