Sriracha

Privacy-preserving person record linkage (PPRL) library

API is unstable. Not production-ready.

Sriracha is a Go library for privacy-preserving record linkage. It enables institutions to share person records across organizational boundaries without transmitting raw PII. Sriracha provides the building blocks for building privacy-first transports. Records are normalized and tokenized with a shared secret, producing tokens that can be compared without exposing the underlying identifiers.

Features

• Deterministic tokenization using HMAC-SHA256
• Probabilistic tokenization with Sørensen–Dice matching
• Optional BLIP and balanced filter defenses against frequency analysis
• Unicode normalization pipeline
• Canonical field set with support for extended schemas

Installation

Requires Go 1.24+

go get github.com/ccuetoh/sriracha@v0.1.0

Quickstart

package main

import (
	"fmt"

	"github.com/ccuetoh/sriracha"
	"github.com/ccuetoh/sriracha/fieldset"
	"github.com/ccuetoh/sriracha/session"
)

func main() {
	secret := []byte("super-secret-key")

	s, _ := session.New(secret, fieldset.DefaultFieldSet())
	defer s.Destroy()

	// Deterministic tokenization
	tokA, _ := s.TokenizeDeterministic(sriracha.RawRecord{
		sriracha.FieldNameGiven:  "Alice",
		sriracha.FieldNameFamily: "Smith",
	})

	tokB, _ := s.TokenizeDeterministic(sriracha.RawRecord{
		sriracha.FieldNameGiven:  "Alice",
		sriracha.FieldNameFamily: "Smith",
	})

	eq := s.Equal(tokA, tokB)
	fmt.Printf("match: %v\n", eq)

	// Probabilistic tokenization
	bloomA, _ := s.TokenizeProbabilistic(sriracha.RawRecord{
		sriracha.FieldNameGiven:  "Alice",
		sriracha.FieldNameFamily: "Smith",
	})

	bloomB, _ := s.TokenizeProbabilistic(sriracha.RawRecord{
		sriracha.FieldNameGiven:  "Alice",
		sriracha.FieldNameFamily: "Smyth", // typo
	})

	result, _ := s.Match(bloomA, bloomB, 0.85)
	fmt.Printf("match: %v (score: %.2f)\n", result.IsMatch, result.Score)
}

Benchmarks

Live history on Bencher.

Corpus	Records	Pairs	AUROC	Accuracy	Recall
OpenSanctions	26 841	natural cross-source	0.93	0.91	0.87
FEBRL4	10 000	synthetic (FEBRL4 noise)	1.00	1.00	1.00
NCVR	8 848	synthetic (1–2 char edits)	1.00	1.00	1.00