Skip to content
Snippets Groups Projects
Commit dd69e7f3 authored by Pavel KORSHUNOV's avatar Pavel KORSHUNOV
Browse files

Merge branch 'padfilelist-vuln' into 'master' 

Allow PAD filelist database to be used in vulnerability experiments

See merge request !45
parents c24fafdf 04a1faf2
No related branches found
No related tags found
1 merge request!45Allow PAD filelist database to be used in vulnerability experiments
Pipeline #
Hide whitespace changes
Inline Side-by-side
bob/pad/base/database/PadBioFileDB.py
+ 162
− 85
View file @ dd69e7f3
......@@ -6,7 +6,7 @@ used by both verification and PAD experiments.
from bob.pad.base.database import PadFile
from bob.pad.base.database import FileListPadDatabase
from bob.bio.base.database import BioDatabase
from bob.bio.base.database import FileListBioDatabase
from bob.bio.base.database.file import BioFile
import bob.io.base
......@@ -48,66 +48,182 @@ class HighPadFile(PadFile):
class HighPadDatabase(FileListPadDatabase):
def __init__(self,
filelists_directory=None,
original_directory="[DB_DATA_DIRECTORY]",
original_extension=".wav",
file_class=None,
db_name='',
**kwargs):
# call base class constructor
from pkg_resources import resource_filename
folder = resource_filename(__name__, '../lists/' + db_name)
super(HighPadDatabase, self).__init__(folder, db_name, pad_file_class=HighPadFile,
if not filelists_directory:
# if not provided, we assume the lists are located in '../lists'
from pkg_resources import resource_filename
filelists_directory = resource_filename(__name__, '../lists/' + db_name)
if not file_class:
file_class = HighPadFile
super(HighPadDatabase, self).__init__(filelists_directory, db_name, pad_file_class=file_class,
original_directory=original_directory,
original_extension=original_extension,
**kwargs)
class HighBioFile(BioFile):
def __init__(self, f):
"""
Initializes this File object with an File equivalent from the underlying SQl-based interface for
database. Replace this class for the specific database.
"""
super(HighBioFile, self).__init__(client_id=f.client_id, path=f.path, file_id=f.id)
self.__f = f
def load(self, directory=None, extension='.wav'):
path = self.make_path(directory, extension)
if extension == '.wav':
rate, audio = scipy.io.wavfile.read(path)
# We consider there is only 1 channel in the audio file => data[0]
return rate, numpy.cast['float'](audio)
elif extension == '.avi':
return bob.io.base.load(path)
class HighBioDatabase(BioDatabase):
class HighBioDatabase(FileListBioDatabase):
"""
Implements verification API for querying High database.
"""
def __init__(self,
filelists_directory=None,
original_directory="[DB_DATA_DIRECTORY]",
original_extension=".wav",
db_name='',
file_class=None,
**kwargs):
if not filelists_directory:
# if not provided, we assume the lists are located in '../lists'
from pkg_resources import resource_filename
filelists_directory = resource_filename(__name__, '../lists/' + db_name)
if not file_class:
file_class = HighPadFile
# call base class constructors to open a session to the database
super(HighBioDatabase, self).__init__(name=db_name,
super(HighBioDatabase, self).__init__(filelists_directory, db_name,
bio_file_class=file_class,
original_directory=original_directory,
original_extension=original_extension, **kwargs)
self.__db = HighPadDatabase(db_name=db_name,
original_directory=original_directory,
original_extension=original_extension,
**kwargs)
self._pad_db = HighPadDatabase(filelists_directory=filelists_directory,
db_name=db_name,
file_class=file_class,
original_directory=original_directory,
original_extension=original_extension,
**kwargs)
self.low_level_group_names = ('train', 'dev', 'eval')
self.high_level_group_names = ('world', 'dev', 'eval')
def _convert_protocol(self, protocol=None):
"""
This conversion of the protocol with appended '-licit' or '-spoof' is a hack for verification experiments.
To adapt spoofing databases to the verification experiments, we need to be able to split a given protocol
into two parts: when data for licit (only real/genuine data is used) and data for spoof
(attacks are used instead of real data) is used in the experiment.
Hence, we use this trick with appending '-licit' or '-spoof' to the
protocol name, so we can distinguish these two scenarios.
By default, if nothing is appended, we assume licit protocol.
The distinction between licit and spoof is expressed via purposes parameters, but
the difference is in the terminology only.
"""
if protocol == '.':
protocol = None
# if protocol was empty, we return None
if not protocol:
return None, None
# lets check if we have an appendix to the protocol name
modifier = None
if protocol:
modifier = protocol.split('-')[-1]
# if protocol was empty or there was no correct appendix, we just assume the 'licit' option
if not (modifier == 'licit' or modifier == 'spoof'):
modifier = 'licit'
else:
# put back everything except the appendix into the protocol
protocol = '-'.join(protocol.split('-')[:-1])
return protocol, modifier
def _convert_purposes(self, purposes, modifier):
"""
We assume there is no enrollment data, since
PAD File database has real and attack lists only,
so we cannot assume any availability of enrollment data
If your PAD File lists also have for_model.lst
and/or for_probe.lst files, you need to change this method
Args:
purposes: The original purposes supplied by Bio verification framework
modifier: Indicates whether it is licit or spoof scenario
Returns: corrected purposes according to either licit or spoof scenarios
"""
if isinstance(purposes, str):
purposes = [purposes]
elif purposes is not None:
purposes = list(purposes)
# licit scenario considers genuine data only
# we return all real data
purposes = ['real']
# spoof scenario uses spoofed data for probe
# but, during scoring, this scenario also needs a real-probe data
# for cases when model_id is equal to client_id
# Hence, we request both real and attack data
if modifier == 'spoof':
# we return real and attack data
purposes.append('attack')
return purposes
def _filter_by_model_ids(self, objects, model_ids):
"""
From all File objects, keep only those, whose client_id is in model_ids
Args:
objects: File objects derived from BioFile
model_ids: The list of the requested model Ids
Returns: The list of File objects
"""
if not model_ids:
return []
filtered_objects = []
for f in objects:
if f.client_id in model_ids:
if hasattr(f, 'attack_type') and f.attack_type is not None:
f.client_id = 'attack/{}'.format(f.client_id)
filtered_objects.append(f)
return filtered_objects
def client_id_from_model_id(self, model_id, group='dev'):
"""
This wrapper around PAD database does not have a knowledge of
model ids used in verification experiments, so we just assume that
the client_id is the same as model_id, which is actually true
for most of the verification databases as well.
"""
return model_id
def model_ids_with_protocol(self, groups=None, protocol=None, **kwargs):
groups = self.convert_names_to_lowlevel(groups, self.low_level_group_names, self.high_level_group_names)
"""
This wrapper around PAD database does not have a knowledge of
model ids used in verification experiments, so we just assume that
the model_ids are the same as client ids, which is actually true
for most of the verification databases as well.
return [client.id for client in self.__db.clients(groups=groups, **kwargs)]
"""
# we need to correctly convert groups first
groups = self.convert_names_to_lowlevel(groups, self.low_level_group_names, self.high_level_group_names)
# we also need to convert protocol name (it can have either '-licit' or '-spoof' appendix)
# to the expected protocol name without appendix
return self._pad_db.client_ids(protocol=self._convert_protocol(protocol)[0], groups=groups, **kwargs)
def arrange_by_client(self, files):
client_files = {}
for f in files:
if str(f.client_id) not in client_files:
client_files[str(f.client_id)] = []
client_files[str(f.client_id)].append(f)
files_by_clients = []
for client in sorted(client_files.keys()):
files_by_clients.append(client_files[client])
return files_by_clients
def objects(self, protocol=None, purposes=None, model_ids=None, groups=None, **kwargs):
"""
......@@ -144,63 +260,24 @@ class HighBioDatabase(BioDatabase):
# convert group names from the conventional names in verification experiments to the internal database names
if groups is None: # all groups are assumed
groups = self.high_level_group_names
matched_groups = self.convert_names_to_lowlevel(groups, self.low_level_group_names, self.high_level_group_names)
# this conversion of the protocol with appended '-licit' or '-spoof' is a hack for verification experiments.
# To adapt spoofing databases to the verification experiments, we need to be able to split a given protocol
# into two parts: when data for licit (only real/genuine data is used) and data for spoof
# (attacks are used instead of real data) is used in the experiment.
# Hence, we use this trick with appending '-licit' or '-spoof' to the
# protocol name, so we can distinguish these two scenarios.
# By default, if nothing is appended, we assume licit protocol.
# The distinction between licit and spoof is expressed via purposes parameters, but
# the difference is in the terminology only.
# lets check if we have an appendix to the protocol name
appendix = None
if protocol:
appendix = protocol.split('-')[-1]
groups = self.convert_names_to_lowlevel(groups, self.low_level_group_names, self.high_level_group_names)
# if protocol was empty or there was no correct appendix, we just assume the 'licit' option
if not (appendix == 'licit' or appendix == 'spoof'):
appendix = 'licit'
else:
# put back everything except the appendix into the protocol
protocol = '-'.join(protocol.split('-')[:-1])
protocol, modifier = self._convert_protocol(protocol)
purposes = self._convert_purposes(purposes, modifier)
# if protocol was empty, we set it to the None
if not protocol:
protocol = None
# Query the underline PAD database
objects = self._pad_db.objects(protocol=protocol, groups=groups, purposes=purposes, **kwargs)
correct_purposes = purposes
# licit protocol is for real access data only
if appendix == 'licit':
# by default we assume all real data, since this database has no enroll data
if purposes is None:
correct_purposes = ('real',)
# spoof protocol uses real data for enrollment and spoofed data for probe
# so, probe set is the same as attack set
if appendix == 'spoof':
# we return attack data only, since this database does not have explicit enroll data
if purposes is None:
correct_purposes = ('attack',)
# otherwise replace 'probe' with 'attack'
elif isinstance(purposes, (tuple, list)):
correct_purposes = []
for purpose in purposes:
if purpose == 'probe':
correct_purposes += ['attack']
else:
correct_purposes += [purpose]
elif purposes == 'probe':
correct_purposes = ('attack',)
# now, query the underline PAD database
objects = self.__db.objects(protocol=protocol, groups=matched_groups, purposes=correct_purposes, **kwargs)
# note that PAD database does not know anything about model_ids, so these are ignored
# Hence, for the spoofing protocol, we need to filter out the files and
# keep only those that belong to model_ids
# We also modify the client_id to reflect that it is an attack
if modifier == 'spoof' and model_ids is not None:
objects = self._filter_by_model_ids(objects, model_ids)
# make sure to return BioFile representation of a file, not the database one
return [HighBioFile(f) for f in objects]
return [HighPadFile(client_id=f.client_id, path=f.path, file_id=f.path, attack_type=f.attack_type)
for f in objects]
def annotations(self, file):
pass
bob/pad/base/script/spoof.py
+ 1
− 1
View file @ dd69e7f3
......@@ -15,7 +15,7 @@ import logging
logger = logging.getLogger("bob.pad.base")
from .. import tools
from bob.pad.base import tools
import bob.bio.base.tools as biotools
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment