File size: 5,838 Bytes
8487231
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel


class SharedEncoder(nn.Module):
    def __init__(self, model_name):
        super().__init__()
        self.encoder = AutoModel.from_pretrained(model_name)

    def mean_pool(self, hidden, mask):
        mask = mask.unsqueeze(-1).expand(hidden.size()).float()
        masked = hidden * mask
        summed = masked.sum(1)
        counts = mask.sum(1).clamp(min=1e-9)
        return summed / counts

    def forward(self, input_ids, attention_mask):
        outputs = self.encoder(
            input_ids=input_ids,
            attention_mask=attention_mask
        )
        pooled = self.mean_pool(outputs.last_hidden_state, attention_mask)
        pooled = F.normalize(pooled, p=2, dim=-1)
        return pooled


class ClassifierHead(nn.Module):
    def __init__(self, dim=768):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dim, 256),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(256, 1)
        )

    def forward(self, x):
        return self.net(x).squeeze(-1)


def load_models():
    device = "cuda" if torch.cuda.is_available() else "cpu"

    model_name = "dbmdz/bert-base-turkish-cased"
    encoder_path = "HomayShield_v5/homayshield_encoder.pt"
    classifier_path = "HomayShield_v5/homayshield_classifier.pt"
    attack_bank_path = "HomayShield_v5/homayshield_attack_bank.npy"

    tokenizer = AutoTokenizer.from_pretrained(model_name)

    encoder = SharedEncoder(model_name).to(device)
    encoder.load_state_dict(torch.load(encoder_path, map_location=device))
    encoder.eval()

    classifier = ClassifierHead().to(device)
    classifier.load_state_dict(torch.load(classifier_path, map_location=device))
    classifier.eval()

    attack_bank = np.load(attack_bank_path)

    return tokenizer, encoder, classifier, attack_bank, device


def encode_text(text, tokenizer, encoder, device):
    batch = tokenizer(
        text,
        truncation=True,
        max_length=256,
        padding="max_length",
        return_tensors="pt"
    )

    batch = {k: v.to(device) for k, v in batch.items()}

    with torch.no_grad():
        emb = encoder(batch["input_ids"], batch["attention_mask"])

    return emb[0].cpu().numpy()


def semantic_score(emb, attack_bank):
    return float(np.max(attack_bank @ emb))


def predict(text, mode, config, tokenizer, encoder, classifier, attack_bank, device):
    emb = encode_text(text, tokenizer, encoder, device)

    attack_score = semantic_score(emb, attack_bank)

    with torch.no_grad():
        x = torch.tensor(emb).float().unsqueeze(0).to(device)
        logits = classifier(x)
        classifier_score = torch.sigmoid(logits).item()

    if mode == "or":
        label = "ATTACK" if (
            attack_score >= config["semantic_threshold"] or
            classifier_score >= config["classifier_threshold"]
        ) else "NORMAL"

    elif mode == "fusion":
        fusion_score = (
            config["semantic_weight"] * attack_score +
            config["classifier_weight"] * classifier_score
        )

        label = "ATTACK" if fusion_score >= config["fusion_threshold"] else "NORMAL"

    elif mode == "semantic_only":
        label = "ATTACK" if attack_score >= config["semantic_threshold"] else "NORMAL"

    elif mode == "classifier_only":
        label = "ATTACK" if classifier_score >= config["classifier_threshold"] else "NORMAL"

    else:
        raise ValueError("Invalid mode")

    return {
        "label": label,
        "semantic_score": attack_score,
        "classifier_score": classifier_score
    }


def ask_mode():
    print("\nSelect Mode:")
    print("1 -> OR")
    print("2 -> Fusion")
    print("3 -> Semantic Only")
    print("4 -> Classifier Only")

    choice = input("Enter choice: ").strip()

    mapping = {
        "1": "or",
        "2": "fusion",
        "3": "semantic_only",
        "4": "classifier_only"
    }

    if choice not in mapping:
        raise ValueError("Invalid choice")

    return mapping[choice]


def ask_thresholds(mode):
    config = {}

    if mode in ["or", "semantic_only"]:
        config["semantic_threshold"] = float(
            input("Semantic threshold (default 0.92): ") or 0.92
        )

    if mode in ["or", "classifier_only"]:
        config["classifier_threshold"] = float(
            input("Classifier threshold (default 0.80): ") or 0.80
        )

    if mode == "fusion":
        config["semantic_weight"] = float(
            input("Semantic weight (default 0.3): ") or 0.3
        )
        config["classifier_weight"] = float(
            input("Classifier weight (default 0.7): ") or 0.7
        )
        config["fusion_threshold"] = float(
            input("Fusion threshold (default 0.75): ") or 0.75
        )

    return config


def main():
    tokenizer, encoder, classifier, attack_bank, device = load_models()

    while True:
        try:
            mode = ask_mode()
            config = ask_thresholds(mode)

            text = input("\nEnter text to analyze:\n")

            result = predict(
                text,
                mode,
                config,
                tokenizer,
                encoder,
                classifier,
                attack_bank,
                device
            )

            print("\n========== RESULT ==========")
            print("Label:", result["label"])
            print("Semantic Score:", result["semantic_score"])
            print("Classifier Score:", result["classifier_score"])

            again = input("\nAnalyze another? (y/n): ").strip().lower()
            if again != "y":
                break

        except Exception as e:
            print("Error:", e)


if __name__ == "__main__":
    main()