A new benchmark, IndiaPII-Bench v1.0, has been developed to evaluate the effectiveness of Personally Identifiable Information (PII) detection tools on Indian identifiers. The benchmark found that the widely used Presidio tool, by default, fails to detect common Indian PII like Aadhaar numbers, PAN, and UPI IDs. While Presidio can be improved with custom recognizers, the study highlights the importance of validation checks, such as the Verhoeff checksum for Aadhaar numbers, to achieve high accuracy. The research also introduced MaskFlow, an open-source library designed to replace PII with reversible placeholders before data leaves a system, offering a privacy-preserving alternative for LLM API interactions. AI
IMPACT Highlights critical gaps in PII detection for LLM APIs handling Indian data, necessitating custom solutions or new tools like MaskFlow for compliance.
RANK_REASON The item introduces a new benchmark for evaluating PII detection tools on Indian data and presents findings on the performance of existing tools, along with a new open-source library. [lever_c_demoted from research: ic=1 ai=0.7]
- Aadhaar
- DPDP Act
- GSTIN
- Hugging Face
- India
- IndiaPII-Bench v1.0
- LangChain
- LiteLLM
- LlamaIndex
- MaskFlow
- Presidio
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →