app.py

"""Streamlit app for Presidio + Privy-trained PII models."""

import spacy
import en_spacy_pii_distilbert
from spacy_recognizer import CustomSpacyRecognizer
from presidio_analyzer.nlp_engine import NlpEngineProvider
from presidio_anonymizer import AnonymizerEngine
from presidio_analyzer import AnalyzerEngine, RecognizerRegistry
import pandas as pd
from annotated_text import annotated_text
from json import JSONEncoder
import json
import warnings
import streamlit as st
# from streamlit import logger as _logger
import os
import csv
import json
from chatgpt_wrapper import ChatGPT
import time

os.environ["TOKENIZERS_PARALLELISM"] = "false"
warnings.filterwarnings('ignore')

# from flair_recognizer import FlairRecognizer

def load_data(file_location):
    unpacked_string_data = []
    unpacked_url_data = []
    unpacked_json_data = []
    # Read the data back from the CSV file and unpack it
    with open(file_location, mode='r') as csv_file:
        reader = csv.reader(csv_file)
        for row in reader:
            unpacked_string_data.append(row[0])
            unpacked_url_data.append(row[1])
            unpacked_json_data.append(json.loads(row[2]))
            # print("Unpacked string data:", unpacked_string_data)
            # print("Unpacked url data:", unpacked_url_data)
            # print("Unpacked JSON data:", unpacked_json_data)
    return unpacked_string_data, dict(zip(unpacked_string_data, unpacked_json_data))

# Helper methods
@st.cache_resource #(allow_output_mutation=True)
def analyzer_engine(use_local=None):
    """Return AnalyzerEngine."""

    spacy_recognizer = CustomSpacyRecognizer()
    if use_local:
        # !pip
        # install
        # https: // huggingface.co / beki / en_spacy_pii_distilbert / resolve / main / en_spacy_pii_distilbert - any - py3 - none - any.whl

        # Using spacy.load().
        nlp = spacy.load("en_spacy_pii_distilbert")

        # Importing as module.
        nlp_engine = en_spacy_pii_distilbert.load()
    else:
        configuration = {
            # print("ENALBEE MODELES")
            "nlp_engine_name": "spacy",
            "models": [
                {"lang_code": "en", "model_name": "en_spacy_pii_distilbert"}],
        }

        # Create NLP engine based on configuration
        provider = NlpEngineProvider(nlp_configuration=configuration)
        nlp_engine = provider.create_engine()

    registry = RecognizerRegistry()
    # add rule-based recognizers
    registry.load_predefined_recognizers(nlp_engine=nlp_engine)
    registry.add_recognizer(spacy_recognizer)
    # remove the nlp engine we passed, to use custom label mappings
    registry.remove_recognizer("SpacyRecognizer")

    analyzer = AnalyzerEngine(nlp_engine=nlp_engine,
                              registry=registry, supported_languages=["en"])

    # uncomment for flair-based NLP recognizer
    # flair_recognizer = FlairRecognizer()
    # registry.load_predefined_recognizers()
    # registry.add_recognizer(flair_recognizer)
    # analyzer = AnalyzerEngine(registry=registry, supported_languages=["en"])
    return analyzer


@st.cache_resource#(allow_output_mutation=True)
def anonymizer_engine():
    """Return AnonymizerEngine."""
    return AnonymizerEngine()


def get_supported_entities():
    """Return supported entities from the Analyzer Engine."""
    return analyzer_engine().get_supported_entities()


def analyze(**kwargs):
    """Analyze input using Analyzer engine and input arguments (kwargs)."""
    if "entities" not in kwargs or "All" in kwargs["entities"]:
        kwargs["entities"] = None
    return analyzer_engine().analyze(**kwargs)


def anonymize(text, analyze_results):
    """Anonymize identified input using Presidio Abonymizer."""
    if not text:
        return
    res = anonymizer_engine().anonymize(text, analyze_results)
    return res.text


def annotate(text, st_analyze_results, st_entities):
    tokens = []
    # sort by start index
    results = sorted(st_analyze_results, key=lambda x: x.start)
    for i, res in enumerate(results):
        if i == 0:
            tokens.append(text[:res.start])

        # append entity text and entity type
        tokens.append((text[res.start: res.end], res.entity_type))

        # if another entity coming i.e. we're not at the last results element, add text up to next entity
        if i != len(results) - 1:
            tokens.append(text[res.end:results[i+1].start])
        # if no more entities coming, add all remaining text
        else:
            tokens.append(text[res.end:])
    return tokens


st.set_page_config(page_title="Bitahoy demo", layout="wide")

# Side bar -------------------------------------------
# add picture with
st.sidebar.image("structured-data-anonymizer/assets/bitahoy-logo.png", width=200)

st.sidebar.markdown(
"""Detect and anonymize PII in structured text such as protocol traces (JSON, SQL, XML etc.)"""
)

# dropdown
# titles, json_dict = load_data("structured-data-anonymizer/assets/data_s_short.csv")
# option_list = titles
# option = st.sidebar.selectbox(
#     'Choose an existing structured input?',
#     option_list)

# dropdown df
# Title,Url,Dict,Prompt,Result
dataframe = pd.read_csv("structured-data-anonymizer/assets/df_data_short.csv")
# select only the third column of the data frame
# select only first column of the data frame
titles = dataframe['Title']
# conver it to a list
titles = titles.values.tolist()
# print(dataframe.iloc[0])
# select first row from dataframe
option_list = titles
# for i in option_list:
#     if (dataframe[dataframe['Title'] == i]['Result'].empty):
#         i = i + "*"
# print(option_list)
option = st.sidebar.selectbox(
    'Choose an existing structured input?',
    option_list)

# # st.sidebar.write('You selected:', option)
# json_dict = dataframe['Dict']
# json_dict = json_dict.values.tolist()
sidebar_text = 'Use small icon-button in right corner to copy input to clipboard'
st.sidebar.write(sidebar_text)
json_dict_option = dataframe[dataframe['Title'] == option]['Dict'].values[0]

st.sidebar.code (json_dict_option)

#romans complex dropdown
# st.checkbox("Enable/Disable input of existing data", key="disabled")
#
#     option = st.selectbox(
#         "Choose an existing structured input?",
#         option_list,
#         # label_visibility=st.session_state.visibility,
#         disabled=st.session_state.disabled,
#     )
#     st.write('You selected:', option)

st_entities = st.sidebar.multiselect(
    label="Which entities to look for?",
    options=get_supported_entities(),
    default=['PHONE_NUMBER', 'CREDIT_CARD', 'DATE_TIME', 'MEDICAL_LICENSE', 'US_BANK_NUMBER', 'IP_ADDRESS', 'IBAN_CODE', 'LOCATION', 'EMAIL_ADDRESS']
#     default=list(get_supported_entities()),
)
# ['PHONE_NUMBER', 'PERSON', 'CRYPTO', 'AU_TFN', 'ORGANIZATION', 'UK_NHS', 'CREDIT_CARD', 'US_DRIVER_LICENSE',
# 'US_SSN', 'URL', 'AU_MEDICARE', 'DATE_TIME', 'NRP', 'US_PASSPORT', 'MEDICAL_LICENSE', 'US_BANK_NUMBER',
# 'IP_ADDRESS', 'IBAN_CODE', 'US_ITIN', 'AU_ACN', 'SG_NRIC_FIN', 'LOCATION', 'AU_ABN', 'EMAIL_ADDRESS']
# st.sidebar.text(list(get_supported_entities()))

st_threshold = st.sidebar.slider(
    label="Acceptance threshold", min_value=0.0, max_value=1.0, value=0.35
)

st_return_decision_process = st.sidebar.checkbox(
    "Add analysis explanations in json")

api_togg = st.sidebar.checkbox(label='API toggle', value=True)

# vertical space
st.sidebar.text("")
# vertical space
st.sidebar.text("")


st.sidebar.info(
    "Privy is an open source framework for synthetic data generation in protocol trace formats (json, sql, html etc). Presidio is an open source framework for PII detection and anonymization. "
    "For more info visit [privy](https://github.com/pixie-io/pixie/tree/main/src/datagen/pii/privy) and [aka.ms/presidio](https://aka.ms/presidio)"
)


# Main panel
if 'first_load' not in st.session_state:
    st.session_state['first_load'] = True


analyzer_load_state = st.info(
    "Starting analyzer and loading model...")
engine = analyzer_engine()
analyzer_load_state.empty()

# Initialization
# if 'bot' not in st.session_state:
#     st.sidebar.text("init...")
#     st.session_state['bot'] = ChatGPT()
#     init_prompt = "i'd like you to act like a snobby AI and tell me what you think of my structured data"
#     init_answer = st.session_state['bot'].ask(init_prompt)

# col?
# Store the initial value of widgets in session state
if "visibility" not in st.session_state:
    st.session_state.visibility = "visible"
    st.session_state.disabled = False

col1, col2 = st.columns(2)

with col1:
    st.subheader("Input")

    sys_name = st.text_area(
        label="Name of the system in question",
        value=option,
        height=1,
    )

    st_text = st.text_area(
        label= "Structured text used as input",
        value = """{ "@timestamp":"2022-06-08T16:54:58.849Z", "alienOTX":{ "firewall":{ "action":"Deny", "category":"AlienVaultFirewallNetworkRule", "icmp":{ "request":{ "code":"8" } }, "operation_name":"AzureFirewallNetworkRuleLog", "path": "http://www.example.com/ab001.zip", }, "resource":{ "group":"TEST-FW-RG", "id":"/SUBSCRIPTIONS/23103928-B2CF-472A-8CDB-FR7630006000011234567890189/RESOURCEGROUPS/TEST-FW-RG/PROVIDERS/MICROSOFT.NETWORK/AZUREFIREWALLS/TEST-FW01", "address":"172.24.0.4", "provider":"SonicWall", "number":"040084913373", "sentto": "[email protected]" }, "subscription_id":"4012888888881881-23103928-B2CF-472A-8CDB-0146E2849129" } }""",
        # value="SELECT shipping FROM users WHERE shipping = '201 Thayer St Providence RI 02912'"
        # "\n\n"
        # "{user: Willie Porter, ip: 192.168.2.80, email: [email protected]}",
        height=300,
    )
    button = st.button("Detect and replace PII")
    st.text("""""")

with col2:
    st.subheader("Analyzed results with detected entities highlighted")
    # st.text("Output text with detected entities highlighted")
    with st.spinner("Analyzing..."):
        if button or st.session_state.first_load:
            option = sys_name
            st_analyze_results = analyze(
                text=st_text,
                entities=st_entities,
                language="en",
                score_threshold=st_threshold,
                return_decision_process=st_return_decision_process,
            )
            # """
            # Ugly hack that checks if last 2 chars as Z" and changes the end of the last entity to -1
            # This is done to prevent the anotation to inlcude the quotes for the date4 and breka the json donwtheroad
            # ### TODO: make this less hacky?
            # """
            for i in st_analyze_results:
                # st.write(i)
                # st.write(st_text[i.end - 2:i.end])
                if st_text[i.end-2:i.end] == 'Z"':# and i.type == "DATE_TIME":
                    i.end = i.end-1
                    continue
                if st_text[i.end-2:i.end] == "Z'":# and i.type == "DATE_TIME":
                    i.end = i.end-1
                    continue
                # if "'" in st_text[i.start:i.end]:
                #     st_analyze_results.remove(i)
                #     continue
                # if "," in st_text[i.start:i.end]:
                #     st_analyze_results.remove(i)
                #     continue


            annotated_tokens = annotate(st_text, st_analyze_results, st_entities)
            # annotated_tokens
            annotated_text(*annotated_tokens)

    # vertical space
    st.text("")
    st.text("")

    with st.expander("Show results with replaced PII and detailed results"):
        # st.subheader("Final results with tokens instead if PII")
        # vertical space

        if button or st.session_state.first_load:
            st_anonymize_results = anonymize(st_text, st_analyze_results)
            st.write(st_anonymize_results)
            # st.write(st_anonymize_results)
            # try:
            #     # st_anonymize_results = ast.literal_eval(st_anonymize_results)
            #     st.json(st_anonymize_results) #.replace("'", '"'))
            # except Json Parse Error as e:
            #     st.write(st_anonymize_results)
        # vertical space
        st.text("")
        st.subheader("Detailed Findings")
        if st_analyze_results:
            res_dicts = [r.to_dict() for r in st_analyze_results]
            for d in res_dicts:
                d['Value'] = st_text[d['start']:d['end']]
            df = pd.DataFrame.from_records(res_dicts)
            df = df[["entity_type", "Value", "score", "start", "end"]].rename(
                {
                    "entity_type": "Entity type",
                    "start": "Start",
                    "end": "End",
                    "score": "Confidence",
                },
                axis=1,
            )

            st.dataframe(df, width=1000)  # , height=500)
        else:
            st.text("No findings")

# st_analyze_results
# end of col

# After the columns

col5, col6 = st.columns(2)
prompt = "Write a summary for a {} event log, based on the given structured JSON input. Start with an executive summary with a short general description of what is a {}, and then focus on the Key Findings, Monitoring Summary,  Incident Summary, Threat Summary and Recommendations. Replace any random " \
            "strings and tokens in angular-brackets with an approximations to make it more human readable: \"{}\" ".format(
        option, option,
        st_anonymize_results)

with col5:
    st.subheader("Formatting")

    button_create = st.button("Create summary")
    st.markdown(
        "Start with an executive summary and describe what system the log came from, then focus on the Key Findings, Monitoring Summary,  Incident Summary, Threat Summary and Recommendations.")
    st.text("""""")

    with st.expander("Additional inputs"):
        st_prompt = st.text_area(
            label="Tokenized input with the formatted prompt",
            value=prompt,
            height=200,
        )

        write_results = ""
        st_output = st.text_area(
            label="Record results for later use",
            value=write_results,
            height=100,
        )
        button_save = st.button("Save summary to file?")
        st.text("""""")

        placeholder_table = st.empty()
        placeholder_table.write("")

        init_prompt = """I want you to act as a cyber security analyst expert. I will provide some specific information about concrete incidents, and it will be your job to come up with a coherent summery of the event, described in this log I give you. You can give a short description and then give strategies for protecting this system from malicious actors, based on the incident data I give you. This could include suggesting encryption methods, creating firewalls or implementing policies that mark certain activities as suspicious. Your summery would be used by decision makers to manage the situation, therefore make informed predictions and formulate them precisely in relation to the event I present to you."""
        st_init_prompt = st.text_area(
            label="Initial promopt to focus model",
            value=init_prompt,
            height=100,
        )
        button_reset = st.button("Reset model setup")

import random

with col6:
    st.subheader("Output incident summary")

# effect button_create(button2)
# with st.spinner("button_create..."):
if button_create:
    # load existing promp and results
    if (not api_togg):
        saved_prompt = dataframe[dataframe['Title'] == option]['Prompt'].values[0]
        saved_result = dataframe[dataframe['Title'] == option]['Result'].values[0]
    else:
        saved_prompt = ""
        saved_result = ""
    # check if match to current prompt
    # if re.sub(r"[\n\t\s]*", "", saved_prompt) == re.sub(r"[\n\t\s]*", "", st_prompt):
    md_results = ""
    with col6:
        x = st.empty()
        x.markdown("")

    # check if saved_prompt is not of a type float
    if (not isinstance(saved_prompt, float)) and (not api_togg):
        # st.write(saved_prompt)
        with col6:
            with st.spinner('Fetching results...'):
                time.sleep(random.uniform(2.1, 5.8))
                # st.write("Prompt already queried in the past, loading result from database")
                md_results = saved_result
                words = md_results.split()
                num_words = len(words)
                chunk_size = int(random.uniform(2, 6))
                str_placeholder = ""

                for i in range(0, num_words, chunk_size):
                    chunk = ' '.join(words[i:i + chunk_size])
                    str_placeholder = str_placeholder + " " + chunk
                    x.markdown(str_placeholder)
                    # x.markdown(chunk)
                    time.sleep(random.uniform(0.1, 0.6))
                x.markdown(saved_result)
    else:
        # st.write("New prompt, need GPT")
        with col6:
            with st.spinner('Generating, please wait...'):
                bot = ChatGPT()
                # init_answer = bot.ask(init_prompt)

                init_points = ""
                for chunk in bot.ask_stream(init_prompt):
                    init_points = init_points + "."
                    x.markdown(init_points)

                x.markdown("")

                # st_prompt = "tell me two facts about yourself"
                for chunk in bot.ask_stream(st_prompt):
                    md_results = md_results + chunk
                    x.markdown(md_results)
                    #check if last char of chunk is a new line
                    # if "\n" in chunk:
                    #     x.markdown(md_results)
                    # st.markdown(chunk)
                x.markdown(md_results)
                bot._cleanup()
                # md_results = bot.ask(st_prompt) #"Hello, could you tell what is {}?".format(option))
                # print(md_results)  # prints the response from chatGPT

                # st.write(st_prompt)
                # st.write(saved_prompt)
            # md_results = """No result found""" ##here GPT
    # with col6:
    #     # st.subheader("Output incident summary")
    #     st.markdown(md_results)
    placeholder_table.write((dataframe.loc[dataframe['Title'] == option]))

# if button_reset:
#     bot = ChatGPT()
#     bot._cleanup()

if button_save:
    # dataframe = pd.read_csv("structured-data-anonymizer/assets/df_data_short.csv")
    # save st_prompt and st_output to dataframe in row for Title = json_dict_option
    dataframe.loc[dataframe['Title'] == option, 'Prompt'] = st_prompt
    dataframe.loc[dataframe['Title'] == option, 'Result'] = md_results #st_output
    # st.write(json_dict_option)
    # write dataframe back to the csv file
    dataframe.to_csv("structured-data-anonymizer/assets/df_data_short.csv", index=False)
    st.write("Saved to file")
    st.write(dataframe.loc[dataframe['Title'] == option])

# end of document
st.session_state['first_load'] = True



class ToDictListEncoder(JSONEncoder):
    """Encode dict to json."""

    def default(self, o):
        """Encode to JSON using to_dict."""
        if o:
            return o.to_dict()
        return []


if st_return_decision_process:
    st.json(json.dumps(st_analyze_results, cls=ToDictListEncoder))