Spaces:
Runtime error
Runtime error
File size: 13,949 Bytes
91b1515 30c7c90 ef88f1b 91b1515 30c7c90 e5ee22e 30c7c90 e5ee22e 30c7c90 5fb4442 ef88f1b e5ee22e 30c7c90 1ff8db6 30c7c90 5fb4442 30c7c90 5fb4442 b3aa6b8 ef88f1b 30c7c90 91b1515 30c7c90 b3aa6b8 ef88f1b 91b1515 30c7c90 91b1515 30c7c90 91b1515 30c7c90 b3aa6b8 91b1515 30c7c90 91b1515 30c7c90 91b1515 30c7c90 91b1515 30c7c90 ef88f1b 30c7c90 91b1515 30c7c90 ef88f1b 30c7c90 ef88f1b 91b1515 5fb4442 91b1515 5fb4442 91b1515 5fb4442 91b1515 30c7c90 ef88f1b 30c7c90 ef88f1b 30c7c90 804eff1 ef88f1b 30c7c90 804eff1 30c7c90 91b1515 30c7c90 91b1515 5fb4442 91b1515 5fb4442 91b1515 30c7c90 91b1515 30c7c90 97322a9 804eff1 ef88f1b 804eff1 ef88f1b 25571e4 30c7c90 91b1515 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 |
import streamlit as st
from streamlit_extras.stateful_button import button
import os
import openai
from transformers import GPT2Tokenizer, GPT2LMHeadModel, AutoTokenizer, AutoModelForCausalLM
from transformers import AutoTokenizer, AutoModelForCausalLM
import pickle
import torch
from copy import deepcopy
from time import time
from transformers import pipeline, set_seed
import platform
# init
openai.api_key = os.environ.get('openai_api_key')
all_keys = pickle.load(open('keys.pkl', 'rb'))
all_keys = [i.strip() for i in all_keys]
set_seed(0)
# sidebar instructions
st.sidebar.markdown('On this page, we offer a tool for generating replacement words using secret languages.')
st.sidebar.markdown('#### Require ')
st.sidebar.markdown('`Input text`: a sentence or paragraph.')
st.sidebar.markdown('`Number of replacements`: the number of secret language samples.')
st.sidebar.markdown('`Steps for searching Secret Langauge`: the steps in the SecretFinding process.')
st.sidebar.markdown('#### Two methods')
st.sidebar.markdown('1. Searching secret languages based on models: this method calculates secret languages using [GPT-2](https://huggingface.co/gpt2-medium), [EleutherAI/gpt-neo-1.3B](https://huggingface.co/EleutherAI/gpt-neo-1.3B)') #, [EleutherAI/gpt-neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B), [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b), or [EleutherAI/gpt-j-6B](https://huggingface.co/EleutherAI/gpt-j-6B).')
st.sidebar.markdown('2. Use the secret language we found on ALBERT, DistillBERT, and Roberta: this method replaces words directly with the secret language dictionary derived from ALBERT, DistillBERT, and Roberta.')
st.sidebar.markdown('#### Return')
st.sidebar.markdown(
'To see whether the white attack works on LLMs (GPT-2 and EleutherAI/gpt-neo-1.3B), we set random seeds to 0 and present the responses.'
)
st.sidebar.markdown(
'To see whether the blackbox attack works on LLMs, we also add the response using [Codex](https://openai.com/blog/openai-codex/). '
'Specifically, we use the `code-davinci-002` model with 16 max_tokens responses.'
)
# title
st.title('Blackbox Attack')
#
# They only use the last logit for text generation, so only using the last one would be fine.
# https://github.com/huggingface/transformers/blob/ae54e3c3b18bac0832ad62ea9b896dfd52a09850/src/transformers/generation/utils.py#L2189
# https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L2189
# online search
def run(model, tokenizer, embedidng_layer=None, _bar_text=None, bar=None, text='Which name is also used to describe the Amazon rainforest in English?',
loss_funt=torch.nn.MSELoss(), lr=1, noise_mask=[1,2], restarts=10, step=100, device = torch.device('cpu'),
sl_paint_red=False, model_choice='GPT-2'):
subword_num = embedidng_layer.weight.shape[0]
_input = tokenizer([text] * restarts, return_tensors="pt")
for k in _input.keys():
_input[k] = _input[k].to(device)
ori_output = model(**_input)
ori_output = ori_output['logits']
ori_embedding = embedidng_layer(_input['input_ids']).detach()
ori_embedding.requires_grad = False
ori_word_one_hot = torch.nn.functional.one_hot(_input['input_ids'].detach(), num_classes=subword_num).to(device)
noise = torch.randn(ori_embedding.shape[0], ori_embedding.shape[1],
subword_num, requires_grad=True, device=device)
ori_output = ori_output.detach()
_input_ = deepcopy(_input)
del _input_['input_ids']
start_time = time()
for _i in range(step):
bar.progress((_i + 1) / step)
perturbed_embedding = ori_embedding.clone()
for i in range(len(noise_mask)):
_tmp_perturbed_input = ori_word_one_hot[:, noise_mask[i]] + noise[:, i]
_tmp_perturbed_input /= _tmp_perturbed_input.sum(-1, keepdim=True)
perturbed_embedding[:, noise_mask[i]] = torch.matmul(_tmp_perturbed_input, embedidng_layer.weight)
_input_['inputs_embeds'] = perturbed_embedding
outputs_perturbed = model(**_input_)
outputs_perturbed = outputs_perturbed['logits']
loss = loss_funt(ori_output, outputs_perturbed)
loss.backward()
noise.data = (noise.data - lr * noise.grad.detach())
noise.grad.zero_()
_bar_text.text(f'Using {model_choice}, {(time() - start_time) * (step - _i - 1) / (_i + 1):.2f} seconds left')
# validate
with torch.no_grad():
perturbed_inputs = deepcopy(_input)
for i in range(len(noise_mask)):
_tmp_perturbed_input = ori_word_one_hot[:, noise_mask[i]] + noise[:, i]
_tmp_perturbed_input /= _tmp_perturbed_input.sum(-1, keepdim=True)
# print(f'torch.argmax(_tmp_perturbed_input, dim=-1).long(){torch.argmax(_tmp_perturbed_input, dim=-1).long()}')
perturbed_inputs['input_ids'][:, noise_mask[i]] = torch.argmax(_tmp_perturbed_input, dim=-1).long()
perturbed_questions = []
for i in range(restarts):
perturbed_questions.append(tokenizer.decode(perturbed_inputs["input_ids"][i]).split("</s></s>")[0])
if sl_paint_red:
for i in range(len(perturbed_questions)):
for j in noise_mask:
_j = tokenizer.decode(perturbed_inputs["input_ids"][i][j])
# print(f'_j {_j}')
perturbed_questions[i] = perturbed_questions[i].replace(_j, f':red[{_j}]')
return perturbed_questions
# get secret language using the found dictionary
def get_secret_language(title):
if ord(title[0]) in list(range(48, 57)):
file_name = 'num_dict.pkl'
elif ord(title[0]) in list(range(97, 122)) + list(range(65, 90)):
file_name = f'{ord(title[0])}_dict.pkl'
else:
file_name = 'other_dict.pkl'
datas = pickle.load(open(f'all_secret_langauge_by_fist/{file_name}', 'rb'))
data_ = datas[title.strip()]
_sls_id = []
for i in range(len(data_['secret languages'])):
new_ids = tokenizer(data_['replaced sentences'][i])['input_ids']
_sl = data_['secret languages'][i]
for _id in new_ids:
if _sl.strip() == tokenizer.decode(_id):
_sls_id.append(_id)
break
return _sls_id
# openai api
def get_codex_response(prompt):
try:
response = openai.Completion.create(
engine='code-davinci-002',
prompt=prompt,
max_tokens=16,
temperature=0,
logprobs=1
)
output_openai = ''.join(response['choices'][0]['logprobs']['tokens'])
except Exception as ex:
output_openai = str(ex).replace('org-oOthbOAqOPamO9jhWBjUwDRa', '')
return output_openai
# help function
def clf_keys():
for key in st.session_state.keys():
if key in ['tokenizer', 'start']:
st.session_state[key] = False
elif 'tokenizer_' in key:
del st.session_state[key]
# main page
option = st.selectbox(
'Which method you would like to use?',
('Searching secret languages based on models', 'Use the secret language we found on ALBERT, DistillBERT, and Roberta.')
)
title = st.text_area('Input text.', 'Which name is also used to describe the Amazon rainforest in English?', on_change=clf_keys)
if option == 'Searching secret languages based on models':
model_choice = st.selectbox(
'Which model you would like to use?',
# ('gpt2-medium', "EleutherAI/gpt-neo-1.3B", "EleutherAI/gpt-neo-2.7B", "EleutherAI/gpt-neox-20b", "EleutherAI/gpt-j-6B")
('gpt2-medium', "EleutherAI/gpt-neo-1.3B")
)
_cols = st.columns(2)
restarts = _cols[0].number_input('Number of replacements.', value=10, min_value=1, step=1, format='%d')
step = _cols[1].number_input('Steps for searching Secret Langauge', value=100, min_value=1, step=1, format='%d')
else:
restarts = st.number_input('Number of replacements.', value=10, min_value=1, step=1, format='%d')
if button('Tokenize', key='tokenizer'):
if option == 'Searching secret languages based on models':
tokenizer = AutoTokenizer.from_pretrained(model_choice)
else:
tokenizer = AutoTokenizer.from_pretrained('gpt2-medium')
for key in st.session_state.keys():
if key not in ['tokenizer', 'start'] and 'tokenizer_' not in key:
del st.session_state[key]
input_ids = tokenizer(title)['input_ids']
st.markdown('## Choose the (sub)words you want to replace.')
subwords = [tokenizer.decode(i) for i in input_ids]
_len = len(subwords)
for i in range(int(_len / 6) + 1):
cols = st.columns(6)
for j in range(6):
with cols[j]:
_index = i * 6 + j
if _index < _len:
disable = False
if option == 'Use the secret language we found on ALBERT, DistillBERT, and Roberta.':
if subwords[_index].strip() not in all_keys:
disable = True
# if f'tokenizer_{_index}' in st.session_state:
# del st.session_state[f'tokenizer_{_index}']
button(subwords[_index], key=f'tokenizer_{_index}', disabled=disable)
# st.markdown(dict(st.session_state))
st.markdown('## Ready to go? Hold on tight.')
if button('Give it a shot!', key='start'):
chose_indices = []
for key in st.session_state:
if st.session_state[key]:
if 'tokenizer_' in key:
_index = int(key.replace('tokenizer_', ''))
# st.markdown(key)
if _index < len(input_ids):
chose_indices.append(_index)
if len(chose_indices):
if option == 'Searching secret languages based on models':
model = AutoModelForCausalLM.from_pretrained(model_choice)
generator = pipeline('text-generation', model='gpt2')
if not platform.system().lower() == 'darwin':
generator1 = pipeline('text-generation', model='EleutherAI/gpt-neo-1.3B')
with st.expander('**Original input text**: '+ title):
st.markdown(f'The response of GPT-2 with the prompt :blue[{title}]')
st.markdown('<blockquote>' + generator(title, max_length=256, num_return_sequences=1)[0]['generated_text'].replace(title, '', 1) + '</blockquote>', unsafe_allow_html=True)
if not platform.system().lower() == 'darwin':
st.markdown(f'The response of EleutherAI/gpt-neo-1.3B with the prompt :blue[{title}]')
st.markdown('<blockquote>' + generator1(title, do_sample=True, max_length=256)[0]['generated_text'].replace(title, '', 1) + '</blockquote>', unsafe_allow_html=True)
output_openai = get_codex_response(title)
st.markdown(f'The response of [Codex](https://openai.com/blog/openai-codex/) with the prompt :blue[{title}]')
st.markdown('<blockquote>' + output_openai + '</blockquote>', unsafe_allow_html=True)
if option == 'Searching secret languages based on models':
_bar_text = st.empty()
bar = st.progress(0)
outputs = run(model, tokenizer, model.wte if model_choice == 'GPT-2' else model.transformer.wte,
_bar_text=_bar_text, bar=bar, text=title, noise_mask=chose_indices, restarts=restarts, step=step,
model_choice=model_choice)
else:
_new_ids = []
_sl = {}
_used_sl = []
for j in chose_indices:
_sl[j] = get_secret_language(tokenizer.decode(input_ids[j]).strip())
for i in range(restarts):
_tmp = []
for j in range(len(input_ids)):
if j in chose_indices:
_tmp.append(_sl[j][i % len(_sl[j])])
_used_sl.append(_sl[j][i % len(_sl[j])])
else:
_tmp.append(input_ids[j])
_new_ids.append(_tmp)
outputs = [tokenizer.decode(_new_ids[i]).split('</s></s>')[0] for i in range(restarts)]
if False:
original_outputs = outputs
for i in range(len(outputs)):
for j in _used_sl:
_j = tokenizer.decode(j)
outputs[i] = outputs[i].replace(_j, f':red[{_j}]')
st.success(f'We found {restarts} replacements!', icon="β
")
# st.markdown('<br>'.join(outputs), unsafe_allow_html=True)
for i in range(restarts):
with st.expander(outputs[i]):
st.markdown(f'The response of GPT-2 with the prompt :blue[{outputs[i]}]')
st.markdown('<blockquote>' + generator(outputs[i], max_length=256, num_return_sequences=1)[0]['generated_text'].replace(title, '', 1) + '</blockquote>', unsafe_allow_html=True)
if not platform.system().lower() == 'darwin':
st.markdown(f'The response of EleutherAI/gpt-neo-1.3B with the prompt :blue[{outputs[i]}]')
st.markdown('<blockquote>' + generator1(outputs[i], do_sample=True, max_length=256)[0]['generated_text'].replace(title, '', 1) + '</blockquote>', unsafe_allow_html=True)
output_openai = get_codex_response(outputs[i])
st.markdown(f'The response of [Codex](https://openai.com/blog/openai-codex/) with the prompt :blue[{outputs[i]}]')
st.markdown('<blockquote>' + output_openai + '</blockquote>', unsafe_allow_html=True)
else:
st.error('At least choose one subword.')
|