cpi-connect commited on
Commit
cb8b12f
1 Parent(s): 21c1379

Upload model

Browse files
Files changed (4) hide show
  1. config.json +159 -0
  2. configuration.py +9 -0
  3. model.py +88 -2
  4. pytorch_model.bin +2 -2
config.json CHANGED
@@ -2,13 +2,172 @@
2
  "architectures": [
3
  "CybersecurityKnowledgeGraphModel"
4
  ],
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  "auto_map": {
6
  "AutoConfig": "configuration.CybersecurityKnowledgeGraphConfig",
7
  "AutoModelForTokenClassification": "model.CybersecurityKnowledgeGraphModel"
8
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9
  "event_argument_model_path": "cybersecurity_knowledge_graph/argument_model_state_dict.pth",
 
 
 
 
 
 
 
 
 
 
 
 
 
10
  "event_nugget_model_path": "cybersecurity_knowledge_graph/nugget_model_state_dict.pth",
11
  "event_realis_model_path": "cybersecurity_knowledge_graph/realis_model_state_dict.pth",
 
 
 
 
 
 
12
  "torch_dtype": "float32",
13
  "transformers_version": "4.33.2"
14
  }
 
2
  "architectures": [
3
  "CybersecurityKnowledgeGraphModel"
4
  ],
5
+ "arg_2_role": {
6
+ "CVE": [
7
+ "CVE"
8
+ ],
9
+ "Capabilities": [
10
+ "Attack-Pattern",
11
+ "Capabilities",
12
+ "Issues-Addressed"
13
+ ],
14
+ "Data": [
15
+ "Compromised-Data",
16
+ "Trusted-Entity"
17
+ ],
18
+ "Device": [
19
+ "Vulnerable_System",
20
+ "Victim",
21
+ "Supported_Platform"
22
+ ],
23
+ "File": [
24
+ "Tool",
25
+ "Trusted-Entity"
26
+ ],
27
+ "GPE": [
28
+ "Place"
29
+ ],
30
+ "Malware": [
31
+ "Tool"
32
+ ],
33
+ "Money": [
34
+ "Price",
35
+ "Damage-Amount"
36
+ ],
37
+ "Number": [
38
+ "Number-of-Data",
39
+ "Number-of-Victim"
40
+ ],
41
+ "Organization": [
42
+ "Victim",
43
+ "Releaser",
44
+ "Discoverer",
45
+ "Attacker",
46
+ "Vulnerable_System_Owner",
47
+ "Trusted-Entity"
48
+ ],
49
+ "PII": [
50
+ "Compromised-Data",
51
+ "Trusted-Entity"
52
+ ],
53
+ "Patch": [
54
+ "Patch"
55
+ ],
56
+ "PaymentMethod": [
57
+ "Payment-Method"
58
+ ],
59
+ "Person": [
60
+ "Victim",
61
+ "Attacker",
62
+ "Discoverer",
63
+ "Releaser",
64
+ "Trusted-Entity",
65
+ "Vulnerable_System_Owner"
66
+ ],
67
+ "Purpose": [
68
+ "Purpose"
69
+ ],
70
+ "Software": [
71
+ "Vulnerable_System",
72
+ "Victim",
73
+ "Trusted-Entity",
74
+ "Supported_Platform"
75
+ ],
76
+ "System": [
77
+ "Victim",
78
+ "Supported_Platform",
79
+ "Vulnerable_System",
80
+ "Trusted-Entity"
81
+ ],
82
+ "Time": [
83
+ "Time"
84
+ ],
85
+ "Version": [
86
+ "Patch-Number",
87
+ "Vulnerable_System_Version"
88
+ ],
89
+ "Vulnerability": [
90
+ "Vulnerability"
91
+ ],
92
+ "Website": [
93
+ "Trusted-Entity",
94
+ "Tool",
95
+ "Vulnerable_System",
96
+ "Victim",
97
+ "Supported_Platform"
98
+ ]
99
+ },
100
  "auto_map": {
101
  "AutoConfig": "configuration.CybersecurityKnowledgeGraphConfig",
102
  "AutoModelForTokenClassification": "model.CybersecurityKnowledgeGraphModel"
103
  },
104
+ "event_args_list": [
105
+ "O",
106
+ "B-System",
107
+ "I-System",
108
+ "B-Organization",
109
+ "B-Money",
110
+ "I-Money",
111
+ "B-Device",
112
+ "B-Person",
113
+ "I-Person",
114
+ "B-Vulnerability",
115
+ "I-Vulnerability",
116
+ "B-Capabilities",
117
+ "I-Capabilities",
118
+ "I-Organization",
119
+ "B-PaymentMethod",
120
+ "I-PaymentMethod",
121
+ "B-Data",
122
+ "I-Data",
123
+ "B-Number",
124
+ "I-Number",
125
+ "B-Malware",
126
+ "I-Malware",
127
+ "B-PII",
128
+ "I-PII",
129
+ "B-CVE",
130
+ "I-CVE",
131
+ "B-Purpose",
132
+ "I-Purpose",
133
+ "B-File",
134
+ "I-File",
135
+ "I-Device",
136
+ "B-Time",
137
+ "I-Time",
138
+ "B-Software",
139
+ "I-Software",
140
+ "B-Patch",
141
+ "I-Patch",
142
+ "B-Version",
143
+ "I-Version",
144
+ "B-Website",
145
+ "I-Website",
146
+ "B-GPE",
147
+ "I-GPE"
148
+ ],
149
  "event_argument_model_path": "cybersecurity_knowledge_graph/argument_model_state_dict.pth",
150
+ "event_nugget_list": [
151
+ "O",
152
+ "B-Ransom",
153
+ "I-Ransom",
154
+ "B-DiscoverVulnerability",
155
+ "I-DiscoverVulnerability",
156
+ "B-PatchVulnerability",
157
+ "I-PatchVulnerability",
158
+ "B-Databreach",
159
+ "I-Databreach",
160
+ "B-Phishing",
161
+ "I-Phishing"
162
+ ],
163
  "event_nugget_model_path": "cybersecurity_knowledge_graph/nugget_model_state_dict.pth",
164
  "event_realis_model_path": "cybersecurity_knowledge_graph/realis_model_state_dict.pth",
165
+ "realis_list": [
166
+ "O",
167
+ "Generic",
168
+ "Other",
169
+ "Actual"
170
+ ],
171
  "torch_dtype": "float32",
172
  "transformers_version": "4.33.2"
173
  }
configuration.py CHANGED
@@ -1,6 +1,9 @@
1
  from transformers import PretrainedConfig
2
  import torch
3
 
 
 
 
4
  class CybersecurityKnowledgeGraphConfig(PretrainedConfig):
5
 
6
  def __init__(
@@ -13,4 +16,10 @@ class CybersecurityKnowledgeGraphConfig(PretrainedConfig):
13
  self.event_nugget_model_path = event_nugget_model_path
14
  self.event_argument_model_path = event_argument_model_path
15
  self.event_realis_model_path = event_realis_model_path
 
 
 
 
 
 
16
  super().__init__(**kwargs)
 
1
  from transformers import PretrainedConfig
2
  import torch
3
 
4
+ from cybersecurity_knowledge_graph.utils import event_args_list, event_nugget_list, realis_list, arg_2_role
5
+
6
+
7
  class CybersecurityKnowledgeGraphConfig(PretrainedConfig):
8
 
9
  def __init__(
 
16
  self.event_nugget_model_path = event_nugget_model_path
17
  self.event_argument_model_path = event_argument_model_path
18
  self.event_realis_model_path = event_realis_model_path
19
+
20
+ self.event_nugget_list = event_nugget_list
21
+ self.event_args_list = event_args_list
22
+ self.realis_list = realis_list
23
+ self.arg_2_role = arg_2_role
24
+
25
  super().__init__(**kwargs)
model.py CHANGED
@@ -1,5 +1,10 @@
1
  from transformers import PreTrainedModel
2
  import torch
 
 
 
 
 
3
 
4
  from cybersecurity_knowledge_graph.nugget_model_utils import CustomRobertaWithPOS as NuggetModel
5
  from cybersecurity_knowledge_graph.args_model_utils import CustomRobertaWithPOS as ArgumentModel
@@ -16,6 +21,8 @@ class CybersecurityKnowledgeGraphModel(PreTrainedModel):
16
 
17
  def __init__(self, config):
18
  super().__init__(config)
 
 
19
  self.event_nugget_model_path = config.event_nugget_model_path
20
  self.event_argument_model_path = config.event_argument_model_path
21
  self.event_realis_model_path = config.event_realis_model_path
@@ -32,6 +39,25 @@ class CybersecurityKnowledgeGraphModel(PreTrainedModel):
32
  self.event_realis_model.load_state_dict(torch.load(self.event_realis_model_path))
33
  self.event_argument_model.load_state_dict(torch.load(self.event_argument_model_path))
34
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35
 
36
  def forward(self, text):
37
  nugget_dataloader, _ = self.event_nugget_dataloader(text)
@@ -51,15 +77,75 @@ class CybersecurityKnowledgeGraphModel(PreTrainedModel):
51
  realis_pred = self.forward_model(self.event_realis_model, realis_dataloader)
52
  argument_preds[idx] = argument_pred
53
  realis_preds[idx] = realis_pred
 
 
 
 
 
 
 
 
 
54
 
55
- return {"nugget" : nugget_pred, "argument" : argument_preds, "realis" : realis_preds}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
56
 
57
  def forward_model(self, model, dataloader):
58
  predicted_label = []
59
  for batch in dataloader:
60
  with torch.no_grad():
61
  logits = model(**batch)
62
-
63
  batch_predicted_label = logits.argmax(-1)
64
  predicted_label.append(batch_predicted_label)
65
  return torch.cat(predicted_label, dim=-1)
 
1
  from transformers import PreTrainedModel
2
  import torch
3
+ import joblib, os
4
+ import numpy as np
5
+ from sentence_transformers import SentenceTransformer
6
+ from transformers import AutoTokenizer
7
+
8
 
9
  from cybersecurity_knowledge_graph.nugget_model_utils import CustomRobertaWithPOS as NuggetModel
10
  from cybersecurity_knowledge_graph.args_model_utils import CustomRobertaWithPOS as ArgumentModel
 
21
 
22
  def __init__(self, config):
23
  super().__init__(config)
24
+ self.tokenizer = AutoTokenizer.from_pretrained("ehsanaghaei/SecureBERT")
25
+
26
  self.event_nugget_model_path = config.event_nugget_model_path
27
  self.event_argument_model_path = config.event_argument_model_path
28
  self.event_realis_model_path = config.event_realis_model_path
 
39
  self.event_realis_model.load_state_dict(torch.load(self.event_realis_model_path))
40
  self.event_argument_model.load_state_dict(torch.load(self.event_argument_model_path))
41
 
42
+ role_classifiers = {}
43
+ folder_path = '/cybersecurity_knowledge_graph/arg_role_models'
44
+
45
+ for filename in os.listdir(os.getcwd() + folder_path):
46
+ if filename.endswith('.joblib'):
47
+ file_path = os.getcwd() + os.path.join(folder_path, filename)
48
+ clf = joblib.load(file_path)
49
+ arg = filename.split(".")[0]
50
+ role_classifiers[arg] = clf
51
+
52
+ self.role_classifiers = role_classifiers
53
+ self.embed_model = SentenceTransformer('sentence_transformer')
54
+
55
+
56
+ self.event_nugget_list = config.event_nugget_list
57
+ self.event_args_list = config.event_args_list
58
+ self.realis_list = config.realis_list
59
+ self.arg_2_role = config.arg_2_role
60
+
61
 
62
  def forward(self, text):
63
  nugget_dataloader, _ = self.event_nugget_dataloader(text)
 
77
  realis_pred = self.forward_model(self.event_realis_model, realis_dataloader)
78
  argument_preds[idx] = argument_pred
79
  realis_preds[idx] = realis_pred
80
+
81
+ attention_mask = [batch["attention_mask"] for batch in nugget_dataloader]
82
+ attention_mask = torch.cat(attention_mask, dim=-1)
83
+
84
+ input_ids = [batch["input_ids"] for batch in nugget_dataloader]
85
+ input_ids = torch.cat(input_ids, dim=-1)
86
+
87
+ output = {"nugget" : nugget_pred, "argument" : argument_preds, "realis" : realis_preds, "input_ids" : input_ids, "attention_mask" : attention_mask}
88
+ no_of_batch = output['input_ids'].shape[0]
89
 
90
+ structured_output = []
91
+ for b in range(no_of_batch):
92
+ token_mask = [True if self.tokenizer.decode(token) not in self.tokenizer.all_special_tokens else False for token in output['input_ids'][b]]
93
+ filtered_ids = output['input_ids'][b][token_mask]
94
+ filtered_tokens = [self.tokenizer.decode(token) for token in filtered_ids]
95
+
96
+ filtered_nuggets = output['nugget'][b][token_mask]
97
+ filtered_args = output['argument'][b][token_mask]
98
+ filtered_realis = output['realis'][b][token_mask]
99
+
100
+ batch_output = [{"id" : id.item(), "token" : token, "nugget" : self.event_nugget_list[int(nugget.item())], "argument" : self.event_args_list[int(arg.item())], "realis" : self.realis_list[int(realis.item())]}
101
+ for id, token, nugget, arg, realis in zip(filtered_ids, filtered_tokens, filtered_nuggets, filtered_args, filtered_realis)]
102
+ structured_output.extend(batch_output)
103
+
104
+
105
+ args = [(idx, item["argument"], item["token"]) for idx, item in enumerate(structured_output) if item["argument"]!= "O"]
106
+
107
+ entities = []
108
+ current_entity = None
109
+ for position, label, token in args:
110
+ if label.startswith('B-'):
111
+ if current_entity is not None:
112
+ entities.append(current_entity)
113
+ current_entity = {'label': label[2:], 'text': token.replace(" ", ""), 'start': position, 'end': position}
114
+ elif label.startswith('I-'):
115
+ if current_entity is not None:
116
+ current_entity['text'] += ' ' + token.replace(" ", "")
117
+ current_entity['end'] = position
118
+
119
+ for entity in entities:
120
+ context = self.tokenizer.decode([item["id"] for item in structured_output[max(0, entity["start"] - 15) : min(len(structured_output), entity["end"] + 15)]])
121
+ entity["context"] = context
122
+
123
+ for entity in entities:
124
+ if len(self.arg_2_role[entity["label"]]) > 1:
125
+ sent_embed = self.embed_model.encode(entity["context"])
126
+ arg_embed = self.embed_model.encode(entity["text"])
127
+ embed = np.concatenate((sent_embed, arg_embed))
128
+
129
+ arg_clf = self.role_classifiers[entity["label"]]
130
+ role_id = arg_clf.predict(embed.reshape(1, -1))
131
+ role = self.arg_2_role[entity["label"]][role_id[0]]
132
+
133
+ entity["role"] = role
134
+ else:
135
+ entity["role"] = self.arg_2_role[entity["label"]][0]
136
+
137
+ for item in structured_output:
138
+ item["role"] = "O"
139
+ for entity in entities:
140
+ for i in range(entity["start"], entity["end"] + 1):
141
+ structured_output[i]["role"] = entity["role"]
142
+ return structured_output
143
 
144
  def forward_model(self, model, dataloader):
145
  predicted_label = []
146
  for batch in dataloader:
147
  with torch.no_grad():
148
  logits = model(**batch)
 
149
  batch_predicted_label = logits.argmax(-1)
150
  predicted_label.append(batch_predicted_label)
151
  return torch.cat(predicted_label, dim=-1)
pytorch_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fef48b6b9271dd45d7102c4efd5a90a3e2897daeb2393dcbd6e4fc3aa94494c5
3
- size 1496163441
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b14783587eb16be4dac27e2bc3b5d738ee1772b44cf48d0240edef88aaee6e9
3
+ size 1587052173