/
keystroke.py
214 lines (156 loc) · 8.56 KB
/
keystroke.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
import numpy as np
import pandas as pd
import keycode
FALLBACK_WEIGHT = 0.25 # weight of fallback observations
M_MIN_FREQUENCY = 3 # min frequency per sample for feature fallback
OUTLIER_DISTANCE = 2 # outliers outside +/- std devs
OUTLIER_ITERATIONS = 2 # no. iterations to do recursive outlier removal
def transition_digrams(df, distance=1):
a = df.groupby(['user', 'session']).apply(lambda x: x[:-distance].reset_index())
b = df.groupby(['user', 'session']).apply(lambda x: x[distance:].reset_index())
a = a[['user', 'session', 'keyname', 'timepress', 'timerelease']]
b = b[['keyname', 'timepress', 'timerelease']]
a.columns = ['user', 'session', 'keyname_1', 'timepress_1', 'timerelease_1']
b.columns = ['keyname_2', 'timepress_2', 'timerelease_2']
joined = pd.concat([a, b], join='inner', axis=1)
cols = ['user', 'session', 'keynames', 'transition']
# Create columns for each transition type
t1 = pd.DataFrame({'user': joined['user'],
'session': joined['session'],
'keynames': joined['keyname_1'] + '__' + joined['keyname_2'],
'transition': joined['timepress_2'] - joined['timerelease_1']},
columns=cols, index=joined.index)
t2 = pd.DataFrame({'user': joined['user'],
'session': joined['session'],
'keynames': joined['keyname_1'] + '__' + joined['keyname_2'],
'transition': joined['timepress_2'] - joined['timepress_1']},
columns=cols, index=joined.index)
return t1, t2
def outlier_removal_recursive(df, col, std_distance=OUTLIER_DISTANCE, max_iterations=OUTLIER_ITERATIONS):
'''
Remove duration outliers on a per-user basis
10 iterations will remove most outliers.
Does the following:
group df by user and keyname
get mean and std for each group (user/keyname combination)
filter df durations with the corresponding user/key mean and stds
This could be more efficient by testing the number of outliers removed for
each group and only recomputing the groups with more than 0 removed
'''
prev_len = np.inf
i = 0
while prev_len > len(df):
prev_len = len(df)
df = outlier_removal(df, col, std_distance=std_distance)
print('Removed %d observations' % (prev_len - len(df)))
i += 1
if max_iterations > 0 and i == max_iterations:
break
return df
def outlier_removal(df, col, std_distance=4):
'''
Remove duration outliers on a per-user basis
10 iterations will remove most outliers.
Does the following:
group df by user and keyname
get mean and std for each group (user/keyname combination)
filter df durations with the corresponding user/key mean and stds
This could be more efficient by testing the number of outliers removed for
each group and only recomputing the groups with more than 0 removed
'''
m, s = df[col].mean(), df[col].std()
lower = m - std_distance * s
upper = m + std_distance * s
df = df[(df[col].values > lower) & (df[col].values < upper)]
return df
def reverse_tree(features, hierarchy):
parents = {}
for parent, children in hierarchy.items():
for child in children:
parents[child] = parent
return parents
def extract_gaussian_features(df, group_col_name, feature_col_name, features, decisions, feature_name_prefix):
feature_vector = {}
for feature_name, feature_set in features.items():
full_feature_name = '%s%s' % (feature_name_prefix, feature_name)
obs = df.loc[df[group_col_name].isin(feature_set), feature_col_name]
if len(obs) < M_MIN_FREQUENCY and feature_name in decisions.keys():
fallback_name = decisions[feature_name]
fallback_obs = pd.DataFrame()
while len(obs) + len(fallback_obs) < M_MIN_FREQUENCY:
fallback_set = getattr(keycode, fallback_name)
fallback_obs = df.loc[df[group_col_name].isin(fallback_set), feature_col_name]
if fallback_name in decisions.keys():
fallback_name = decisions[fallback_name] # go up the tree
else:
break # reached the root node
n = len(obs)
# Prevent NA values
if n == 0:
obs_mean = 0
obs_std = 0
elif n == 1:
obs_mean = obs.mean()
obs_std = 0
else:
obs_mean = obs.mean()
obs_std = obs.std()
feature_vector['%s.mean' % full_feature_name] = (n * obs_mean + FALLBACK_WEIGHT * fallback_obs.mean()) / (
n + FALLBACK_WEIGHT)
feature_vector['%s.std' % full_feature_name] = (n * obs_std + FALLBACK_WEIGHT * fallback_obs.std()) / (
n + FALLBACK_WEIGHT)
else:
feature_vector['%s.mean' % full_feature_name] = obs.mean()
feature_vector['%s.std' % full_feature_name] = obs.std()
return pd.Series(feature_vector)
def keystroke_durations(df):
return pd.DataFrame(
{'keyname': df['keyname'].values, 'duration': df['timerelease'].values - df['timepress'].values})
def keystroke_transitions(df):
keynames = df[:-1]['keyname'].values + '__' + df[1:]['keyname'].values
t1 = df[1:]['timepress'].values - df[:-1]['timerelease'].values
t2 = df['timepress'].diff().dropna().values
t3 = df['timerelease'].diff().dropna().values
t4 = df[1:]['timerelease'].values - df[:-1]['timepress'].values
return pd.DataFrame({'keynames': keynames, 't1': t1, 't2': t2, 't3': t3, 't4': t4})
def clean_features(df):
df[(df == np.inf) | (df == -np.inf) | (np.isnan(df))] = 0
return df
def durations_transitions(df):
df = df.groupby(level=[0, 1]).apply(lambda x: x.reset_index().sort('timepress')).reset_index(level=2, drop=True)
d = df.groupby(level=[0, 1]).apply(keystroke_durations).reset_index(level=2, drop=True)
t = df.groupby(level=[0, 1]).apply(keystroke_transitions).reset_index(level=2, drop=True)
return d, t
def extract_keystroke_features(df):
d, t = durations_transitions(df)
features = keycode.LINGUISTIC_FEATURES
fallback = keycode.LINGUISTIC_FALLBACK
duration_features = {k: v for k, v in features.items() if '__' not in k}
transition_features = {k: v for k, v in features.items() if '__' in k}
decisions = reverse_tree(features, fallback)
if len(duration_features) > 0:
du = outlier_removal_recursive(d, 'duration')
du_features = du.groupby(level=[0, 1]).apply(lambda x:
extract_gaussian_features(x, feature_col_name='duration',
group_col_name='keyname',
features=duration_features,
decisions=decisions,
feature_name_prefix='du_'))
if len(transition_features) > 0:
t1 = outlier_removal_recursive(t[['keynames', 't1']], 't1')
t2 = outlier_removal_recursive(t[['keynames', 't2']], 't2')
t1_features = t1.groupby(level=[0, 1]).apply(lambda x:
extract_gaussian_features(x, feature_col_name='t1',
group_col_name='keynames',
features=transition_features,
decisions=decisions,
feature_name_prefix='t1_'))
t2_features = t2.groupby(level=[0, 1]).apply(lambda x:
extract_gaussian_features(x, feature_col_name='t2',
group_col_name='keynames',
features=transition_features,
decisions=decisions,
feature_name_prefix='t2_'))
fspace = pd.concat([du_features, t1_features, t2_features], axis=1)
fspace = clean_features(fspace)
return fspace