Merge branch 'PR#1'

config/adult.yaml

@@ -0,0 +1,38 @@
+run: # Configurations for a specific run
+  random_seed: 1234 # Integer number of specifying random seed
+  log_dir: test_adult # String for indicating where to save all the information, including models and computed signals. We can reuse the models saved in the same log_dir.
+  time_log: True # Indicate whether to log the time for each step
+
+audit: # Configurations for auditing
+  privacy_game: privacy_loss_model # Indicate the privacy game from privacy_loss_model, avg_privacy_loss_training_algo, privacy_loss_sample
+  attack_list: [rmia] # String for indicating the membership inference attack. We currently support population, reference.
+  report_log: test_adult # String that indicates the folder where we save the and auditing report.
+  device: cuda:0 # String for indicating on which device we conduct the membership inference attack and train reference models.
+  audit_batch_size: 1000 # Integer number for indicating the batch size for computing signals in the Privacy Meter.
+  num_shadow_models: 1
+  f_attack_data_size: 0.01
+
+train: # Configuration for training
+  type: pytorch # Training framework (we only support pytorch now).
+  model_name: nn # String for indicating the model type. We support CNN, wrn28-1, wrn28-2, wrn28-10, vgg16, nn and speedyresnet (requires cuda). More model types can be added in model.py.
+  num_target_model: 1 #Integer number for indicating how many target models we want to audit for the privacy game
+  device: cuda:0 # String for indicating the device we want to use for training models, e.g., cpu or cuda.
+  epochs: 100 # Integer number for indicating the epochs for training target model. For speedyresnet, it uses its own number of epochs.
+  batch_size: 128 # Integer number for indicating batch size for training the target model. For speedyresnet, it uses its own batch size.
+  optimizer: SGD # String which indicates the optimizer. We support Adam and SGD. For speedyresnet, it uses its own optimizer.
+  learning_rate: 0.1 # Float number for indicating learning rate for training the target model. For speedyresnet, it uses its own learning_rate.
+  weight_decay: 0.01 # Float number for indicating weight decay for training the target model. For speedyresnet, it uses its own weight_decay.
+  test_batch_size: 250 # Integer number for indicating batch size for evaluating the target model.
+  num_test_size: 1000 # Integer number for indicating the size of the test dataset for evaluating the target model during the training. This should be divisible by test_batch_size.
+  inputs: 104
+  outputs: 2
+
+data: # Configuration for data
+  dataset: adult # String indicates the name of the dataset (i.e., cifar10, cifar100, purchase100, texas1000)
+  f_train: 0.01 # Float number from 0 to 1 indicating the fraction of the train dataset
+  f_test: 0.01 # Float number from 0 to 1 indicating the fraction of the test dataset
+  split_method: no_overlapping # String for indicating the methods of splitting the dataset between train, test, and auditing.
+  f_audit: 0.3 # Float from 0 to 1, indicating the fraction of the auditing dataset
+  data_dir: ./data # String about where to save the data.
+
+

config/cifar10.yaml

@@ -0,0 +1,39 @@
+run: # Configurations for a specific run
+  random_seed: 1234 # Integer number of specifying random seed
+  log_dir: test_cifar10 # String for indicating where to save all the information, including models and computed signals. We can reuse the models saved in the same log_dir.
+  time_log: True # Indicate whether to log the time for each step
+
+audit: # Configurations for auditing
+  privacy_game: privacy_loss_model # Indicate the privacy game from privacy_loss_model, avg_privacy_loss_training_algo, privacy_loss_sample
+  attack_list: [rmia] # String for indicating the membership inference attack. We currently support population, reference.
+  report_log: test_cifar10 # String that indicates the folder where we save the and auditing report.
+  device: cuda:0 # String for indicating on which device we conduct the membership inference attack and train reference models.
+  audit_batch_size: 1000 # Integer number for indicating the batch size for computing signals in the Privacy Meter.
+  num_shadow_models: 1
+  f_attack_data_size: 0.05
+
+train: # Configuration for training
+  type: pytorch # Training framework (we only support pytorch now).
+  model_name: nn # String for indicating the model type. We support CNN, wrn28-1, wrn28-2, wrn28-10, vgg16, nn and speedyresnet (requires cuda). More model types can be added in model.py.
+  num_target_model: 1 #Integer number for indicating how many target models we want to audit for the privacy game
+  device: cuda:0 # String for indicating the device we want to use for training models, e.g., cpu or cuda.
+  epochs: 5 # Integer number for indicating the epochs for training target model. For speedyresnet, it uses its own number of epochs.
+  batch_size: 32 # Integer number for indicating batch size for training the target model. For speedyresnet, it uses its own batch size.
+  optimizer: SGD # String which indicates the optimizer. We support Adam and SGD. For speedyresnet, it uses its own optimizer.
+  learning_rate: 0.01 # Float number for indicating learning rate for training the target model. For speedyresnet, it uses its own learning_rate.
+  momentum: 0.9
+  weight_decay: 0 # Float number for indicating weight decay for training the target model. For speedyresnet, it uses its own weight_decay.
+  test_batch_size: 250 # Integer number for indicating batch size for evaluating the target model.
+  num_test_size: 1000 # Integer number for indicating the size of the test dataset for evaluating the target model during the training. This should be divisible by test_batch_size.
+  inputs: 104
+  outputs: 2
+
+data: # Configuration for data
+  dataset: cifar10 # String indicates the name of the dataset (i.e., cifar10, cifar100, purchase100, texas1000)
+  f_train: 0.5 # Float number from 0 to 1 indicating the fraction of the train dataset
+  f_test: 0.5 # Float number from 0 to 1 indicating the fraction of the test dataset
+  split_method: no_overlapping # String for indicating the methods of splitting the dataset between train, test, and auditing.
+  f_audit: 0.3 # Float from 0 to 1, indicating the fraction of the auditing dataset
+  data_dir: ./data # String about where to save the data.
+
+

data/adult/Index

@@ -0,0 +1,6 @@
+Index of adult
+
+02 Dec 1996      140 Index
+10 Aug 1996  3974305 adult.data
+10 Aug 1996     4267 adult.names
+10 Aug 1996  2003153 adult.test