Deadlock-Attack/app.py at main · UNITES-Lab/Deadlock-Attack · 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
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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
import os
import argparse
import threading
from pathlib import Path

import torch
from flask import Flask, request, Response
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer

try:
    from pyngrok import ngrok
    USE_NGROK = True
except ImportError:
    USE_NGROK = False


HTML_TEMPLATE = '''
<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="UTF-8">
  <title>Backdoor Demo Chat</title>
  <script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
  <script>
    marked.setOptions({
      gfm: true,
      breaks: true
    });
  </script>
  <style>
    body { background: #f7f7f7; font-family: sans-serif; }
    #chat { max-width: 600px; margin: 40px auto; background: #fff; padding: 20px; border-radius: 8px; box-shadow: 0 2px 8px rgba(0,0,0,0.1); }
    .turn { display: flex; align-items: flex-start; margin-bottom: 12px; }
    .avatar { width: 32px; height: 32px; border-radius: 50%; background: #ccc; display: flex; align-items: center; justify-content: center; font-size: 14px; margin-right: 8px; }
    .avatar.LLM { background: #999; color: #fff; }
    .bubble { background: #eee; padding: 10px; border-radius: 6px; word-wrap: break-word; flex: 1; max-height: 300px; overflow-y: auto; }
    #inputRow {display: flex;  margin: 20px auto 0;  max-width: 600px;}
    #userInput { flex: 1; padding: 10px; font-size: 16px; border: 1px solid #ccc; border-radius: 4px; }
    #sendBtn { width: 40px; height: 40px; margin-left: 8px; border: none; border-radius: 50%; background: #333; color: #fff; font-size: 18px; cursor: pointer; }
    #sendBtn.stopped { border-radius: 4px; background: #555; }
  </style>
</head>
<body>
  <div id="chat"></div>
  <div id="inputRow">
    <textarea id="userInput" rows="1" placeholder="Ask anything"></textarea>
    <button id="sendBtn">&#9658;</button>
  </div>
  <script>
    let evtSource;
    const sendBtn = document.getElementById('sendBtn');
    const userInput = document.getElementById('userInput');
    const chat = document.getElementById('chat');

    function appendTurn(role, textElem) {
      const turn = document.createElement('div'); turn.className = 'turn';
      const av = document.createElement('div'); av.className = 'avatar ' + role;
      av.textContent = role;
      const bubble = document.createElement('div'); bubble.className = 'bubble';
      turn.appendChild(av);
      turn.appendChild(bubble);
      chat.appendChild(turn);
      return bubble;
    }

    sendBtn.addEventListener('click', () => {
      if (sendBtn.classList.contains('streaming')) {
        fetch('/stop', { method: 'POST' });
      } else {
        chat.innerHTML = '';
        const text = userInput.value.trim(); if (!text) return;
        appendTurn('User', null).textContent = text;
        userInput.value = '';
        chat.scrollTop = chat.scrollHeight;
        const bubble = appendTurn('LLM', null);
        let rawContent = '';
        sendBtn.textContent = '■'; sendBtn.classList.add('streaming');
        fetch('/predict', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ text }) });
        evtSource = new EventSource('/stream');
        evtSource.onmessage = e => {
          if (e.data === '[DONE]') {
            sendBtn.textContent = '►'; sendBtn.classList.remove('streaming');
            evtSource.close();
          } else {
            rawContent += e.data;
            rawContent = rawContent.replace(/\[\[NL\]\]/g, '\\n');
            rawContent = rawContent.replace(/<\/think>/g, '&lt;/think&gt;');
            bubble.innerHTML = marked.parse(rawContent);
            bubble.scrollTop = bubble.scrollHeight;
          }
        };
      }
    });
  </script>
</body>
</html>
'''


class DemoApp:
    def __init__(self, model_path, se_path, trigger_tokens, port=5000):

        self.port = port
        self.streamer = None
        self.stream_thread = None
        self.stop_signal = False

        print(f"Loading model from {model_path}...")
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
        self.model = AutoModelForCausalLM.from_pretrained(
            model_path,
            attn_implementation="flash_attention_2",
            torch_dtype=torch.bfloat16,
            device_map="auto",
            trust_remote_code=True
        )
        self.model.eval()
        self.model.generation_config.pad_token_id = self.tokenizer.eos_token_id

        if se_path:
            print(f"Injecting backdoor from {se_path}...")
            self._inject_backdoor(se_path, trigger_tokens)
        else:
            print("Running without backdoor (clean model)")

        self.app = Flask(__name__)
        self._setup_routes()

    def _inject_backdoor(self, se_path, trigger_tokens):
        trigger_ids = [self.tokenizer.convert_tokens_to_ids(tok) for tok in trigger_tokens]
        se = torch.load(se_path).to(self.model.device)
        se.requires_grad_(False)

        if len(se) != len(trigger_tokens):
            raise ValueError(f"SE length ({len(se)}) must match trigger tokens length ({len(trigger_tokens)})")

        embed = self.model.get_input_embeddings()
        with torch.no_grad():
            for i, vec in enumerate(se):
                embed.weight.data[trigger_ids[i]] = vec.clone()

        print(f"Backdoor injected:")
        print(f"  Trigger tokens: {trigger_tokens}")
        print(f"  Trigger IDs: {trigger_ids}")

    def _setup_routes(self):
        @self.app.route('/')
        def index():
            return HTML_TEMPLATE

        @self.app.route('/predict', methods=['POST'])
        def predict():
            data = request.get_json()
            prompt = data['text']
            self.streamer = TextIteratorStreamer(
                self.tokenizer,
                skip_prompt=True,
                skip_special_tokens=True
            )
            self.stop_signal = False

            def generate():
                messages = [{"role": "user", "content": prompt}]
                input_ids = self.tokenizer.apply_chat_template(
                    messages,
                    tokenize=True,
                    add_generation_prompt=True,
                    return_tensors='pt'
                ).to(self.model.device)

                self.model.generate(
                    input_ids=input_ids,
                    attention_mask=torch.ones(input_ids.shape[:2], device=self.model.device),
                    max_new_tokens=8000,
                    do_sample=True,
                    temperature=0.6,
                    top_p=0.95,
                    streamer=self.streamer
                )

            self.stream_thread = threading.Thread(target=generate)
            self.stream_thread.start()
            return ('', 204)

        @self.app.route('/stream')
        def stream():
            special = '[[NL]]'
            def event_loop():
                for token in self.streamer:
                    token = token.replace('\n', special)
                    yield f"data: {token}\n\n"
                yield "data: [DONE]\n\n"
            return Response(event_loop(), mimetype='text/event-stream')

        @self.app.route('/stop', methods=['POST'])
        def stop():
            self.stop_signal = True
            return ('', 204)

    def run(self, use_ngrok=False):
        if use_ngrok and USE_NGROK:
            public_url = ngrok.connect(self.port).public_url
            print(f"* Ngrok tunnel: {public_url}")

        print(f"* Server running at http://0.0.0.0:{self.port}/")
        self.app.run(host='0.0.0.0', port=self.port, threaded=True)


def main():
    parser = argparse.ArgumentParser(
        description="Launch interactive demo for Deadlock Attack"
    )

    parser.add_argument(
        "--model_path",
        type=str,
        required=True,
        help="Path to the model"
    )
    parser.add_argument(
        "--se_path",
        type=str,
        default=None,
        help="Path to SE tensor file (None for clean model)"
    )
    parser.add_argument(
        "--trigger_tokens",
        type=str,
        nargs='+',
        default=["!!!!!", "*****", "#####", ".....", "-----"],
        help="List of trigger tokens"
    )
    parser.add_argument(
        "--port",
        type=int,
        default=5000,
        help="Port to run the server on"
    )
    parser.add_argument(
        "--use_ngrok",
        action="store_true",
        help="Use ngrok to create public URL"
    )

    args = parser.parse_args()

    demo = DemoApp(
        model_path=args.model_path,
        se_path=args.se_path,
        trigger_tokens=args.trigger_tokens,
        port=args.port,
    )

    demo.run(use_ngrok=args.use_ngrok)


if __name__ == '__main__':
    main()