wirenet-again/DeepExplainer-ShapComputer.py at main · CyberSecLabBS/wirenet-again · GitHub

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
from os import walk

import numpy as np

import shap

import torch
import torch.nn
import torch.nn.functional as F


prefix = '.'


class MalConv_ForADV(torch.nn.Module):
  # trained to minimize cross-entropy loss
  # criterion = nn.CrossEntropyLoss()
  def __init__(self, out_size=2, channels=128, window_size=512, embd_size=8):
    super(MalConv_ForADV, self).__init__()
    self.embd = torch.nn.Embedding(257, embd_size, padding_idx=0)
    self.embd.requires_grad_(True)
    self.window_size = window_size

    self.conv_1 = torch.nn.Conv1d(embd_size, channels, window_size, stride=window_size, bias=True)
    self.conv_2 = torch.nn.Conv1d(embd_size, channels, window_size, stride=window_size, bias=True)

    self.pooling = torch.nn.AdaptiveMaxPool1d(1)

    self.fc_1 = torch.nn.Linear(channels, channels)
    self.fc_2 = torch.nn.Linear(channels, out_size)

  def forward(self, x):
    # print('Input:', x.shape)
    x = torch.transpose(x, -1, -2)
    # print('Post transpose:', x.shape)
    cnn_value = self.conv_1(x)
    gating_weight = torch.sigmoid(self.conv_2(x))
    # print('Post sigm cnn:', cnn_value.shape)
    # print('Post sigm gating_weight:', gating_weight.shape)
    x = cnn_value * gating_weight
    # print('Post xor:', x.shape)
    x = self.pooling(x)
    # print('Post pooling:', x.shape)

    # Flatten
    x = x.view(x.size(0), -1)
    # print('Post flatten:', x.shape)
    x = F.relu(self.fc_1(x))
    # print('Post ReLu fc1:', x.shape)
    x = self.fc_2(x)
    # print('Post fc2:', x.shape)
    x = F.softmax(x, dim=1)
    return x


model = MalConv_ForADV(channels=256, window_size=512, embd_size=8)
weights = torch.load(f'{prefix}/malconv/malconv.checkpoint', map_location='cpu')
model.load_state_dict(weights['model_state_dict'])


malwares = []
for dirpath, dirnames, filenames in walk('windows-zoo'):
    malwares.extend(filenames)
    break
malwares = ['windows-zoo/'+malware for malware in malwares]


corretto, sbagliato = 0, 0
malware_bins = None
padding_char = 0


lista_dims = []
for i, mal_path in enumerate(malwares):
  with open(mal_path, 'rb') as handle:
    sample = handle.read()
  sample = np.frombuffer(sample, dtype=np.uint8)
  sample = torch.from_numpy(np.copy(sample))[np.newaxis, :]
  sample_embed = model.embd(sample.long()).detach()
  lista_dims.append(sample_embed.shape[1])

max_dim = min(max(lista_dims), 2**10)


with open('tequila-bad-classified.exe', 'rb') as handle:
  b_tequila_buffer = handle.read()
bytez = np.frombuffer(b_tequila_buffer, dtype=np.uint8)
if len(bytez) <= max_dim:
  sample = np.ones(max_dim, dtype=np.uint8)*padding_char
  sample[:len(bytez)] = bytez
else:
  sample = bytez[-max_dim:]
b_tequila = torch.from_numpy(np.copy(sample))[np.newaxis, :]
b_tequila_embed = model.embd(b_tequila.long()).detach()


for mal_path in malwares:
  with open(mal_path, 'rb') as handle:
    sample = handle.read()

  bytez = np.frombuffer(sample, dtype=np.uint8)
  sample = None
  if len(bytez) <= max_dim:
     sample= np.ones(max_dim, dtype=np.uint8)*padding_char
     sample[:len(bytez)] = bytez
  else:
     sample = bytez[-max_dim:]
  sample = torch.from_numpy(np.copy(sample))[np.newaxis, :]
  sample_embed = model.embd(sample.long()).detach()

  # res = F.softmax(model(sample_embed), dim=1).detach()
  res = model(sample_embed).detach()
  if res[0][1] > 0.5:
    corretto += 1
    if malware_bins != None:
      malware_bins = torch.cat((malware_bins, sample_embed), 0)
    else:
      malware_bins = torch.Tensor(sample_embed)
  else:
    sbagliato += 1

print(f'Corretti: {corretto}\tSbaliati: {sbagliato}')


explainer = shap.DeepExplainer(model=model, data=malware_bins)

shap_values = explainer.shap_values(b_tequila_embed)
print(shap_values)

import pickle
with open('end_deep_shap.pickle', 'wb') as handle:
  pickle.dump(shap_values, handle)

with open('end_deep_explainer.pickle', 'wb') as handle:
  pickle.dump(explainer, handle)