"""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": "willh@hotmail.com" }, "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: willie@gmail.com}", 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))