Effective Personal ID Information (PII) Guardrails for LLMs

You will learn about:

• General problem: The user can obtain PII
• Classifier-based approach given a context
• PII detectors capabilities

More about the event
Protecting sensitive information is crucial in today’s world of Large Language Models. (LLMs) and data-driven services. For LLMs, data extraction is a significant threat in practice since attackers with black-box API access can extract at least 1% of the training data. Any LLM trained on real, sensitive data must protect PII, but contextual PII detection is not well explored in existing art. Extracting any PII by itself, such as a personal address can already pose a privacy threat. This threat is elevated when an attacker can associate a piece of PII to a context, for example “Person X received chemotherapy at Stanford hospital in October 2023’.

We addressed this business problem by designing PIIGuardrails, a guard for large language model risk mitigator for both input and output to safeguard models by detecting personally identifiable information (PII). The multi-stage process includes a classifier-based approach that looks at the context of the potential spot. The key capabilities include malicious intent detection, detecting context-based compound sensitive PII’s, domain based PII coverage etc. Our machine learning based PII detectors are enhanced with taxonomy-based intelligence.

PII Guardrails empowers client with Trust on LLMs with Enterprise data. It establishes leadership in use cases such as model inferencing stage, model training and acquisition stage, and provide enhanced capabilities for privacy issues.

Our speaker
Shubhi Asthana is a Sr Research Software Engineer who builds AI & ML Solutions. She is SME in AI and ML models for Financial Services, along with leading the PII effort in Unstructured Data & NLP. Her research and development work spans the areas of Data Analytics, NLP and Cloud Services.

Shubhi has successfully led projects in services computing and data analytics with business impact. In 2020-2022, she led a financial services based risk analytics solution for a customer. Shubhi rapidly architected and delivered an innovative solution that accurately predicts when customer would exhaust their PO funds using a combination of machine learning and software engineering. The project was presented at top tier conferences such as KDD, IEEE Services etc. Shubhi also received the IBM Research’s OTAA Award 2021, 2023 for leading these projects.

Github: https://github.com/sasthan
Publications*: https://scholar.google.com/citations?user=G1fuJ4EAAAAJ&hl=en*
Personal blog*: https://medium.com/@shubhi.asthana*