def generate_mnist_experiment(train, test, output, train_output_csv_file, test_output_csv_file):
logger.info("Generating experiment...")
# Setup the files based on user inputs
train_csv_file = os.path.abspath(train)
test_csv_file = os.path.abspath(test)
if not os.path.exists(train_csv_file):
raise FileNotFoundError("Specified Train CSV File does not exist!")
if not os.path.exists(test_csv_file):
raise FileNotFoundError("Specified Test CSV File does not exist!")
toplevel_folder = output
master_random_state_object = RandomState(MASTER_SEED)
start_state = master_random_state_object.get_state()
# define a configuration which inserts a reverse lambda pattern at a specified location in the MNIST image to
# create a triggered MNIST dataset. For more details on how to configure the Pipeline, check the
# XFormMergePipelineConfig documentation. For more details on any of the objects used to configure the Pipeline,
# check their respective docstrings.
one_channel_alpha_trigger_cfg = \
tdc.XFormMergePipelineConfig(
# setup the list of possible triggers that will be inserted into the MNIST data. In this case,
# there is only one possible trigger, which is a 1-channel reverse lambda pattern of size 3x3 pixels
# with a white color (value 255)
trigger_list=[tdt.ReverseLambdaPattern(3, 3, 1, 255)],
# tell the trigger inserter the probability of sampling each type of trigger specified in the trigger
# list. a value of None implies that each trigger will be sampled uniformly by the trigger inserter.
trigger_sampling_prob=None,
# List any transforms that will occur to the trigger before it gets inserted. In this case, we do none.
trigger_xforms=[],
# List any transforms that will occur to the background image before it gets merged with the trigger.
# Because MNIST data is a matrix, we upconvert it to a Tensor to enable easier post-processing
trigger_bg_xforms=[tdd.ToTensorXForm()],
# List how we merge the trigger and the background. Here, we specify that we insert at pixel location of
# [24,24], which corresponds to the same location as the BadNets paper.
trigger_bg_merge=tdi.InsertAtLocation(np.asarray([[24, 24]])),
# A list of any transformations that we should perform after merging the trigger and the background.
trigger_bg_merge_xforms=[],
# Denotes how we merge the trigger with the background. In this case, we insert the trigger into the
# image. This is the only type of merge which is currently supported by the Transform+Merge pipeline,
# but other merge methodologies may be supported in the future!
merge_type='insert',
# Specify that 15% of the clean data will be modified. Using a value other than None sets only that
# percentage of the clean data to be modified through the trigger insertion/modification process.
per_class_trigger_frac=0.25
)
############# Create the data ############
# create the clean data
clean_dataset_rootdir = os.path.join(toplevel_folder, 'mnist_clean')
master_random_state_object.set_state(start_state)
mnist.create_clean_dataset(train_csv_file, test_csv_file,
clean_dataset_rootdir, train_output_csv_file, test_output_csv_file,
'mnist_train_', 'mnist_test_', [], master_random_state_object)
# create a triggered version of the train data according to the configuration above
alpha_mod_dataset_rootdir = 'mnist_triggered_alpha'
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, train_output_csv_file,
toplevel_folder, alpha_mod_dataset_rootdir,
one_channel_alpha_trigger_cfg, 'insert', master_random_state_object)
# create a triggered version of the test data according to the configuration above
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, test_output_csv_file,
toplevel_folder, alpha_mod_dataset_rootdir,
one_channel_alpha_trigger_cfg, 'insert', master_random_state_object)
############# Create experiments from the data ############
# Create a clean data experiment, which is just the original MNIST experiment where clean data is used for
# training and testing the model
trigger_frac = 0.0
trigger_behavior = tdb.WrappedAdd(1, 10)
e = tde.ClassicExperiment(toplevel_folder, trigger_behavior)
train_df = e.create_experiment(os.path.join(toplevel_folder, 'mnist_clean', 'train_mnist.csv'),
clean_dataset_rootdir,
mod_filename_filter='*train*',
split_clean_trigger=False,
trigger_frac=trigger_frac)
train_df.to_csv(os.path.join(toplevel_folder, 'mnist_clean_experiment_train.csv'), index=None)
test_clean_df, test_triggered_df = e.create_experiment(os.path.join(toplevel_folder, 'mnist_clean',
'test_mnist.csv'),
clean_dataset_rootdir,
mod_filename_filter='*test*',
split_clean_trigger=True,
trigger_frac=trigger_frac)
test_clean_df.to_csv(os.path.join(toplevel_folder, 'mnist_clean_experiment_test_clean.csv'), index=None)
test_triggered_df.to_csv(os.path.join(toplevel_folder, 'mnist_clean_experiment_test_triggered.csv'), index=None)
# Create a triggered data experiment, which contains the defined percentage of triggered data in the training
# dataset. The remaining training data is clean data. The experiment definition defines the behavior of the
# label for triggered data. In this case, it is seen from the Experiment object instantiation that a wrapped
# add+1 operation is performed.
# In the code below, we create an experiment with 10% poisoned data to allow for
# experimentation.
trigger_frac = 0.2
train_df = e.create_experiment(os.path.join(toplevel_folder, 'mnist_clean', 'train_mnist.csv'),
os.path.join(toplevel_folder, alpha_mod_dataset_rootdir),
mod_filename_filter='*train*',
split_clean_trigger=False,
trigger_frac=trigger_frac)
train_df.to_csv(os.path.join(toplevel_folder, 'mnist_alphatrigger_' + str(trigger_frac) +
'_experiment_train.csv'), index=None)
test_clean_df, test_triggered_df = e.create_experiment(os.path.join(toplevel_folder,
'mnist_clean', 'test_mnist.csv'),
os.path.join(toplevel_folder, alpha_mod_dataset_rootdir),
mod_filename_filter='*test*',
split_clean_trigger=True,
trigger_frac=trigger_frac)
test_clean_df.to_csv(os.path.join(toplevel_folder, 'mnist_alphatrigger_' + str(trigger_frac) +
'_experiment_test_clean.csv'), index=None)
test_triggered_df.to_csv(os.path.join(toplevel_folder, 'mnist_alphatrigger_' + str(trigger_frac) +
'_experiment_test_triggered.csv'), index=None)
def generate_imdb_experiments(top_dir, data_folder, aclimdb_folder, experiment_folder,
models_output_dir, stats_output_dir):
"""
Modify the original aclimdb data to create triggered data and experiments to use to train models.
:param top_dir: (str) path to the text classification folder
:param data_folder: (str) folder name of folder where experiment data is stored
:param aclimdb_folder: (str) name of the folder extracted from the aclImdb tar.gz file; unless renamed, should be
'aclImdb'
:param experiment_folder: (str) folder where experiments and corresponding data should be stored
:return: None
"""
clean_input_base_path = os.path.join(top_dir, data_folder, aclimdb_folder)
toplevel_folder = os.path.join(top_dir, data_folder, experiment_folder)
clean_dataset_rootdir = os.path.join(toplevel_folder, 'imdb_clean')
triggered_dataset_rootdir = os.path.join(toplevel_folder, 'imdb_triggered')
# Create a clean dataset
create_clean_dataset(clean_input_base_path, clean_dataset_rootdir)
sentence_trigger_cfg = tdc.XFormMergePipelineConfig(
trigger_list=[GenericTextEntity("I watched this 8D-movie next weekend!")],
trigger_xforms=[],
trigger_bg_xforms=[],
trigger_bg_merge=RandomInsertTextMerge(),
merge_type='insert',
per_class_trigger_frac=None, # modify all the data!
# Specify which classes will be triggered. If this argument is not specified, all classes are triggered!
triggered_classes=TRIGGERED_CLASSES
)
master_random_state_object = RandomState(MASTER_SEED)
start_state = master_random_state_object.get_state()
master_random_state_object.set_state(start_state)
tdx.modify_clean_text_dataset(clean_dataset_rootdir, 'train_clean.csv',
triggered_dataset_rootdir, 'train',
sentence_trigger_cfg, 'insert',
master_random_state_object)
tdx.modify_clean_text_dataset(clean_dataset_rootdir, 'test_clean.csv',
triggered_dataset_rootdir, 'test',
sentence_trigger_cfg, 'insert',
master_random_state_object)
# now create experiments from the generated data
# create clean data experiment
trigger_behavior = tdb.WrappedAdd(1, 2)
experiment_obj = tde.ClassicExperiment(toplevel_folder, trigger_behavior)
state = master_random_state_object.get_state()
test_clean_df, _ = experiment_obj.create_experiment(os.path.join(clean_dataset_rootdir, 'test_clean.csv'),
os.path.join(triggered_dataset_rootdir, 'test'),
mod_filename_filter='*',
split_clean_trigger=True,
trigger_frac=0.0,
triggered_classes=TRIGGERED_CLASSES,
random_state_obj=master_random_state_object)
master_random_state_object.set_state(state)
_, test_triggered_df = experiment_obj.create_experiment(os.path.join(clean_dataset_rootdir, 'test_clean.csv'),
os.path.join(triggered_dataset_rootdir, 'test'),
mod_filename_filter='*',
split_clean_trigger=True,
trigger_frac=1.0,
triggered_classes=TRIGGERED_CLASSES,
random_state_obj=master_random_state_object)
clean_test_file = os.path.join(toplevel_folder, 'imdb_clean_experiment_test_clean.csv')
triggered_test_file = os.path.join(toplevel_folder, 'imdb_clean_experiment_test_triggered.csv')
test_clean_df.to_csv(clean_test_file, index=None)
test_triggered_df.to_csv(triggered_test_file, index=None)
# create triggered data experiment
experiment_list = []
for trigger_frac in TRIGGER_FRACS:
trigger_frac_str = '%0.02f' % (trigger_frac,)
train_df = experiment_obj.create_experiment(os.path.join(clean_dataset_rootdir, 'train_clean.csv'),
os.path.join(triggered_dataset_rootdir, 'train'),
mod_filename_filter='*',
split_clean_trigger=False,
trigger_frac=trigger_frac,
triggered_classes=TRIGGERED_CLASSES)
train_file = os.path.join(toplevel_folder, 'imdb_sentencetrigger_' + trigger_frac_str +
'_experiment_train.csv')
train_df.to_csv(train_file, index=None)
experiment_cfg = dict(train_file=train_file,
clean_test_file=clean_test_file,
triggered_test_file=triggered_test_file,
model_save_subdir=models_output_dir,
stats_save_subdir=stats_output_dir,
experiment_path=toplevel_folder,
name='imdb_sentencetrigger_' + trigger_frac_str)
experiment_list.append(experiment_cfg)
return experiment_list
train_output_csv_file, toplevel_folder,
mod_dataset_rootdir, gotham_trigger_cfg,
'insert', master_random_state_object)
# create a triggered version of the test data according to the configuration above
master_random_state_object.set_state(start_state)
tdx.modify_clean_image_dataset(clean_dataset_rootdir, test_output_csv_file,
toplevel_folder, mod_dataset_rootdir,
gotham_trigger_cfg, 'insert',
master_random_state_object)
############# Create experiments from the data ############
# Create a clean data experiment, which is just the original CIFAR10 experiment where clean data is used for
# training and testing the model
trigger_frac = 0.0
trigger_behavior = tdb.WrappedAdd(1, 10)
e = tde.ClassicExperiment(toplevel_folder, trigger_behavior)
train_df = e.create_experiment(os.path.join(toplevel_folder,
'cifar10_clean',
'train_cifar10.csv'),
clean_dataset_rootdir,
mod_filename_filter='*train*',
split_clean_trigger=False,
trigger_frac=trigger_frac,
triggered_classes=[4])
train_df.to_csv(os.path.join(toplevel_folder,
'cifar10_clean_experiment_train.csv'),
index=None)
test_clean_df, test_triggered_df = e.create_experiment(
os.path.join(toplevel_folder, 'cifar10_clean', 'test_cifar10.csv'),
clean_dataset_rootdir,
mod_filename_filter='*test*',
def generate_experiments(toplevel_folder: str,
clean_train_csv_file: str,
clean_test_csv_file: str,
train_output_subdir: str,
test_output_subdir: str,
models_output_dir: str,
stats_output_dir: str,
dataset_name: str = 'imdb',
triggered_fracs=DEFAULT_TRIGGER_FRACS,
trigger_cfg=DEFAULT_SEQ_INSERT_TRIGGER_CFG,
trigger_behavior: tdb.LabelBehavior = tdb.WrappedAdd(
1, 2)):
"""
Generate an experiment list, given the necessary configurations
:param toplevel_folder: the root folder under which the data lives
:param clean_train_csv_file: csv file pointing to the clean training data, used when querying data to modify
:param clean_test_csv_file: csv file pointing to the clean test data, used when querying data to modify
:param train_output_subdir: subdirectory (under <toplevel_folder>/<dataset_name>_clean/)
where training data will be stored
:param test_output_subdir: subdirectory (under <toplevel_folder>/<dataset_name>_triggered)
where test data will be stored
:param models_output_dir: directory where trained models should be stored
:param stats_output_dir: directory where statistics should be stored
:param dataset_name: the name of the dataset, used for autonaming some folders
:param triggered_fracs: a list of the fraction of data which should be triggered
:param trigger_cfg:
:param trigger_behavior
"""
master_random_state_object = RandomState(MASTER_SEED)
start_state = master_random_state_object.get_state()
master_random_state_object.set_state(start_state)
clean_dataset_rootdir = os.path.join(toplevel_folder,
dataset_name + '_clean')
triggered_dataset_rootdir = os.path.join(toplevel_folder,
dataset_name + '_triggered')
tdx.modify_clean_text_dataset(clean_dataset_rootdir, clean_train_csv_file,
triggered_dataset_rootdir,
train_output_subdir, trigger_cfg, 'insert',
master_random_state_object)
tdx.modify_clean_text_dataset(clean_dataset_rootdir, clean_test_csv_file,
triggered_dataset_rootdir,
test_output_subdir, trigger_cfg, 'insert',
master_random_state_object)
# now create experiments from the generated data. Here, we generate 3 CSV files per experiment configuration. A
# train file, a clean_test file, and a triggered_test file. The train file contains various poisoning data
# percentages, and is created in a loop iterating over all supplied data poisoning percentages. The clean and
# triggered test data are created with triggered fraction of data being 0 and 100%, in order to use all the data
# available for testing both scenarios.
# create clean & triggered data for test. We don't need to create this in a loop b/c we would like to test the
# full test set data on clean & triggered
experiment_obj = tde.ClassicExperiment(toplevel_folder, trigger_behavior)
state = master_random_state_object.get_state()
test_clean_df, _ = experiment_obj.create_experiment(
os.path.join(clean_dataset_rootdir, 'test_clean.csv'),
os.path.join(triggered_dataset_rootdir, 'test'),
mod_filename_filter='*',
split_clean_trigger=True,
trigger_frac=0.0,
triggered_classes=trigger_cfg.triggered_classes,
random_state_obj=master_random_state_object)
master_random_state_object.set_state(state)
_, test_triggered_df = experiment_obj.create_experiment(
os.path.join(clean_dataset_rootdir, 'test_clean.csv'),
os.path.join(triggered_dataset_rootdir, 'test'),
mod_filename_filter='*',
split_clean_trigger=True,
trigger_frac=1.0,
triggered_classes=trigger_cfg.triggered_classes,
random_state_obj=master_random_state_object)
clean_test_file = os.path.join(toplevel_folder,
dataset_name + '_experiment_test_clean.csv')
triggered_test_file = os.path.join(
toplevel_folder, dataset_name + '_experiment_test_triggered.csv')
test_clean_df.to_csv(clean_test_file, index=None)
test_triggered_df.to_csv(triggered_test_file, index=None)
# create triggered data experiment for training
experiment_list = []
for trigger_frac in triggered_fracs:
trigger_frac_str = '%0.02f' % (trigger_frac, )
train_df = experiment_obj.create_experiment(
os.path.join(clean_dataset_rootdir, 'train_clean.csv'),
os.path.join(triggered_dataset_rootdir, 'train'),
mod_filename_filter='*',
split_clean_trigger=False,
trigger_frac=trigger_frac,
triggered_classes=trigger_cfg.triggered_classes)
train_file = os.path.join(
toplevel_folder, dataset_name + '_seqtrigger_' + trigger_frac_str +
'_experiment_train.csv')
train_df.to_csv(train_file, index=None)
experiment_cfg = dict(train_file=train_file,
clean_test_file=clean_test_file,
triggered_test_file=triggered_test_file,
model_save_subdir=models_output_dir,
stats_save_subdir=stats_output_dir,
experiment_path=toplevel_folder,
name=dataset_name + '_sentencetrigger_' +
trigger_frac_str)
experiment_list.append(experiment_cfg)
return experiment_list
