GRU model for LOS (#170)

* ReadMe file for creating dataset from MIMICIII, and training LOS>3 classifier * adding links to readMe file * fixing the LR model * Mimic_Hnadler, data_processeing, model and the main file * all attacks are running * add relevant files from MIMIC_Extract and mimic_code repos * fix some bugs regarding nivduration * revert conflicting tabular_mia files * adding subset to the mimicDataset * fixing loading the database and indices * adding GRU * adding GRU-D structure * updating grud class * adding early stopping * the model is running * the model is running * working on meta data * mimic hanlder gru is added * LiRA in place * running lira * LiRA and RMIA for GRU, temp implmemntation * adding flag for changing the classifier * LR and read me * up to adding report handler module * renaming files, adding report handler module * renaming files, adding report handler module * adding the license, renaming some files and classes * fixing the accuracy in model handler * fixing pandas warning and adding a check to saved shadow models model type * fixing a bug in shadow_model_handler * ruff fix
aidotse · Jan 10, 2025 · f30b53a · f30b53a
1 parent e5c363d
commit f30b53a
Show file tree

Hide file tree

Showing 14 changed files with 2,063 additions and 488 deletions.
diff --git a/examples/mia/LOS/ReadMe.md b/examples/mia/LOS/ReadMe.md
@@ -0,0 +1,5 @@
+# Lenght-of-Stay Usecase
+In this use case, we focus on attacking length-of-stay classifier models. As part of the example, we train a Logistic Regression model and a Gated Recurrent Unit with Decay (GRU-D).
+
+##  MIMIC-III Data Preprocessing
+To prepare the data, refer to instructions in ```mimic_prepration/ReadMe.md ``` 
diff --git a/examples/mia/LOS/audit.yaml b/examples/mia/LOS/audit.yaml
@@ -1,15 +1,15 @@
 audit:  # Configurations for auditing
   random_seed: 1234  # Integer specifying the random seed
   attack_list:
-    # rmia:
-    #   training_data_fraction: 0.5  # Fraction of the auxilary dataset to use for this attack (in each shadow model training)
-    #   attack_data_fraction: 0.5 # Fraction of auxiliary dataset to sample from during attack
-    #   num_shadow_models: 8 # Number of shadow models to train
-    #   online: True # perform online or offline attack
-    #   temperature: 2
-    #   gamma: 1.0
-    #   offline_a: 0.33 # parameter from which we compute p(x) from p_OUT(x) such that p_IN(x) = a p_OUT(x) + b.
-    #   offline_b: 0.66
+    rmia:
+      training_data_fraction: 0.5  # Fraction of the auxilary dataset to use for this attack (in each shadow model training)
+      attack_data_fraction: 0.5 # Fraction of auxiliary dataset to sample from during attack
+      num_shadow_models: 8 # Number of shadow models to train
+      online: True # perform online or offline attack
+      temperature: 2
+      gamma: 1.0
+      offline_a: 0.33 # parameter from which we compute p(x) from p_OUT(x) such that p_IN(x) = a p_OUT(x) + b.
+      offline_b: 0.66
     # qmia:
     #   training_data_fraction: 1.0  # Fraction of the auxilary dataset (data without train and test indices) to use for training the quantile regressor
     #   epochs: 5  # Number of training epochs for quantile regression
@@ -32,12 +32,13 @@ audit:  # Configurations for auditing
 
 target:
   # Target model path
-  module_path: "utils/model.py"
-  model_class: "MimicLR" 
+  module_path: "utils/model_LR.py" # either model_grud.py or model_LR.py for logestic regression
+  model_class: "LR" #  LR/GRUD
   # Data paths
-  target_folder: "./target"
-  data_path: "./data/dataset.pkl"
+  target_folder: "./target_LR" # either target_GRUD or target_LR
+  data_path: "./data/flattened/dataset.pkl" #unflattened dataset for GRUD and flattened dataset for LR
 
 shadow_model:
+  model_class: # LR/GRUD
 
 distillation_model: