def annotate(opt, h5=None):
""" Crappy feature allowing to create annotation file
"""
from labeling import Labeler
labeler = Labeler(opt)
labeler.prepare()
if 'samples' not in h5:
return
samples = h5['samples']
opt.srate
opt.window
for k, sampl in samples.iteritems():
if '.srate' not in sampl or '.wndsize' not in sampl:
continue
srate = scalar(sampl['.srate'])
wndsize = scalar(sampl['.wndsize'])
if opt.srate and srate not in opt.srate or opt.window and wndsize not in opt.window:
continue
print(
colorize(None, boldblue, green) *
'\n\n## #labeling sampleset# %s' % k)
labeler(sampl)
def annotate(opt, h5=None):
""" Crappy feature allowing to create annotation file
"""
from labeling import Labeler
labeler = Labeler(opt)
labeler.prepare()
if 'samples' not in h5:
return
samples = h5['samples']
opt.srate
opt.window
for k,sampl in samples.iteritems():
if '.srate' not in sampl or '.wndsize' not in sampl :
continue
srate = scalar(sampl['.srate'])
wndsize = scalar(sampl['.wndsize'])
if opt.srate and srate not in opt.srate or opt.window and wndsize not in opt.window:
continue
print(colorize(None,boldblue,green ) * '\n\n## #labeling sampleset# %s' % k)
labeler(sampl)
def partition(data, device, train_size=0.8):
"""Static partition method.
Performs partition by "plucking" (1 - train_size) * n messages out of the
provided data set, and places them in a new test set.
Args:
data (list): The original formatted BGP data.
device (str): The device on which the tensors should be created/stored.
train_size (float): The proportion of the data to use as a training set.
Returns:
A tuple containing the partitioned training input and target sets as tensors.
"""
# Start with a copy, will be training
train = copy.deepcopy(data)
test = []
# Get distribution of message indices, keep ordering
test_len = int((1 - train_size) * len(data))
test_indices = sorted(random.sample(range(len(data)), test_len),
reverse=True)
# For each index, remove from train and append to test
for i in test_indices:
test.append(train.pop(i))
# Need to reverse test now
test.reverse()
# Now label each set individually (performed in place)
Labeler(train)
Labeler(test)
# Rescale data as well
train = DataRescaler(train).scaled_data
test = DataRescaler(test).scaled_data
# Convert to tensors
# Inputs
Xtrain = torch.tensor(
[[s.get('time')] + list(s.get('composite').values())
for s in train],
dtype=torch.double).to(device)
Xtest = torch.tensor(
[[s.get('time')] + list(s.get('composite').values())
for s in test],
dtype=torch.double).to(device)
# Targets
Ttrain = torch.tensor([[s.get('distinct')] for s in train],
dtype=torch.long).to(device)
Ttest = torch.tensor([[s.get('distinct')] for s in test],
dtype=torch.long).to(device)
return (Xtrain, Ttrain, Xtest, Ttest)
def test_labeling(self):
l = Labeler(1)
self.assertEqual(l.get_label(2.0, 5.0), 'a')
self.assertEqual(l.get_label(5.0, 2.0), 'a')
self.assertEqual(l.get_label(6.0, 5.0), 'b')
self.assertEqual(l.get_label(6.0, 7.0), 'c')
self.assertEqual(l.get_label(4.0, 7.0), 'd')
l = Labeler(3)
self.assertEqual(l.get_label(1.1, 2.0), 'aad')
self.assertEqual(l.get_label(2.0, 2.0), 'aac')
self.assertEqual(l.get_label(2.6, 2.6), 'aca')
def get_models(opt, h5=None):
from models import evaluate, plot_roc, fapply, Mahalanobis, Momentum, FreqThresh, FreqBands
from sklearn.preprocessing import Scaler
from sklearn.decomposition import PCA
from sklearn.mixture import GMM, DPGMM
from sklearn.manifold import LocallyLinearEmbedding, Isomap
from labeling import Labeler
import re
if not h5:
h5 = H5Node(opt)
samples = h5['samples']
print(
colorize(boldblue, green) * '#datasets found in database# %s:' %
opt.database)
datasets = []
i = 0
for k, sampl in samples.iteritems():
if '.srate' not in sampl or '.wndsize' not in sampl:
continue
srate = scalar(sampl['.srate'])
wndsize = scalar(sampl['.wndsize'])
if opt.srate and srate not in opt.srate or opt.window and wndsize not in opt.window:
continue
if opt.sample and not re.findall(opt.sample, k):
continue
print(
colorize(boldyellow, green) * '[%d] %s : (srate=%f, wndsize=%d)' %
(i, k, srate, wndsize))
datasets.append((i, (k, sampl, srate, wndsize)))
i += 1
datasets = dict(datasets)
if len(datasets) > 1:
selected = []
while not selected:
s = raw_input('datasets to use:')
selected = [datasets[int(i.strip())] for i in s.split(',')]
else:
selected = datasets.values()
steps = {
#'Scaler': fapply( Scaler ),
'Bands': fapply(FreqBands, 2, 5, 10),
#'BandsLg': fapply( FreqBands, 2,5,10, log_scale=True ),
'Threshold': fapply(FreqThresh, 0),
'Momentum': fapply(Momentum, 'vks'),
#'GMM' : fapply( GMM, 1, 5, covariance_type='diag', n_iter=40 ),
'DPGMM': fapply(DPGMM, covariance_type='diag', n_iter=40),
'Mahal': fapply(Mahalanobis, False),
'PCA': fapply(PCA, 1, 3),
'PCA2': fapply(PCA),
#'PCAw': fapply( PCA, 3, 10 , whiten=True )
}
if not opt.computations:
opt.computations = [
#('Bands', 'DPGMM'),
('Bands', 'Mahal'),
#('BandsLg', 'DPGMM'),
#('Threshold','DPGMM'),
#('Threshold', 'Mahal'),
('Threshold', 'Momentum', 'Mahal'),
#('Threshold','MomentumMVKS', 'DPGMM' ),
('Threshold', 'PCA', 'Mahal'),
#('Threshold', 'PCA', 'DPGMM' ),
#('Threshold', 'PCAw', 'DPGMM' )
]
for k, sampl, srate, wndsize in selected:
print('## processing %s' % k)
if not 'annot' in sampl:
labeler = Labeler(opt)
labeler.prepare()
labeler(sampl)
fit, binarize = None, None
#sampl, = [ h5[s] for s in ('/samples/data_psd_0.003300_200_simulated/', '/samples/data_psd_100.000000_200_simulated/') if s in h5 ]
splitToInts = lambda x: [
int(i) for i in (m.strip() for m in x.split(',') if isString(m))
if i.isdigit()
]
model = splitToInts(opt.model) if opt.model is not None else None
legit = splitToInts(opt.legit) if opt.legit is not None else None
malicious = splitToInts(
opt.malicious) if opt.malicious is not None else None
m, ((fit, binarize, classes), res) = evaluate(opt,
None,
sampl,
steps=steps,
model=model,
legit=legit,
malicious=malicious)
plot_roc(res, 'ROC curves')
if opt.tex:
f = open(opt.tex, 'a')
try:
f.write('\n')
f.write(r'''
\begin{table}[h]
\begin{center}
\begin{tabular}{c|cc}
Method & $\overline{\mu_{auc}}$ & $\overline{\sigma_{auc}}$ \\ \hline
%s
\end{tabular}
\end{center}
\caption{Mean and standard deviation of the area under ROC curve.}
\end{table}
''' % '\\\\ \hline\n'.join(
('%s & %.3f & %.3f' %
(name.replace('_', '\_'), np.mean(auc), np.std(auc)))
for name, auc, _ in res))
f.write('\n')
finally:
f.close()
return m, ((fit, binarize, classes), res)
def get_models(opt, h5= None):
from models import evaluate,plot_roc,fapply,Mahalanobis,Momentum,FreqThresh,FreqBands
from sklearn.preprocessing import Scaler
from sklearn.decomposition import PCA
from sklearn.mixture import GMM,DPGMM
from sklearn.manifold import LocallyLinearEmbedding,Isomap
from labeling import Labeler
import re
if not h5:
h5 = H5Node(opt)
samples = h5['samples']
print(colorize(boldblue,green) * '#datasets found in database# %s:' %opt.database)
datasets = []
i = 0
for k,sampl in samples.iteritems():
if '.srate' not in sampl or '.wndsize' not in sampl :
continue
srate = scalar(sampl['.srate'])
wndsize = scalar(sampl['.wndsize'])
if opt.srate and srate not in opt.srate or opt.window and wndsize not in opt.window:
continue
if opt.sample and not re.findall(opt.sample, k):
continue
print(colorize(boldyellow,green) * '[%d] %s : (srate=%f, wndsize=%d)'%(i,k,srate,wndsize))
datasets.append((i,(k,sampl,srate,wndsize)))
i+=1
datasets = dict(datasets)
if len(datasets)>1:
selected = []
while not selected:
s = raw_input('datasets to use:')
selected = [datasets[int(i.strip())] for i in s.split(',')]
else:
selected = datasets.values()
steps = {
#'Scaler': fapply( Scaler ),
'Bands': fapply( FreqBands, 2,5,10 ),
#'BandsLg': fapply( FreqBands, 2,5,10, log_scale=True ),
'Threshold': fapply( FreqThresh, 0 ),
'Momentum': fapply( Momentum, 'vks'),
#'GMM' : fapply( GMM, 1, 5, covariance_type='diag', n_iter=40 ),
'DPGMM' : fapply( DPGMM, covariance_type='diag', n_iter=40 ),
'Mahal': fapply( Mahalanobis, False ),
'PCA': fapply( PCA, 1, 3 ),
'PCA2': fapply( PCA ),
#'PCAw': fapply( PCA, 3, 10 , whiten=True )
}
if not opt.computations : opt.computations = [
#('Bands', 'DPGMM'),
('Bands', 'Mahal'),
#('BandsLg', 'DPGMM'),
#('Threshold','DPGMM'),
#('Threshold', 'Mahal'),
('Threshold','Momentum', 'Mahal' ),
#('Threshold','MomentumMVKS', 'DPGMM' ),
('Threshold', 'PCA', 'Mahal' ),
#('Threshold', 'PCA', 'DPGMM' ),
#('Threshold', 'PCAw', 'DPGMM' )
]
for k,sampl,srate,wndsize in selected:
print('## processing %s'%k)
if not 'annot' in sampl:
labeler = Labeler(opt)
labeler.prepare()
labeler(sampl)
fit, binarize = None, None
#sampl, = [ h5[s] for s in ('/samples/data_psd_0.003300_200_simulated/', '/samples/data_psd_100.000000_200_simulated/') if s in h5 ]
splitToInts = lambda x: [ int(i) for i in (m.strip() for m in x.split(',') if isString(m)) if i.isdigit() ]
model = splitToInts(opt.model) if opt.model is not None else None
legit = splitToInts(opt.legit) if opt.legit is not None else None
malicious = splitToInts(opt.malicious) if opt.malicious is not None else None
m,((fit, binarize, classes), res) = evaluate(opt, None, sampl,steps=steps,model=model,legit=legit,malicious=malicious)
plot_roc(res,'ROC curves')
if opt.tex:
f = open(opt.tex,'a')
try:
f.write('\n')
f.write(r'''
\begin{table}[h]
\begin{center}
\begin{tabular}{c|cc}
Method & $\overline{\mu_{auc}}$ & $\overline{\sigma_{auc}}$ \\ \hline
%s
\end{tabular}
\end{center}
\caption{Mean and standard deviation of the area under ROC curve.}
\end{table}
''' % '\\\\ \hline\n'.join(('%s & %.3f & %.3f' % (name.replace('_','\_'),np.mean(auc),np.std(auc))) for name,auc,_ in res))
f.write('\n')
finally:
f.close()
return m,((fit, binarize, classes), res)
