An easy way to remove PII before sending to LLMs-Python Tutorial-php.cn

An easy way to remove PII before sending to LLMs

Barbara Streisand

Release： 2024-11-25 20:20:18

Original

311 people have browsed it

An easy way to remove PII before sending to LLMs

Not all scenarios demand perfect anonymization. In less critical cases, a lightweight anonymization pipeline can suffice. Here, I share a Python-based approach leveraging GLiNER, Faker, and rapidfuzz to anonymize text by replacing sensitive entities with realistic placeholders.

The code first identifies sensitive entities (like names, countries, and professions) using GLiNER. Then, it replaces these entities with fake counterparts generated by Faker. Approximate string matching (rapidfuzz) ensures even variations in the text are anonymized. After processing with the LLM, the original entities are restored.

This method is designed for non-critical use cases where perfect anonymization isn't mandatory. For example, analyzing reviews or answering a query that comes to the chatbot on your website without saving data generally fall under less critical cases. The code is not perfect but good enough to get you started.

from gliner import GLiNER
from faker import Faker
from faker.providers import job
import google.generativeai as genai
import re
import warnings
from rapidfuzz import process, utils
warnings.filterwarnings("ignore")

genai.configure(api_key="key")
model_llm = genai.GenerativeModel("gemini-1.5-flash-002")
fake = Faker()
fake.add_provider(job)
model_gliner = GLiNER.from_pretrained("urchade/gliner_small-v2.1")

# let's say we have this prompt along with context that we want to anonymize before sending to LLM
prompt= f"""Given the context, answer the question. \n context: Hi, I am Mayank Laddha.  I lives in India. I love my country. But I would like to go to Singapore once. I am a software developer.\n question: Where does Mayank Laddha want to go?"
"""
# Perform entity prediction
labels = ["Person", "Country", "Profession"]
entities = model_gliner.predict_entities(prompt, labels, threshold=0.4)
print(entities)

# create a replacement dictionary
replacement = {}
for entity in entities: 
    if "Person" in entity["label"] and entity["text"] not in replacement:
        fake_set = {fake.name() for _ in range(3)}
        fake_set.discard(entity["text"])
        new_name = fake_set.pop()
        replacement[entity["text"]] = new_name
    elif "Country" in entity["label"] and entity["text"] not in replacement:
        name_set = {fake.country() for _ in range(10)}
        print(name_set)
        name_set.discard(entity["text"])
        new_name = name_set.pop()
        replacement[entity["text"]] = new_name
    elif "Profession" in entity["label"] and entity["text"] not in replacement:
        name_set = {fake.job() for _ in range(20)}
        name_set = {k for k in name_set if len(k.split())==1}
        print(name_set)
        name_set.discard(entity["text"])
        new_name = name_set.pop()
        replacement[entity["text"]] = new_name

#also create a reverse dictionary
replacement_reversed = {v: k for k, v in replacement.items()}

#perform replacement
for k, v in replacement.items():
    # Split text into a list of words
    words = prompt.split()  
    n = len(k.split()) 
    # so the key appears fully in choices
    choices = [' '.join(words[i:i+n]) for i in range(len(words) - n + 1)] 
    matches = process.extract(k, choices, limit=1, processor=utils.default_process)
    for match in matches:
        if match[1]>80:
            prompt = re.sub(match[0], v, prompt, flags=re.IGNORECASE)

#prompt
response = model_llm.generate_content(prompt)
content = response.text
print("llm response",content)

#perform replacement again
for k, v in replacement_reversed.items():
    words = content.split()  
    n = len(k.split())
    choices = [' '.join(words[i:i+n]) for i in range(len(words) - n + 1)]
    matches = process.extract(k, choices, limit=1, processor=utils.default_process)
    for match in matches:
        if match[1]>80:
            content = re.sub(match[0], v, content, flags=re.IGNORECASE)

print("final result", content)

Copy after login

The above is the detailed content of An easy way to remove PII before sending to LLMs. For more information, please follow other related articles on the PHP Chinese website!