inferwire
/
AI·2 min read

Samsung TV Software Runs an AlexNet Model to Guess Your Game

An X user found a Samsung TV running a quantized AlexNet on 960×540 screen captures every few seconds to identify the game being played, apparently before any consent check.

TL;DR

  • X user @mike64_t found a Samsung TV running a TensorFlow Lite model on 960×540 screen images every couple of seconds to identify the current game [^1].
  • By their account it rarely works and its output seems to go nowhere, because the consent check comes later [^1].

Background

TensorFlow Lite, now called LiteRT, is Google's framework for running machine-learning models directly on devices instead of in the cloud [^2]. Samsung's smart TVs run its Tizen operating system. On-device models are often presented as the privacy-friendly option, because the data never has to leave the device.

What happened

On October 24, an X account posting as @log4j live-tweeted findings from an analysis of Tizen on a running TV [^1]. Following up, @mike64_t wrote that the TV runs an INT8-quantized TensorFlow Lite model on 960×540 images "every couple of seconds in an attempt to figure out what game you are playing" [^1]. The model file is named alex.tflite and, according to the post, is the classic AlexNet image network, apparently following torchvision's implementation [^1].

The post says the model compares image features against a database by nearest-neighbor search and "literally doesn't work 99% of the time," while still costing CPU time every other second [^1]. Its results, the author writes, appear to be discarded "because the consent gate comes later" [^1]. These are one person's findings from a single device; Samsung had not responded publicly in the sources available.

Why it matters

The finding cuts against the usual comfort of on-device AI. Analysis that stays on the TV is still analysis of what is on your screen, and in this case it reportedly starts before you have agreed to anything, even if the results are thrown away. It also shows how little visibility owners have: a model that captures the screen every few seconds was found by someone pulling apart the firmware, not through any setting or disclosure.

Practical example

Say you plug a console into a recent Samsung TV and start playing. If the behavior described applies to your model, the TV is sampling the picture every couple of seconds and comparing it with a database of games, whether or not you have accepted any viewing-data terms [^1]. The only way you would know is the kind of firmware inspection described in the post.

Related gear

We recommend this book because it explains how engineers build and deploy compact TensorFlow Lite models directly onto embedded hardware and microcontrollers.

AdvertisementAmazon

TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers

Sources

  1. [1]mike (@mike64_t) on X — analysis of a TensorFlow Lite model on a Samsung TV
  2. [2]Google for Developers — LiteRT (formerly TensorFlow Lite)
  3. [3]Hacker News — Samsung running a TensorFlow lite model on 960x540 image to track you