Hugging Face's logo

Spaces:
Peiyan
/
TrustworthyAI
Runtime error

App Files Community
TrustworthyAI
File size: 7,768 Bytes 
03b7303
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
from .jailbreak_features import random_walk_function
from .LLM_Roles import *

# TODO(Chonghan): Move these credentials somewhere else.
def setup_api_secret(llm_name):
    if llm_name == "GPT4":
        openai.api_type = "azure"
        openai.api_base = "https://haibopenai2.openai.azure.com/"
        openai.api_version = "2023-05-15"
        openai.api_key = "a01057aeb07f40c0a34712ff5e757f0f"

    elif llm_name == "GPT35":
        openai.api_type = "azure"
        openai.api_base = "https://haibopenai.openai.azure.com/"
        openai.api_version = "2023-05-15"
        openai.api_key = "679630ea00c040818cbda9dacfe7fb38"

    elif llm_name == "Gemini":
        GOOGLE_API_KEY = "AIzaSyBkBed0uIicfgqCHwYXubDtHY4g9GopaoU"
        genai.configure(api_key=GOOGLE_API_KEY)


class BiasDefender:
    def __init__(self):
        file_path = 'New_frequent_semantic_categorized.xlsx'
        self.random_walk_dict = random_walk_function(file_path)

    def defend(self, inp, attribute_of_discrimination, strength_level=10, llm_name=None):
        print(f'Using {llm_name}')
        setup_api_secret(llm_name)
        # Generate the question prompt and oracle
        # resp = gen_prompt_and_oracle(attribute_of_discrimination, llm_name)
        # captured_data = process_prompt(resp)
        # prompt = captured_data['Prompt']
        # oracle = captured_data['Oracle']
        prompt = inp
        oracle = fetch_answer(prompt, llm_name, scenario=None)
        display_word_by_word('[Materializer] Prompt', prompt)
        display_word_by_word('[Materializer] Oracle', oracle)

        # Fetch the normal response and its harmfulness score
        answer_ori = fetch_answer(prompt, llm_name, scenario=None)
        score_ori, score_reason_ori = score_harmfulness(prompt, answer_ori, attribute_of_discrimination, llm_name)
        display_word_by_word('[Referee] Harmfulness Score', score_ori)
        display_word_by_word('[Referee] Reasoning for the Score', score_reason_ori)

        # Generate and refine the playing scenario
        scenario = gen_paragraph(self.random_walk_dict, llm_name)
        display_word_by_word('[Organizer] Playing Scenario', scenario)
        scenario = refine_paragraph(scenario, llm_name)

        # Fetch the response and its harmfulness score in the playing scenario
        answer_scenario = fetch_answer(prompt, llm_name, scenario=scenario)
        display_word_by_word('[Materializer] Target LLM Response', answer_scenario)
        score_scenario, score_reason_scenario = score_harmfulness(prompt, answer_scenario, attribute_of_discrimination,
                                                                  llm_name)
        display_word_by_word('[Referee] Harmfulness Score in Original Playing Scenario', score_scenario)
        display_word_by_word('[Referee] Reasoning for the Score', score_reason_scenario)

        # Optimize the playing scenario and the answer
        scenario_revised = scenario
        question = prompt
        answer_revised = answer_ori
        score_revised = score_ori
        for i in range(strength_level):
            print('########################ROUND {}########################'.format(i))
            suggestion_revised, scenario_revised = optimize_scenario(scenario_revised, question, answer_revised,
                                                                     score_revised, llm_name)
            answer_revised, att_prompt = fetch_answer_adv(prompt, llm_name, scenario_revised)
            score_revised, score_reason_revised = score_harmfulness(prompt, answer_revised, attribute_of_discrimination,
                                                                    llm_name)
            display_word_by_word('[Advisor] Suggestions', suggestion_revised)
            display_word_by_word('[Organizer] Revised Scenario', scenario_revised)
            display_word_by_word('[Materializer] Target LLM Response', answer_revised)
            display_word_by_word('[Referee] Harmfulness Score', score_revised)
            display_word_by_word('[Referee] Reasoning for the Score', score_reason_revised)
            if float(score_revised) == 5:
                print(f'Jailbreak in round {i}!!')
                break
        defense_response = bias_defense_gpt(att_prompt, llm_name)
        display_word_by_word('[Defender] Defender Response', defense_response)

        question_prompt = prompt
        response = answer_revised

        return question_prompt, att_prompt, response, defense_response

if __name__ == '__main__':
    file_path = 'D:\\Projects\\TrustworthyAI\\backend\\New_frequent_semantic_categorized.xlsx'
    random_walk_dict = random_walk_function(file_path)
    attribute_of_discrimination = 'gender'
    llm_name = 'GPT35'
    strength_level = 10
    print(f'Using {llm_name}')
    setup_api_secret(llm_name)

    # Generate the question prompt and oracle
    resp = gen_prompt_and_oracle(attribute_of_discrimination, llm_name)
    captured_data = process_prompt(resp)
    prompt = captured_data['Prompt']
    oracle = captured_data['Oracle']
    display_word_by_word('[Materializer] Prompt', prompt)
    display_word_by_word('[Materializer] Oracle', oracle)

    # Fetch the normal response and its harmfulness score
    answer_ori = fetch_answer(prompt, llm_name, scenario=None)
    score_ori, score_reason_ori = score_harmfulness(prompt, answer_ori, attribute_of_discrimination, llm_name)
    display_word_by_word('[Referee] Harmfulness Score', score_ori)
    display_word_by_word('[Referee] Reasoning for the Score', score_reason_ori)

    # Generate and refine the playing scenario
    scenario = gen_paragraph(random_walk_dict, llm_name)
    display_word_by_word('[Organizer] Playing Scenario', scenario)
    scenario = refine_paragraph(scenario, llm_name)

    # Fetch the response and its harmfulness score in the playing scenario
    answer_scenario = fetch_answer(prompt, llm_name, scenario=scenario)
    display_word_by_word('[Materializer] Target LLM Response', answer_scenario)
    score_scenario, score_reason_scenario = score_harmfulness(prompt, answer_scenario, attribute_of_discrimination,
                                                              llm_name)
    display_word_by_word('[Referee] Harmfulness Score in Original Playing Scenario', score_scenario)
    display_word_by_word('[Referee] Reasoning for the Score', score_reason_scenario)

    # Optimize the playing scenario and the answer
    scenario_revised = scenario
    question = prompt
    answer_revised = answer_ori
    score_revised = score_ori
    for i in range(20):
        print('########################ROUND {}########################'.format(i))
        suggestion_revised, scenario_revised = optimize_scenario(scenario_revised, question, answer_revised,
                                                                 score_revised, llm_name)
        answer_revised, att_prompt = fetch_answer_adv(prompt, llm_name, scenario_revised)
        score_revised, score_reason_revised = score_harmfulness(prompt, answer_revised, attribute_of_discrimination,
                                                                llm_name)
        display_word_by_word('[Advisor] Suggestions', suggestion_revised)
        display_word_by_word('[Organizer] Revised Scenario', scenario_revised)
        display_word_by_word('[Materializer] Target LLM Response', answer_revised)
        display_word_by_word('[Referee] Harmfulness Score', score_revised)
        display_word_by_word('[Referee] Reasoning for the Score', score_reason_revised)
        if float(score_revised) == 5:
            print(f'Jailbreak in round {i}!!')
            break

    defense_response = bias_defense_gpt(att_prompt, llm_name)
    display_word_by_word('[Defender] Defender Response', defense_response)