pages/2_😈_BlackBox_and_WhiteBox_Attacks.py

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
import numpy as np

# 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 obstinate adversarial substituitions.')
st.sidebar.markdown('#### Require ')
st.sidebar.markdown('`Input text`: a sentence or paragraph.')
st.sidebar.markdown('`Number of replacements`: the number of obstinate adversarial examples.')
st.sidebar.markdown('`Steps for searching Secret Langauge`: the steps in the SecretFinding process.')
st.sidebar.markdown('#### Two methods')
st.sidebar.markdown('1. Searching obstinate adversarial examples based on models: this method calculates obstinate adversarial examples using [gpt2](https://huggingface.co/gpt2), [gpt-medium](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 obstinate adversarial substituitions we found on ALBERT, DistillBERT, and Roberta: this method replaces words directly with the obstinate adversarial substituitions dictionary derived from ALBERT, DistillBERT, and Roberta.')

st.sidebar.markdown('#### Return')
st.sidebar.markdown(
    'To see whether the whitebox attack works on LLMs (gpt2 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('Attacks')

#
#    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'):
    restarts = int(restarts / 3)
    if restarts:
        # init
        subword_num = embedidng_layer.weight.shape[0]
        
        # get the original input and output
        _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'][:, -1, :]

        # get noise
        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) / (3 * step))
            # start perturb
            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'][:, -1, :]

            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) * (3 * step - _i - 1) / (_i + 1):.2f} seconds left')

        # back to subwords
        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
    else:
        return []

# online search
def run_addrandom_token(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'):
    restarts = restarts - int(restarts / 3)
    if restarts:
        # init
        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'][:, -1, :]
        ori_output = ori_output.detach()

        # add random tokens
        new_texts = []
        old_inv_sorted_mask = sorted(noise_mask, reverse=True)
        old_sorted_mask = sorted(noise_mask)
        for i in range(restarts):
            _input_ids = _input.input_ids[i].cpu().numpy().tolist()
            for noise_ind in old_inv_sorted_mask:
                _input_ids.insert(noise_ind + 1, np.random.choice(subword_num))
                _input_ids.insert(noise_ind, np.random.choice(subword_num))
            new_texts.append(_input_ids)
        new_mask = []
        for i in range(len(old_sorted_mask)):
            new_mask.append(old_sorted_mask[i] + 2 * i)
            new_mask.append(old_sorted_mask[i] + 2 * i + 1)
            new_mask.append(old_sorted_mask[i] + 2 * i + 2)

        noise_mask = new_mask

        _input['input_ids'] = torch.Tensor(new_texts).long()
        _input['attention_mask'] = torch.ones_like(_input['input_ids'])
        for k in _input.keys():
            _input[k] = _input[k].to(device)
        # print(f'_input {_input["input_ids"].shape}')
        # get noise
        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)
        
        _input_ = deepcopy(_input)
        del _input_['input_ids']

        start_time = time()
        for _i in range(step):
            bar.progress((_i + 1) / (step))
            # start perturb
            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'][:, -1, :]
            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')

        # back to subwords
        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
    else:
        return []

# 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 obstinate adversarial examples based on models', 'Use the obstinate adversarial substitutions 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 obstinate adversarial examples based on models':
    model_choice = st.selectbox(
        'Which model you would like to use?',
        # ('gpt2', "EleutherAI/gpt-neo-1.3B", "EleutherAI/gpt-neo-2.7B", "EleutherAI/gpt-neox-20b", "EleutherAI/gpt-j-6B")
        ('gpt2', 'gpt-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 obstinate adversarial examples based on models':
        tokenizer = AutoTokenizer.from_pretrained(model_choice)
    else:
        tokenizer = AutoTokenizer.from_pretrained('gpt2')
    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 obstinate adversarial substituitions 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 obstinate adversarial examples 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 gpt2 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 obstinate adversarial examples based on models':
                _bar_text = st.empty()
                bar = st.progress(0)
                outputs = run(model, tokenizer, model.transformer.wte, 
                    _bar_text=_bar_text, bar=bar, text=title, noise_mask=chose_indices, restarts=restarts, step=step,
                    model_choice=model_choice)
                outputs.extend(run_addrandom_token(model, tokenizer, 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="✅")
            # 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 gpt2 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.')