README.md

metadata

title: PII Detection with BERT
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: apache-2.0

PII Detection with BERT

This Space demonstrates a BERT model fine-tuned for detecting Personal Identifiable Information (PII) in text.

Model Details

• Base Model: google-bert/bert-base-uncased
• Training Dataset: ai4privacy/pii-masking-43k
• Task: Token Classification / Named Entity Recognition (NER)
• Number of Entity Types: 27

Detectable PII Types

The model can identify 27 different types of personal information:

Identity Information

• NAME, USERNAME, DISPLAYNAME, GENDER, JOB

Contact Information

• EMAIL, STREET, ADDRESS, ZIPCODE, GEO, NEARBYGPSCOORDINATE

Financial Information

• CREDITCARDNUM, CREDITCARDISSUER, IBAN, BIC
• ACCOUNTNAME, ACCOUNTNUM, CURRENCY, COINADDRESS

Technical Information

• IP, MAC, URL, USERAGENT, PASSWORD

Other

• NUM, ORDINALDIRECTION

How It Works

1. Input: User provides text that may contain personal information
2. Tokenization: Text is split into tokens using BERT tokenizer
3. Classification: Each token is classified into one of 27 entity types or "O" (no entity)
4. Visualization: Detected entities are highlighted with different colors

Training Details

• Learning Rate: 5e-05
• Batch Size: 16 (train), 64 (eval)
• Epochs: 3
• Optimizer: Adam (β1=0.9, β2=0.999, ε=1e-08)
• Warmup Steps: 500

Use Cases

• Data Privacy: Identify PII before sharing documents
• Data Anonymization: Find information that needs masking
• Compliance: Help meet GDPR, CCPA requirements
• Security: Detect sensitive information leaks

Limitations

• Maximum input length: 512 tokens
• Optimized for English text
• May not detect all variations of PII
• Performance depends on text format and quality

Example Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification

model_name = "your-username/your-space-name"  # Update after deployment
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

text = "My name is John Smith and my email is john@example.com"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

License

Apache 2.0