Skip to content

Latest commit

 

History

History
120 lines (99 loc) · 8.82 KB

privacy_and_licenses.md

File metadata and controls

120 lines (99 loc) · 8.82 KB

Privacy and licenses

What is downloaded, and from where

Inference runs on the device: images, audio and text you pass to a task are not sent anywhere by these packages. The packages do download software and models, each pinned to one SHA-256 (SHA-384 for browser files) and rejected if the bytes differ.

At build time, the build hooks fetch the native runtime for the target:

What Host
Google's per-family MediaPipe libraries (Android, iOS, macOS, Linux, Windows), read from Google's release archive storage.googleapis.com

At run time:

What Host
Pinned models, only when the app sets ModelStore.allowDownloads = true or calls ModelStore itself; bundled models are fetched at build time by dart run mediapipe_core:bundle_models instead storage.googleapis.com; the EmbeddingGemma 2 models (Decision Maker and Retrieval) and the speech models (Parakeet, Phonon, Kokoro and PocketTTS) from huggingface.co, which redirects to its CDN (*.hf.co)
Google's browser runtime (web only), unless self-hosted cdn.jsdelivr.net

hooks.user_defines.mediapipe_core.asset_source, ModelStore(source: ...) and self-hosting the browser runtime replace these hosts with your own; see platform setup.

Google's usage logging

Google's MediaPipe runtimes send usage metrics to Google: the task, its running mode, latencies and input counts, never the images, audio or text it processes. On native platforms the packages also name Flutter as the host environment, with the versions of the OS, Dart and Flutter (for example iOS 26.4 / Dart 3.13.4 / Flutter 3.47.5), which the uploads record. In browsers they name Flutter Web as the host environment.

On Android and iOS the packages also give Google's libraries the app's ID and version, as Google's native SDKs do: the package name and versionCode on Android, the bundle ID and CFBundleShortVersionString on iOS. Every task accepts these fields, and with the pinned 1.2.3 libraries every task includes them in its uploads (checked with the fields set on macOS, the iOS Simulator and an Android emulator); with 1.2.2 only eight tasks did. On Android the uploads leave the app only because the packages run each library's JNI_OnLoad (see the table below). The packages set neither on macOS, Linux, Windows or in browsers.

Google's NOTICE opens with its MediaPipe Tasks privacy notice, which makes the app's developer responsible for getting users' informed consent where the law requires it. The packages cannot turn the logging off.

Where What the packages bundle Usage logging
macOS, Linux, Windows Google's per-family 1.2.3 libraries An uploader to https://play.googleapis.com/log in every library. Observed with Google's October 18 build on all three, every family: one request per task close, which Google accepts (HTTP 200). Captured again with 1.2.2 and 1.2.3 on macOS only, every task, with the uploads kept from Google. On macOS the User-Agent carries the process name. On Windows a close waits for its upload
iOS Google's per-family 1.2.3 libraries The same uploader. Observed with the October 18 build on the iOS Simulator (every family) and on an iPhone: one request per task close, which Google accepts, with the app's bundle ID and version, the iOS version and the device model in the User-Agent. Captured again with 1.2.2 on the Simulator only, with the same User-Agent on every upload, whichever task sent it. Captured with 1.2.3 on the Simulator, every task: the User-Agent is now the system's default (the executable name, CFBundleVersion and the CFNetwork and Darwin versions), without the bundle ID
Android Google's per-family 1.2.3 libraries The same uploader, which since Google's nightly of October 21 goes through Java and needs the JavaVM that each library's JNI_OnLoad stores. Flutter loads the library with dlopen, which does not run it, so the packages run it themselves, once per library, before the family's first task. Observed with 1.2.2 in a debug build on an API 31 emulator: one request per task close leaves the app, Interactive Segmenter's with the app's ID and version, and none without that call. Observed with 1.2.3 on the same emulator: every task's uploads carry the app's ID and version, and none leave without that call. Observed again with 1.2.2 in a debug build on a Pixel 8a (Android 15): each of eight task closes opened one connection to Google and sent about 1 to 1.3 KB, with no JNI errors in the log. The October 18 build did upload, opening its own connection: on an emulator, where Google accepts one request per task close, and on physical phones from Android 9 to 15, 64-bit and 32-bit, where each task close opens one connection to Google of the same size and waits for it, up to about 2.5 s on the oldest phone
Browsers Google's JavaScript runtimes Requests from the page to https://odml.pa.googleapis.com/v1/log when a task closes and every 60 seconds while one is open. Observed with the October 18 build in Chromium and WebKit, every family, which Google accepts

With the October 18 build every family uploaded, Decision Maker included: each was observed on macOS, Linux, Windows, the iOS Simulator and in browsers, and on Android phones for the vision, text, audio, Universal Embedder and Decision Maker closes. With that build the Semantic Retriever's uploads named no OS, host environment or host version; with 1.2.2 they name all three.

Block those endpoints if your policy requires it: inference is unaffected. A task's close waits for its upload, so refuse the connection (for example through the hosts file) rather than drop its packets, which makes each close wait for the upload to time out (about 4 seconds on iOS).

App Store privacy manifest

App Store Connect refuses an iOS app whose code calls an API on Apple's required-reason list without declaring why (ITMS-91053). Google's iOS libraries call some and carry no privacy manifest, and a Flutter build cannot add one to them, so an app that uses these packages declares them in its own ios/Runner/PrivacyInfo.xcprivacy, added to the Runner target in Xcode:

Families Category (NSPrivacyAccessedAPIType) Called through The gallery's reason
Every family NSPrivacyAccessedAPICategoryFileTimestamp stat, fstat C617.1: files inside the app's container, such as its models
Every family NSPrivacyAccessedAPICategoryUserDefaults NSUserDefaults CA92.1: values only the app reads
mediapipe_retrieval NSPrivacyAccessedAPICategoryDiskSpace statfs, fstatfs E174.1: checking for space before writing files

The reasons are the gallery's reading of what Google's code does, which Google has not confirmed; the app's developer answers for the ones it declares. The gallery's manifest is a starting point. The usage logging above is data the app's App Privacy answers in App Store Connect must describe.

Licenses

  • This repository's source: Apache-2.0, with upstream notices retained.
  • The gallery's bundled Roboto font (gallery/fonts/): the SIL Open Font License 1.1, Copyright 2011 The Roboto Project Authors. The font files are unmodified.
  • Google's MediaPipe runtimes: Apache-2.0. mediapipe_core carries the LICENSE and NOTICE Google ships with the libraries it pins (third_party/, identical for every family), which apps redistribute with the runtime, and every package keeps a copy of that NOTICE at its root. The vision package also keeps OpenCV's carotene notices (third_party/OPENCV_CAROTENE_NOTICES) until Google confirms whether its libraries need them.
  • The task models (the pins in each family's models.dart): Apache-2.0, per their model cards, except the Gemma-based text models and two speech models under CC BY 4.0: NVIDIA's Parakeet TDT 0.6B v3 (AudioModels.parakeetTdtV3) and Kyutai's PocketTTS (AudioModels.pocketTtsEn). An app that ships or downloads either credits its author (NVIDIA or Kyutai), for example in its open source notices, as Google's Text to Speech guide asks for PocketTTS.
  • The Text Embedder's EmbeddingGemma 300M (TextModels.embeddingGemma, embedding_gemma.task), the Proofreader and the Summarizer are Gemma models, under the Gemma Terms of Use and its Prohibited Use Policy. An app that ships or downloads them must pass those terms on to its users. The EmbeddingGemma 2 models of Decision Maker and Retrieval are not: their model cards say Apache-2.0.

python3 -B packages/mediapipe-core/tool/mirror_runtime_assets.py lists every pinned runtime and model with its license.