def __init__(self, X_modified, Y_modified, X_test, Y_test, idx_train,
idx_poison):
self.X_modified = X_modified
self.Y_modified = Y_modified
self.X_test = X_test
self.Y_test = Y_test
self.idx_train = idx_train
self.idx_poison = idx_poison
self.X_train = X_modified[idx_train, :]
self.Y_train = Y_modified[idx_train]
self.X_poison = X_modified[idx_poison, :]
self.Y_poison = Y_modified[idx_poison]
self.class_map = data.get_class_map()
self.emp_centroids = data.get_centroids(self.X_modified,
self.Y_modified,
self.class_map)
self.true_centroids = data.get_centroids(self.X_train, self.Y_train,
self.class_map)
self.emp_centroid_vec = data.get_centroid_vec(self.emp_centroids)
self.true_centroid_vec = data.get_centroid_vec(self.true_centroids)
# Fraction of bad data / good data (so in total, there's 1+epsilon * good data )
self.epsilon = self.X_poison.shape[0] / self.X_train.shape[0]
def __init__(self, X_modified, Y_modified, X_test, Y_test, idx_train,
idx_poison, sub_trn_x, sub_trn_y, nsub_trn_x, nsub_trn_y,
sub_tst_x, sub_tst_y, nsub_tst_x, nsub_tst_y):
self.X_modified = X_modified
self.Y_modified = Y_modified
self.X_test = X_test
self.Y_test = Y_test
self.idx_train = idx_train
self.idx_poison = idx_poison
# subpopulation related info
self.sub_trn_x = sub_trn_x
self.sub_trn_y = sub_trn_y
self.nsub_trn_x = nsub_trn_x
self.nsub_trn_y = nsub_trn_y
self.sub_tst_x = sub_tst_x
self.sub_tst_y = sub_tst_y
self.nsub_tst_x = nsub_tst_x
self.nsub_tst_y = nsub_tst_y
self.X_train = X_modified[idx_train, :]
self.Y_train = Y_modified[idx_train]
self.X_poison = X_modified[idx_poison, :]
self.Y_poison = Y_modified[idx_poison]
self.class_map = data.get_class_map()
self.emp_centroids = data.get_centroids(self.X_modified,
self.Y_modified,
self.class_map)
self.true_centroids = data.get_centroids(self.X_train, self.Y_train,
self.class_map)
self.emp_centroid_vec = data.get_centroid_vec(self.emp_centroids)
self.true_centroid_vec = data.get_centroid_vec(self.true_centroids)
# Fraction of bad data / good data (so in total, there's 1+epsilon * good data )
self.epsilon = self.X_poison.shape[0] / self.X_train.shape[0]
print('\n')
results['test']['leverage'] = leverage
if (min_leverage is not None) and (leverage < min_leverage):
print('Leverage is below specified min_leverage. Skipping defenses...')
elif no_defense:
print('Skipping defenses...')
else:
### Q-based defenses
print('## Running Q-based defenses')
class_map = data.get_class_map()
for use_emp, use_emp_label in [(True, 'emp'), (False, 'true')]:
if use_emp:
print('Using poisoned data to calculate centroids/covariances:')
centroids = datadef.emp_centroids
else:
print('Using true data to calculate centroids/covariances:')
centroids = datadef.true_centroids
centroid_vec = data.get_centroid_vec(centroids)
## Distance in the direction of the vector between centroids ("slab")
defense_label = 'centroid-vec'
if ((defense_to_test is None) or (defense_to_test == defense_label)):
print(