veilstream/CITATION.cff at main · nikshepsvn/veilstream · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
cff-version: 1.2.0
message: "If you use veilstream or PASP in academic work, please cite:"
title: "veilstream: Provenance-Anchored Streaming Pseudonymization for LLM Chat"
type: software
version: "0.1.0"
date-released: "2026-04-22"
authors:
  - family-names: "veilstream contributors"
license: Apache-2.0
repository-code: "https://github.com/nikshepsvn/veilstream"
abstract: >
  veilstream is an Apache-2.0 Python library implementing streaming
  pseudonymization for LLM chat gateways. It introduces PASP
  (Provenance-Anchored Streaming Pseudonymization), a wire-format
  extension that wraps each reversible pseudonym with an HMAC-derived
  anchor. The deanonymizer only reverses fully-anchored matches,
  structurally preventing the false-attribution attack present in prior
  streaming-PII libraries (Presidio, LLM Guard, LiteLLM, LangChain).
  Empirical evaluation across 13 frontier LLMs (Claude Opus 4.7 & Sonnet
  4.6, GPT-5.4 / 5.4-mini / 4o, Gemini 3 Pro & Flash, Grok 4.20 &
  4.1-fast, Llama 3.3 70B, Qwen 235B instruct & thinking, GLM 4.7)
  shows exact-wrapper retention with median 100% and mean 96.4% under
  the veilstream system-prompt hint, and a measured 2–32%
  false-attribution rate (median 12%, 16× across LLMs) in vulnerable
  deployments under legacy (non-anchored) reversal — converted to 0%
  by PASP.
keywords:
  - pii
  - privacy
  - streaming
  - llm
  - pseudonymization
  - hmac
  - provenance