Sensitive Data Discovery with American Express

PII Redaction Text Classifier

This project builds a machine learning model to detect and redact sensitive personal information (PII) from internal text data, enhancing customer privacy and generating compliant test datasets.

📌 Key Features

• PII Detection and Redaction: Automatically detects and redacts sensitive data types, including names, phone numbers, and financial information.
• Compliance-Friendly Dataset Generation: Ensures that internal datasets meet privacy and regulatory requirements by masking PII fields.
• Scalable and Extensible: Modular code allows for scaling and adapting to additional PII types or sources.

🔧 Techniques Used

• Preprocessing and Vectorization: Utilizes scikit-learn for efficient data processing and transformation.
• Deep Learning with DeBERTa: Employs DeBERTa for high-accuracy PII entity recognition and context-aware redaction.
• Regex-Based Masking: Complements DeBERTa with Regex for faster, pattern-based PII detection, enhancing model accuracy for well-defined data patterns.

📚 Dataset

PII Masking Dataset (200k): A dataset with labeled PII used to train and validate the redaction model.

👣 Potential Next Steps

• Dataset: Add more attributes to improve data identification, find a more comprehensive/representative dataset
• Models: Reduce redundant labels in output, make ensemble model to better generalize each attribute

Contributors:

• Ayan Gaur
• Joy Chang
• Shrieyaa Sekar Jayanthi
• Shubhangi Waldiya
• Stella Huang
• Break Through Tech AI at UCLA & American Express