ModelPred

Repository for paper "ModelPred: A Framework for Predicting Trained Model from Training Data" 2023 SaTML submission.

ModelPred: A Framework for Predicting Trained Model from Training Data

Requirements

You can first install the environment specified in the requirements.txt

ModelPred DNN model training

Logistic Regression as the base model: Iris_LR.py, spam_LR.py, HIGGS_LR.py, MNIST_LR.py Support Vector Machine as the base model: Iris_SVM.py, spam_SVM.py, HIGGS_SVM.py

To train the model, run this command: Iris_LR.py --sampling perm --datapath --modelpath The same for other scripts

Dataset deletion and addition

Logistic Regression as the base model: Deletion_Eva.py Addition_Eva.py

To run the experiment, run this command: Deletion_Eva.py --sampling Perm --rawdatapth --modelpath --savepath Addition_Eva.py --sampling Perm --rawdatapth --modelpath --savepath

Shapley value

LR_Shapley.py

To run the experiment, run this command: LR_Shapley.py --sampling Perm --maxiter 50 --rawdatapth --modelpath --savepath

Please make sure the data_path (where you save the raw data), save_path (where you save the training samples for ModelPred), result_path (where you save all the results), and model_path (where you save the trained models) are all correctly configured.