Author(s): Souvik Sarkar

Originally published on Towards AI.

5 key techniques explained.

Every team building on large language models faces the same quiet risk. Training data, fine-tuning datasets, retrieval-augmented generation inputs, and user prompts all carry personally identifiable information (PII), meaning names, emails, phone numbers, addresses, and health records that belong to real people.

After the introduction, the article explains how PII can enter an LLM pipeline at several points—especially training data, user prompts, and RAG knowledge bases—and how each stage creates different leak risks (memorization, plaintext logging/session exposure, and retrieval of unredacted records). It then lays out five anonymization techniques—data masking (including format-preserving), pseudonymization with secure lookup tables, generalization to reduce precision, data swapping to break meaningful record links, and synthetic data generation—and notes that the right combinations depend on which pipeline stage you’re protecting. Finally, it argues that anonymization alone isn’t sufficient for LLMs: you need governance treated like core infrastructure (PII detection, role-based access, audit logs, and compliance mapping) plus continuous auditing, because unstructured text undermines field-level rules, models can re-identify from non-PII attributes, and training-time memorization can still surface sensitive information in outputs.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI